Imbalanced bearing data fault diagnosis method and device based on improved condition Wasserstein generative adversarial network

Abstract

The invention relates to an unbalanced bearing data fault diagnosis method based on an improved condition Wasserstein generative adversarial network. The problem of bearing data imbalance under a small sample condition is effectively relieved. According to the technical scheme, the method comprises the following steps: firstly, processing a bearing vibration signal by adopting a sliding window overlapping sampling strategy, and converting the bearing vibration signal into a time-frequency image by applying continuous wavelet transform so as to enhance fault feature representation; secondly, an improved condition Wasserstein generative adversarial network is used for carrying out data enhancement, the network measures sample distribution difference through a Wasserstein distance, introduces gradient penalty and L1 loss, combines spectrum normalization and a self-attention mechanism, stabilizes a training process and prevents gradient explosion, and meanwhile, generates a specific category of high-quality samples through condition information; and finally, carrying out accurate diagnosis by adopting a convolutional neural network equipped with a global attention mechanism. According to the method, the common collapse and instability problems of the generative adversarial network can be avoided, and the precise diagnosis of the rolling bearing can still be realized under the conditions of small samples and unbalanced data.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and device for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network. Background Art

[0002] Rolling bearings are core components of mechanical equipment, and their performance directly impacts the stable operation of industrial systems. In modern industry, due to long-term exposure to factors such as thermal fatigue, alternating loads, mechanical vibration, and wear, rolling bearing failure rates remain high, becoming a key bottleneck restricting the manufacture of high-end equipment. Authoritative statistics show that the failure rate of rolling bearings in the industrial sector is as high as 35%, with typical failures such as raceway spalling and cage fracture accounting for over 87%. These failures not only cause precision equipment to lose over 20% of its service life annually but can also pose systemic safety risks in key industries such as electricity and aviation. Therefore, research on rolling bearing fault diagnosis methods is crucial for ensuring the accuracy, reliability, and safety of mechanical equipment and extending its service life.

[0003] Deep learning is currently being widely used for bearing condition monitoring and fault diagnosis in rotating machinery. Traditional methods rely on expert experience to extract features. While deep models possess adaptive feature mining capabilities and can achieve high accuracy, their reliance on training data creates a new technical bottleneck. In actual engineering scenarios, the cost of collecting normal operating data accounts for only 1 / 10 of that of fault data, while the difficulty of acquiring fault samples increases exponentially. This extreme imbalance in data distribution results in a significant decrease in the fault identification accuracy of existing models in practical applications compared to laboratory environments. Consequently, it is difficult to collect samples of fault conditions, resulting in data imbalance and severely impacting the performance of deep learning-based mechanical fault diagnosis models. Summary of the Invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an unbalanced bearing data fault diagnosis method and device based on an improved conditional Wasserstein generative adversarial network. The original one-dimensional vibration signal is processed by sliding window overlapping sampling technology, and then converted into a two-dimensional time-frequency image using continuous wavelet transform. The generative adversarial network is then used to generate high-quality time-frequency pseudo images with clear texture details and color information to balance the fault data distribution, thereby improving the accuracy and stability of the fault diagnosis model.

[0005] In a first aspect, an embodiment of the present invention provides a method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network, comprising the following steps:

[0006] S1: Perform sliding window overlapping sampling on the original one-dimensional vibration signal, and use continuous wavelet transform to convert the processed signal into two-dimensional time-frequency image samples. Normal samples and fault samples are divided according to a certain ratio to obtain an unbalanced data set;

[0007] S2: Establish the LCWGAN-GP model;

[0008] S3: Use the unbalanced dataset to train the LCWGAN-GP model and generate time-frequency image samples with clear texture details and color information through the LCWGAN-GP model;

[0009] S4: Expand the unbalanced dataset to obtain a balanced dataset;

[0010] S5: Establish a GAM-CNN model, input the balanced training set time-frequency image samples into the GAM-CNN classifier to train the fault diagnosis model, and then input the test set into the trained GAM-CNN classifier to achieve accurate prediction of the bearing status.

[0011] In one embodiment, in S1, the sliding window overlapping sampling is intended to divide the time series data into multiple windows of fixed length, and there is a certain overlap between adjacent windows, thereby achieving continuous sampling of the data;

[0012] The continuous wavelet transform is used to extract the processed signal obtained by the sliding window overlapping sampling technique, which is to analyze the characteristics of the signal at different times and frequencies by convolving the signal with a set of mother wavelets of different scales and translations.

[0013] In one embodiment, the continuous wavelet transform specifically includes the following:

[0014] Assume that the original vibration image signal x(t)∈L 2 (R), x(t) represents the original vibration signal, L 2 (R) represents the space of square integrable functions;

[0015] The continuous wavelet transform is expressed as follows:

[0016]

[0017] Where x(t) is the original signal, ψ a,b (t) is the wavelet basis function, a and b are the scale parameter and translation parameter of the wavelet basis function respectively.

[0018] In one embodiment, in S2, establishing the LCWAN-GP model includes the following steps:

[0019] S2-1: Based on the Wasserstein generative adversarial network, category condition information is added to control the model to generate samples of specified categories;

[0020] S2-2: Introduce a gradient penalty term, use spectral normalization to stabilize training, add a self-attention mechanism to enhance the network learning ability, and add L1 loss to the generator loss of the LCWGAN-GP loss function.

[0021] In one embodiment, the LCWAN-GP model includes a generator module G and a discriminator module D. The generator module G accepts noise and class labels and generates false samples. The discriminator module D accepts both false samples and true samples and determines whether they are true or false.

[0022] The loss function of the LCWGAN-GP includes generator loss and discriminator loss;

[0023] The S2-2 specifically includes:

[0024] S2-2-1: The discriminator module introduces spectral normalization and gradient penalty to improve model training. The discriminator loss function is as follows:

[0025]

[0026] Where: L D represents the loss function of the LCWGAN-GP discriminator, Represents the mathematical expectation of the generated sample, z is the noise vector input to the generator, c is the category label corresponding to each noise vector, G(z|c) represents the sample generated by the generator, represents the mathematical expectation of the true sample, represents the gradient penalty term; λ2 represents the gradient penalty coefficient, It is the formula between the real sample x and the generated sample G(z) That is, the interpolation samples obtained by random interpolation sampling, where ε obeys the uniform distribution on [0,1];

[0027] S2-2-2: LCWGAN-GP enhances the generator module loss through the L1 loss module, which calculates the absolute difference between each pixel of the input image and the target image, making the generated image clearer and more realistic;

[0028] The generator loss function is:

[0029]

[0030]

[0031] Where: L GRepresents the loss function of the generator, L1 represents the L1 loss function, λ1 is the hyperparameter of the L1 loss, and λ1 is set to 100.

[0032] In one embodiment, in S4, expanding the bearing imbalance dataset to obtain a balanced dataset includes the following steps:

[0033] S4-1: The imbalanced dataset is fed into the proposed LCWAGN-GP model. Through adversarial training, the generator module G generates samples that are highly similar to the original data.

[0034] S4-2: The generated data samples and the original samples are then combined to form a balanced data set, where the number of normal samples and samples of each fault type in the balanced data is the same.

[0035] In one embodiment, in S5, GAM-CNN is a convolutional neural network equipped with a global attention mechanism, which is used to enable the classification model to focus on key parts and enhance attention to key features.

[0036] In one embodiment, the global attention mechanism first inputs the features into the channel attention mechanism, then weights the output and input and sends them into the spatial attention mechanism, and then weights the output and input;

[0037] The channel attention mechanism uses 3D permutation to preserve 3D information, and then uses a double-layer fully connected layer to amplify the spatial dependency of information across dimensions, multiplying it with the input feature map to strengthen the focus on important channels.

[0038] The spatial attention mechanism uses two convolutional layers for spatial fusion based on the channel attention mechanism, and then multiplies it with the feature map, so as to capture important features in three dimensions.

[0039] In a second aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the method for diagnosing unbalanced bearing data faults based on the Wasserstein generative adversarial network as described in the above embodiment are implemented.

[0040] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the unbalanced bearing data fault diagnosis method based on the Wasserstein generative adversarial network as described in the above embodiment are implemented.

[0041] The beneficial effects of the above technical solution are:

[0042] (1) This paper improves the conditional generative adversarial network model, uses the Wasserstein distance to measure distribution differences, introduces a gradient penalty term, applies gradient penalties in the interpolation region between data distributions, and constrains the gradient norm of the discriminator.

[0043] (2) The present invention equips the generator with an L1 loss module to enhance the image generation capability. In addition, a self-attention mechanism is embedded in the network to fully extract features and improve the authenticity of the generated samples. At the same time, spectral normalization is used in the discriminator network to stabilize model training.

[0044] (3) The classification model GAM-CNN proposed in this paper can effectively capture the key information in bearing fault data by combining the global attention mechanism and convolutional neural network, thereby effectively improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the overall framework of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network of the present invention;

[0046] Figure 2 This is a GAM structure diagram of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network of the present invention;

[0047] Figure 3 The present invention generates an adversarial network structure diagram for the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network;

[0048] Figure 4 Schematic diagram showing the comparison between the original sample and the production samples at different scales of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network of the present invention;

[0049] Figure 5 This is a flow chart of the method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network according to the present invention;

[0050] Figure 6 This is the confusion matrix and t-SNE visualization of the classification results of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network on CWRU different unbalanced data sets in the present invention. DETAILED DESCRIPTION

[0051] The aforementioned and other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of the embodiments with reference to the drawings of this application. The structural contents mentioned in the following embodiments are all referenced to the drawings in the specification.

[0052] Reference Figures 1-6 As shown, in a first aspect, an embodiment of the present invention provides a method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network, which specifically includes the following steps:

[0053] 1. Using sliding window overlapping sampling and continuous wavelet transform, the one-dimensional original vibration signal is converted into a two-dimensional time-frequency image sample through feature extraction

[0054] Sliding window overlapping sampling refers to setting a fixed-size window on the data sequence, and then moving the window according to a certain step size. After each movement, data samples are extracted from the window (that is, the data obtained by processing the original one-dimensional vibration data through the sliding window overlapping sampling technology, such as: the time domain and frequency domain characteristics of the one-dimensional vibration signal); during the movement process, there will be some data areas overlapping between two adjacent windows, thereby realizing continuous sampling of the data; the window size and overlap rate need to be determined. The overlap rate determines the step size of each sliding, which can usually be set as a part of the window length. For example, if the window length is L and the overlap rate is r, the step size is L×(1-r).

[0055] Continuous wavelet transform is a powerful time-frequency analysis method. Specifically, it analyzes the characteristics of a signal at different times and frequencies by convolving the signal with a set of mother wavelets of different scales and translations. Continuous wavelet transform performs well in analyzing non-stationary signals. Specifically, in high-frequency regions, continuous wavelet transform can provide fine time resolution, which helps to identify rapidly changing signal characteristics. In low-frequency regions, it exhibits better frequency resolution, which helps to accurately analyze the spectral characteristics of the signal, making it widely used in the field of bearing fault diagnosis.

[0056] Assume that the original vibration image signal x(t)∈L 2 (R), x(t) represents the original vibration signal, L 2 (R) represents the square integrable function space; the original vibration signal refers to the vibration signals generated by normal and faulty bearings in the public dataset. To simulate the difficulty in collecting faulty samples in actual industrial environments, the processed time-frequency image samples are divided according to a certain imbalance ratio to obtain an unbalanced bearing dataset.

[0057] The continuous wavelet transform is expressed as follows:

[0058]

[0059] Where x(t) is the original signal, ψ a,b (t) is the wavelet basis function, a and b are the scale parameter and translation parameter of the wavelet basis function respectively; Morlet wavelet is selected as the basis function of continuous wavelet transform; compared with other wavelet functions, Morlet wavelet basis function can better adapt to different types of vibration signals and extract more detailed features.

[0060] The specific process is as follows Figure 1 As shown in the data preprocessing part; taking the Case Western Reserve University bearing test bench as an example, the vibration signals contained in the test bench include normal categories and fault sizes of 0.007, 0.014, and 0.021 inches, respectively. Each size includes inner ring, outer ring and spherical faults. Therefore, plus the normal category, there are a total of 10 different state categories of signals (category condition information), which are represented as NC, IR007, IR014, IR021, BL007, BL014, BL021, OR007, OR014, and OR021, respectively. These categories are classified into different category labels (the corresponding labels are 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9, respectively). In the generator network, the category label is mapped to a high-dimensional embedding space through the class label, and after splicing with the noise, it is used as the input of the generator, thereby controlling the category of the image sample generated by the generator.

[0061] Specifically, the vibration sensor installed on the test bench is first used to collect the fault vibration signal and its characteristics, and then the one-dimensional vibration signal is converted into a two-dimensional time-frequency image using the above-mentioned continuous wavelet transform method after sliding window overlapping sampling.

[0062] 2. Using the LCKANG-GP model to generate time-frequency image samples with clear texture details and color information to expand and balance the bearing imbalance dataset

[0063] WGAN-GP is an improved model of GAN, specifically an unsupervised generative model proposed by Ian Goodfellow et al. in 2014. It trains the model through a game process, enabling the model to generate high-quality fake samples that are infinitely close to real data. The GAN network structure is as follows: Figure 3 As shown in the figure; specifically, first a random noise vector z is generated, which is the input of the generator module to generate fake samples; then, the discriminator module receives the input of the real sample and the fake sample generated by the generator module, and outputs the probability that the sample is a real sample; the training goal is to make the fake samples generated by the generator module as similar as possible to the real samples, while making it difficult for the discriminator module to distinguish whether the input sample is a real sample or a forged fake sample.

[0064] The objective function of GAN can be expressed by the following formula:

[0065]

[0066] Among them, D(x) is the output of the discriminator module for the real data x, G(z) is the sample generated by the generator module for the noise z, and P r is the data distribution, P z is the noise distribution.

[0067] However, during training, GAN uses JS divergence to measure the difference between the distribution of real samples and generated samples, which may lead to gradient vanishing and mode collapse. Therefore, to improve the stability of GAN training, WGAN-GP uses Wasserstein distance to measure the distribution difference and adds a gradient penalty term to the objective function to limit the gradient of the discriminator module, thereby further stabilizing the training process. The objective function of WGAN-GP can be expressed as:

[0068]

[0069] in, It is an interpolation sample between the real sample and the generated sample; is the discriminator module pair The gradient of is a hyperparameter that controls the intensity of the gradient penalty; this penalty term keeps the gradient of the discriminator close to 1, thus avoiding instability during training.

[0070] The LCWAN-GP mentioned in this article is a further improvement on this basis.

[0071] The loss function of LCWGAN-GP includes the generator module loss and the discriminator module loss. The discriminator module introduces spectral normalization and gradient penalty to improve model training; the discriminator loss function is as follows:

[0072]

[0073]

[0074] Where: L D represents the loss function of the LCWGAN-GP discriminator module, represents the mathematical expectation of the generated samples, z is the noise vector input to the generator module, c is the category label corresponding to each noise vector, G(z|c) represents the sample generated by the generator module, represents the mathematical expectation of the true sample, represents the gradient penalty term;

[0075] λ2 represents the gradient penalty coefficient, It is the formula between the real sample x and the generated sample G(z) That is, the interpolation samples obtained by random interpolation sampling, where ε obeys the uniform distribution on [0,1];

[0076] LCWGAN-GP enhances the generator module loss through the L1 loss module, which calculates the absolute difference between each pixel of the input image and the target image, making the generated image clearer and more realistic;

[0077] The generator module loss function is:

[0078]

[0079]

[0080] Where: L G Represents the loss function of the generator module, L1 represents the L1 loss function, λ1 is the hyperparameter of L1 loss, and λ1 is set to 100.

[0081] Under different imbalance ratios, LCWAGN-GP generates time-frequency image samples with clear texture details and color information. Figure 4 As shown in the figure, the bearing imbalance dataset is expanded to obtain a balanced dataset. Specifically, the imbalanced dataset is input into the proposed LCWGAN-GP model. Through the adversarial training of the GAN model, the generator module generates samples that are highly similar to the original data. The generated data samples are then mixed with the original samples so that the number of normal samples and samples of each fault type are the same.

[0082] Specifically, the generator module G input is a 100-dimensional Gaussian random noise vector z and a category label c, which are used to generate 64×64 fake image samples; first, a fully connected layer is used to convert the concatenated noise and label vectors into a higher-dimensional feature map; then, four layers of transposed convolution are performed, and batch normalization and ReLU activation functions are applied after the first three layers of transposed convolution to speed up training and prevent overfitting. A self-attention mechanism is introduced between the second and third layers of transposed convolution in the model; finally, the Tanh activation function is used to limit the output range.

[0083] The discriminator module D takes as input real time-frequency image samples, fake samples generated by the generator module G, and a class label c, and outputs a scalar. The input class label is first mapped to a high-dimensional embedding space, which is then expanded and concatenated with the image sample as the input to the subsequent sub-network. The core of the discriminator module D consists of four layers of convolution and the LeakyReLU function, enabling it to more accurately distinguish sample types. Spectral normalization is also used to normalize the weights of each convolution module in the network. A self-attention mechanism is added between the second and third convolution layers to improve their resolution. The sixth sub-network consists of fully connected layers, mapping the extracted high-dimensional features to a scalar output.

[0084] The detailed parameters of the LCWANG-GP network structure are shown in Table 1:

[0085] Table 1 Network parameters of the generator and discriminator

[0086]

[0087]

[0088] 3. Input the balanced training set time-frequency samples into the GAM-CNN classifier to train the fault diagnosis model, and then input the test set into the trained GAM-CNN classifier to achieve accurate prediction of the bearing status.

[0089] The GAM-CNN model is used for fault diagnosis. Specifically, the balanced training set time-frequency samples are input into the GAM-CNN classifier to train the fault diagnosis model, and the test set is then input into the trained GAM-CNN classifier to achieve accurate prediction of the bearing status. The detailed parameters of the GAM-CNN network structure are shown in Table 2:

[0090] Table 2 Network parameters of GAM-CNN

[0091] layer Convolution kernel step length Feature Map size Activation Function GAM-CNN — — — — — convolution 3×3 1×1 32 64×64 ReLU GAM — — 32 64×64 — Max Pooling 2×2 2×2 32 32×32 — convolution 3×3 1×1 64 32×32 ReLU GAM — — 64 32×32 — Max Pooling 2×2 2×2 64 16×16 — convolution 3×3 1×1 128 16×16 ReLU GAM — — 128 16×16 — Max Pooling 2×2 2×2 128 8×8 — AdaptiveAvgPool — — 128 1×1 — Fully connected layer — — 1 — —

[0092] The GAM-CNN classifier is a convolutional neural network equipped with a global attention mechanism. Specifically, the classification model first performs a convolution operation on the input image to extract features of different scales and directions. Then, the GAM (global attention) module performs global information modeling on the feature map to enhance the expressiveness of the features. After three repetitions, the adaptive average pooling layer further reduces the feature dimension and retains the globally important feature information. The fully connected layer classifies the features after the adaptive average pooling, maps the 128-dimensional features to a 1-dimensional output, and obtains the classification result; the GAM-CNN structure is as follows: Figure 1 As shown in the fault diagnosis section, the GAM structure is as follows Figure 2 shown.

[0093] In the second aspect, an embodiment of the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, the various processes of the above-mentioned embodiment of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0094] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the various processes of the above-mentioned embodiment of the unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0095] The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0096] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] IV. Experimental Verification

[0098] A preferred embodiment of the present invention is an unbalanced bearing data fault diagnosis method based on an improved conditional Wasserstein generative adversarial network, which is applied to the Case Western Reserve University Bearing Fault Diagnosis Dataset (CWRU Bearing DataCenter).

[0099] CWRU is a widely recognized dataset in the field of bearing faults and is a landmark benchmark dataset. The experimental platform consists of equipment such as motors, torque sensors, and decoders. The bearing operating speed is 1797 r / min and the motor load is 1.5 kW. In order to collect vibration signals of different fault types, a vibration accelerometer with a sampling frequency of 12 kHz is used. These fault categories include normal state (NC), inner raceway (IR), outer raceway (OR), and ball fault (BL). Three different damage levels are included in each fault type, namely 0.007, 0.014, and 0.021 inches. A total of ten types of bearing fault samples were collected, marked as 0-9, and represented as NC, IR007, IR014, IR021, BL007, BL014, BL021, OR007, OR014, and OR021.

[0100] In this experiment, each bearing data sample was 1024 long, and the original vibration signal samples were enhanced using a sliding window method with a sliding window interval of 512. A total of 235 faulty samples and 475 normal samples were obtained for each class. To simulate the insufficient fault data in actual rolling bearing fault diagnosis, the number of training sets was limited, while the number of test sets remained unchanged to avoid compromising diagnostic accuracy. Specifically, a balanced dataset consisting of 150 randomly selected samples from each class was used for fault diagnosis testing, with 80 samples used for fault diagnosis testing. Subsequently, three unbalanced datasets (AC) were generated, with imbalance ratios of 1:2, 1:5, and 1:10, respectively. The imbalance ratio is defined as the ratio of the number of normal samples to the number of samples in each fault class. More information about the datasets is shown in Table 3.

[0101] Table 3 Description of CWRU experimental data

[0102]

[0103] The generated time-frequency image will affect the fault diagnosis effect; specifically, in the case of unbalanced data, if the time-frequency samples provided by the generated image can accurately reflect the data distribution provided by the original time-frequency image, it will be beneficial for the fault diagnosis model to grasp the characteristics of the data.

[0104] The experiment quantitatively analyzes the generation of high-quality samples and uses the SSIM and FID metrics to evaluate the feature connections between generated samples and real samples. One hundred generated samples and real samples are selected for SSIM evaluation, and the average of the comparison results is taken to obtain the final results. The FID metric is used to evaluate the one hundred generated samples and one hundred real samples. In addition to the proposed method, the proposed model is also compared with the generative adversarial networks (CGAN) and CDCGAN. The comparison results are shown in Table 4.

[0105] Table 4 Evaluation indicators of original samples and generated samples

[0106]

[0107] The data in the table shows that, given the same imbalanced dataset, LCWGAN-GP is able to generate new, high-quality samples, and both the SSIM and FID metrics are improved compared to CGAN and CDCGAN. To further investigate the quality of generated samples, we quantitatively evaluate the quality of samples generated from different imbalanced datasets; the results are shown in Table 5.

[0108] Table 5 Evaluation indicators of original samples and generated samples at different scales

[0109]

[0110] The results show that the samples generated by each dataset have high authenticity and can contribute to the balanced dataset. In addition, the three fault categories in the table above were selected from each dataset and the generated images were compared with the original images. The image comparison results are shown in Figure 2. Figure 4 shown.

[0111] from Figure 4 It can be clearly seen that the images generated by LCWGAN-GP under three different balance ratios retain key features such as texture, clarity, and spectral brightness, while also showing satisfactory diversity.

[0112] After expanding the three imbalanced datasets A, B, and C, the GAM-CNN classifier is retrained using the balanced datasets. After training, the performance of the classifier is tested using the test dataset, thereby completing the fault diagnosis of the corresponding imbalanced datasets. The confusion matrix and t-SNE visualization of the classification results of the test set under the three balanced datasets are shown below. Figure 5 As shown, Figure 5 The confusion matrix in Figure 1 intuitively reflects the accuracy of fault classification. The horizontal and vertical coordinates of the confusion matrix represent the model classification results and the actual labels of the fault categories, respectively.

[0113] from Figure 5 It can be clearly seen that when the imbalance ratio is 2:1, except for two samples of class OR014 that are misclassified, all other fault classes can be correctly classified. When the imbalance ratio is 10:1, the diagnostic accuracy of four categories reaches 100%, and only a small number of samples of the remaining fault classes are misclassified. This shows that the proposed method has the ability to generate high-quality samples and expand the imbalanced dataset, effectively improving the accuracy of fault diagnosis.

[0114] The above description is only for illustrating the present invention. It should be understood that the present invention is not limited to the above embodiments, and various variations that conform to the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of unbalanced bearing data based on an improved conditional Wasserstein generative adversarial network, comprising the following steps: S1: Perform sliding window overlapping sampling on the original one-dimensional vibration signal, and use continuous wavelet transform to convert the processed signal into two-dimensional time-frequency image samples. Normal samples and fault samples are divided according to a certain ratio to obtain an unbalanced data set; S2: Establish the LCWGAN-GP model; S3: Use the unbalanced dataset to train the LCWGAN-GP model and generate time-frequency image samples with clear texture details and color information through the LCWGAN-GP model; S4: Expand the unbalanced dataset to obtain a balanced dataset; S5: Establish a GAM-CNN model, input the balanced dataset time-frequency image samples into the GAM-CNN classifier to train the fault diagnosis model, and then input the test set into the trained GAM-CNN classifier to achieve accurate prediction of the bearing status.

2. The unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network according to claim 1 is characterized in that: In S1, the sliding window overlapping sampling is intended to divide the time series data into multiple windows of fixed length, and there is a certain overlap between adjacent windows, so as to achieve continuous sampling of the data; The continuous wavelet transform is used to extract the processed signal obtained by the sliding window overlapping sampling technique, which is to analyze the characteristics of the signal at different times and frequencies by convolving the signal with a set of mother wavelets of different scales and translations.

3. The unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network according to claim 2 is characterized in that: The continuous wavelet transform specifically includes the following: Assume that the original vibration image signal x(t)∈L 2 (R), x(t) represents the original vibration signal, L 2 (R) represents the space of square integrable functions; The continuous wavelet transform is expressed as follows: Where x(t) is the original signal, ψ a,b (t) is the wavelet basis function, a and b are the scale parameter and translation parameter of the wavelet basis function respectively.

4. The unbalanced bearing data fault diagnosis method based on the improved conditional Wasserstein generative adversarial network according to claim 2, characterized in that: In S2, establishing the LCWAN-GP model includes the following steps: S2-1: Based on the Wasserstein generative adversarial network, category condition information is added to control the model to generate samples of specified categories; S2-2: Introduce a gradient penalty term, use spectral normalization to stabilize training, add a self-attention mechanism to enhance the network learning ability, and add L1 loss to the generator loss of the LCWGAN-GP loss function.

5. The method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network according to claim 4, characterized in that: The LCWAN-GP model includes a generator module and a discriminator module. The generator module accepts noise and class labels and generates false samples. The discriminator module accepts false samples and real samples and judges whether they are true or false. The loss function of the LCWGAN-GP includes generator loss and discriminator loss; The S2-2 specifically includes: S2-2-1: The discriminator module introduces spectral normalization and gradient penalty to improve model training. The discriminator loss function is as follows: Where: L D represents the loss function of the LCWGAN-GP discriminator, Represents the mathematical expectation of the generated sample, z is the noise vector input to the generator G, c is the category label corresponding to each noise vector, G(z|c) represents the sample generated by the generator, represents the mathematical expectation of the true sample, represents the gradient penalty term; λ2 represents the gradient penalty coefficient, It is the formula between the real sample x and the generated sample G(z) That is, the interpolation samples obtained by random interpolation sampling, where ε obeys the uniform distribution on [0,1]; S2-2-2: LCWGAN-GP enhances the generator loss through the L1 loss module, which calculates the absolute difference between each pixel of the input image and the target image, making the generated image clearer and more realistic; The generator loss function is: Where: L G Represents the loss function of the generator, L1 represents the L1 loss function, λ1 is the hyperparameter of the L1 loss, and λ1 is set to 100.

6. The method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network according to claim 5, characterized in that: In S4, the bearing imbalance data set is expanded to obtain a balanced data set, including the following steps: S4-1: The imbalanced dataset is fed into the proposed LCWAGN-GP model. Through adversarial training, the generator module generates samples that are highly similar to the original data. S4-2: The generated data samples and the original samples are then combined to form a balanced data set, where the number of normal samples and samples of each fault type in the balanced data is the same.

7. The method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network according to claim 1, wherein: In the S5, GAM-CNN is a convolutional neural network equipped with a global attention mechanism, which is used to enable the classification model to focus on key parts and enhance attention to key features.

8. The method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network according to claim 7, characterized in that: The global attention mechanism first inputs the features into the channel attention mechanism, then weights the output and input and sends them into the spatial attention mechanism, and then weights the output and input; The channel attention mechanism uses 3D permutation to preserve 3D information, and then uses a double-layer fully connected layer to amplify the spatial dependency of information across dimensions, multiplying it with the input feature map to strengthen the focus on important channels. The spatial attention mechanism uses two convolutional layers for spatial fusion based on the channel attention mechanism, and then multiplies it with the feature map, so as to capture important features in three dimensions.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the method for diagnosing unbalanced bearing data faults based on an improved conditional Wasserstein generative adversarial network as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the unbalanced bearing data fault diagnosis method based on an improved conditional Wasserstein generative adversarial network as described in any one of claims 1 to 8.

Images