Rolling bearing fault diagnosis method based on improved generative adversarial network

Abstract

The invention provides a rolling bearing fault diagnosis method based on an improved generative adversarial network, and belongs to the technical field of bearing fault diagnosis. A traditional generative adversarial network generally achieves a good effect after development for many years, and causes such as great difficulty, fault data scarcity and tedious and complex marking work exist in a rolling bearing fault signal collection process, so that application of the rolling bearing fault signal collection method in a real scene is limited, and some difficulties are brought to fault diagnosis work. Therefore, the rolling bearing fault diagnosis method based on the improved generative adversarial network is provided in order to reduce the dependence of the diagnosis process on a large number of marked samples. According to the method, one-dimensional vibration signal data obtained through sampling is converted into a two-dimensional time-frequency image to serve as network input, the improved generative adversarial network (IGAN) is trained for data enhancement, and the diagnosis precision of different faults under the condition of few label samples is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing fault diagnosis, and specifically relates to a rolling bearing fault diagnosis method based on an improved generative adversarial network under the condition of few-label samples, which includes a feature extractor and two generators and a discriminator composed of a residual module, a CBAM attention mechanism, and a spectral normalization module. Background Art

[0002] Rotating machinery is widely used in industry. As an important component in rotating machinery, rolling bearings play a crucial role in the normal operation of rotating machinery. According to statistical data, the failures of rotating machinery and electrical systems caused by rolling bearing failures account for a relatively high proportion. Therefore, accurate fault diagnosis of rolling bearings is crucial for preventing catastrophic accidents. In recent years, fault diagnosis, as an effective method to ensure the safe operation of mechanical systems, has received more and more attention and research.

[0003] The monitoring and fault diagnosis of rolling bearings is an important step for the industry to move towards intelligence and automation. However, obtaining data usually requires high costs and a long acquisition process. Especially under multiple restrictions in terms of equipment, time, and environment, the difficulty of labeling fault images is even more prominent. Therefore, how to effectively expand the training data set under the condition of insufficient data volume and few-label samples, effectively extract data features, and make accurate discrimination has become the key to solving the problem.

[0004] The emergence of generative adversarial networks has attracted the attention of many experts and scholars, and its ability in data augmentation cannot be underestimated. Therefore, it is a feasible method to improve on the basis of the original generative adversarial network and use the improved generative adversarial network to solve the above problems.

[0005] After retrieval, the application publication number is CN114858455A, a rolling bearing fault diagnosis method and system based on an improved GAN-OSNet, which relates to the technical field of bearing fault diagnosis. The present invention first improves the existing generative adversarial network for data augmentation, then improves the full-scale network, and uses the augmented data to train and optimize the improved full-scale network, so as to obtain the optimal fault diagnosis model. Finally, the optimal fault diagnosis model can be used for the fault diagnosis of rolling bearings. The present invention solves the problems of low accuracy, high cost, and low diagnostic efficiency existing in the existing bearing fault diagnosis technology when the samples are unbalanced.

[0006] In the above patent, although it can solve the problems of low accuracy, high cost and low diagnostic efficiency when the samples are unbalanced, it fails to fully solve the problem of insufficient labels. Therefore, in this patent, the problem of few-label samples is mainly improved. After the improvement, high-precision rolling bearing fault diagnosis can be achieved under the condition of few sample labels. The CBAM attention mechanism and residual module are used to fully improve the feature extraction and recognition capabilities of the generator and discriminator. Due to the skip link effect of the residual module, the speed of information circulation in the network is accelerated; the discriminator loss is further controlled through spectral normalization technology; the AdaptiveMix loss and L1 regularization term are used in network training to improve the quality of generated images. Summary of the Invention

[0007] The present invention aims to solve the above problems in the prior art. It proposes a rolling bearing fault diagnosis method based on an improved generative adversarial network. The technical solution of the present invention is as follows:

[0008] A rolling bearing fault diagnosis method based on an improved generative adversarial network comprises the following steps:

[0009] Step S1: Using sensors to collect rolling bearing fault vibration signals under different fault states to obtain signal samples; segmenting the signal samples piece by piece, and using continuous wavelet transform (CWT) to perform time-frequency image conversion on the segmented one-dimensional vibration signals to convert them into three-channel wavelet scaling images;

[0010] Step S2: Build an improved generative adversarial network, which includes a feature extractor, two generators based on a residual module, a CBAM attention mechanism, and a spectral normalization module, and a discriminator;

[0011] Step S3: The converted three-channel wavelet scalogram is used as the input of the improved generative adversarial network, where the input of the generator is the feature information output by the feature extractor, and the input of the discriminator is the original wavelet scalogram and the image generated by the generator as well as a small amount of label information. During the training process, the AdaptiveMix mechanism is applied to gradually reduce the losses of the generator and discriminator until a Nash equilibrium is reached.

[0012] Step S4: After the training of the improved generative adversarial network as a fault diagnosis model is completed, the network effect is verified using the test set samples.

[0013] Furthermore, the step S1 uses a sensor to collect rolling bearing fault vibration signals under different fault states to obtain signal samples; segmenting the signal samples piece by piece, and using continuous wavelet transform (CWT) to perform time-frequency image conversion on the segmented one-dimensional vibration signal to convert it into a three-channel wavelet scaling map, specifically including the following steps:

[0014] Step S1.1: The sampling process is to segment the original vibration signal with a length of N into groups of every m data points, with an overlapping length of z;

[0015] Step S1.2: Use cmor3-3 wavelet as the mother wavelet to decompose the segmented original vibration signals in number of groups, and extract the corresponding wavelet scale maps;

[0016] Step S1.3: Divide the data set into a training set and a test set according to the needs of network training for network training, and the division ratio of the training set to the test set is 3:1.

[0017] Further, in the step S2, the specific structure of the improved generative adversarial network composed of a feature extractor, two generators and a discriminator based on residual modules, CBAM attention mechanism and spectral normalization module is as follows:

[0018] In the feature extractor, its input is the original image. Use the pre-trained ResNet50 network as the baseline network for feature extraction, remove the classification layer, process the output of the network through global average pooling and fully connected layers, and apply the ReLU activation function. Finally, compress the feature map into a 128-dimensional feature representation, and the output feature representation is used as the input of the generator;

[0019] In the generator, it contains four groups of residual structures. Each group of residual structures consists of two convolutional layers, two normalization layers and one ReLU activation function layer. The convolutional kernels of each convolutional layer are 3×3 convolutional kernels; after being processed by four groups of residual structures, it is input into the CBAM attention mechanism for processing;

[0020] In the discriminator, it contains two groups of residual structures. Similar to the generator, each group of residual structures consists of two convolutional layers, two normalization layers and one ReLU activation function layer. The convolutional kernels of each convolutional layer are 3×3 convolutional kernels; after being processed by two groups of residual structures, it is input into the CBAM attention mechanism for processing; the outputs of the discriminator are category judgment and true / false judgment respectively.

[0021] Further, in the step S3, the loss function of the improved generative adversarial network model is as follows:

[0022] Input the wavelet scale map obtained by continuous wavelet transform into the improved generative adversarial network for training to generate high-quality generated images; among them, the discriminator training involves two modes: supervised learning and unsupervised learning. The loss function of the discriminator consists of supervised learning loss L1 and unsupervised learning L2 loss, as shown in formulas (1), (2) and (3):

[0023] L D = L1 + L2 (1)

[0024]

[0025] where x represents the input, y represents its corresponding label, and P r(x,y) represents the data distribution of the original image, and P g represents the data distribution of the mixed generated image g mix where g mix = α·g1 + (1 - α)·g2, where g1 and g2 are images generated by two generators respectively, and α is a dynamic weight;

[0026] In the supervised learning part, Softmax is used as the activation function of the discriminator, and the categorical cross-entropy loss L1 is used; in the unsupervised learning part, the discriminator is the same as that of the ordinary GAN. The only thing to judge is the authenticity of the output samples. The activation function used is Sigmoid, and the loss is L2. The generator and the discriminator can be trained by minimizing the loss function. The loss function of the generator is shown in formula (4):

[0027]

[0028] where represents the data distribution of the original image, and represent the data distributions of the images generated by two generators, and g mix is the mixed generated image, where g mix = α·g1 + (1 - α)·g2, where g1 and g2 are images generated by two generators respectively, α is a dynamic weight, λ is a regularization parameter, and ‖θ‖1 represents the L1 norm of the model parameter θ. λ‖θ‖1 represents the L1 loss;

[0029] The introduction of the AdaptiveMix module proposes a soft loss, as shown in formula (5):

[0030]

[0031] where x i and x j represent the i-th and j-th training images respectively, represents the hard sample generated by x i and x j σ is a noise term sampled from a Gaussian distribution to prevent overfitting, and D v (·,·) refers to the metric used to evaluate the distance, such as the L1 norm and the L2 norm. λ is a hyperparameter sampled from a Beta distribution λ ∈ Beta(α, α).

[0032] Therefore, the total loss function of the improved GAN after applying the AdaptiveMix module is as shown in Equation (6):

[0033] L IGAN = L D + L G + L Ada (6).

[0034] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the rolling bearing fault diagnosis method based on the improved generative adversarial network as described in any one of the above items.

[0035] A non-transitory computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the rolling bearing fault diagnosis method based on the improved generative adversarial network as described in any one of the above items.

[0036] The advantages and beneficial effects of the present invention are as follows:

[0037] 1) The present invention integrates a spectral normalization module, a CBAM attention mechanism, and a residual module, applies a simple convolutional structure, enhances the network's feature extraction and generated feature capabilities, and improves the accuracy of classification and recognition. Among them, the spectral normalization module is used to alleviate the common gradient disappearance and mode collapse problems in traditional generative adversarial networks; the CBAM attention mechanism makes the model more focused on the key areas of fault signals, enhancing the model's stability and sensitivity to fault details; the residual module improves the model's feature extraction ability and better captures complex patterns in bearing fault data. 2) During the training process, the AdaptiveMix loss and the L1 regularization term are used. By introducing new loss functions and regularization terms, the training process of the generative adversarial network is further optimized, and the quality of the generated samples and the robustness of the model are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the method for improving the generative adversarial network provided by the present invention;

[0039] Figure 2 is a schematic structural diagram of the feature extractor in the improved generative adversarial network;

[0040] Figure 3 is a schematic structural diagram of the generator network in the improved generative adversarial network;

[0041] Figure 4 is a schematic structural diagram of the discriminator network in the improved generative adversarial network;

[0042] Figure 5It is the discriminator classification confusion matrix diagram; Specific implementation manner

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0044] The technical solution for the present invention to solve the above technical problems is:

[0045] See Figure 1 , the flowchart of the research on the rolling bearing fault diagnosis method based on the improved generative adversarial network proposed by the present invention. The specific steps of this embodiment are as follows:

[0046] Step S1: Use sensors to collect the rolling bearing fault vibration signals under different fault states to obtain signal samples. Segment the signal samples one by one, and use continuous wavelet transform (CWT) to perform time-frequency image conversion on the segmented one-dimensional vibration signals, converting them into three-channel wavelet scale diagrams.

[0047] Step S1.1: The sampling process is to segment the original vibration signal with a length of N into groups of every m data points, and the overlapping length is z.

[0048] Table 1 Experimental data representation method

[0049]

[0050] Step S1.2: Use cmor3-3 wavelet as the mother wavelet to decompose the segmented number of groups of original vibration signals, and extract the corresponding wavelet scale diagrams. The experimental data used in this embodiment is the SKF6205 deep groove ball bearing dataset in the rolling bearing fault vibration signals opened by Case Western Reserve University (CWRU) in the United States. In the experiment, an acceleration sensor installed at the motor drive end is used to collect vibration signals. Three types of faults are introduced into the bearing through electrical discharge machining: outer race fault (ORF), rolling element fault (BF), and inner race fault (IRF). The fault degree of each fault type is different (fault diameter: 0.007 inches, 0.014 inches, and 0.021 inches). Select the vibration signals under three motor speed conditions (1772 rpm, 1750 rpm, and 1730 rpm) at a working load of 0 HP as experimental data. The experimental bearing model is SKF6205. The data sampling frequency is 12000 Hz. The experimental data representation method is shown in Table 1.

[0051] Step S1.3: Divide the dataset into a training set and a test set according to the network training needs for network training, and the division ratio of the training set and the test set is 3:1.

[0052] Step S2: Build an improved generative adversarial network, which includes a feature extractor and two generators and a discriminator composed of residual modules, CBAM attention mechanism, and spectral normalization module.

[0053] In the feature extractor, the input is the original image. The pre-trained ResNet50 network is used as the baseline network for feature extraction. The classification layer is removed, and the output of the network is processed through global average pooling and a fully connected layer, and the ReLU activation function is applied. Finally, the feature map is compressed into a 128-dimensional feature representation, and the output feature representation is used as the input of the generator.

[0054] In the generator, there are four groups of residual structures. Each group of residual structures consists of two convolutional layers, two normalization layers, and one ReLU activation function layer. The convolutional kernel of each convolutional layer is a 3×3 convolutional kernel. After being processed by four groups of residual structures, it is input into the CBAM attention mechanism for processing. The specific parameters of the generator network are shown in Table 2.

[0055] Table 2 Improved Generative Adversarial Network - Generator Architecture

[0056]

[0057]

[0058] In the discriminator, there are two groups of residual structures. Similar to the generator, each group of residual structures consists of two convolutional layers, two normalization layers, and one ReLU activation function layer. The convolutional kernel of each convolutional layer is a 3×3 convolutional kernel. After being processed by two groups of residual structures, it is input into the CBAM attention mechanism for processing. The outputs of the discriminator are class judgment and true / false judgment respectively. The specific parameters of the discriminator network are shown in Table 3.

[0059] Table 3 Improved Generative Adversarial Network - Discriminator Architecture

[0060]

[0061] Step S3: Use the transformed three-channel wavelet scale map as the input of the improved generative adversarial network. Among them, the input of the generator is the feature information output by the feature extractor, and the input of the discriminator is the original wavelet scale map, the image generated by the generator, and a small part of the label information. During the training process, the AdaptiveMix mechanism is applied to gradually reduce the losses of the generator and the discriminator until the Nash equilibrium is reached.

[0062] Among them, the discriminator training involves two modes: supervised learning and unsupervised learning. The loss function of the discriminator consists of the supervised learning loss L1 and the unsupervised learning loss L2, as shown in formulas (1), (2), and (3):

[0063] L D = L1 + L2 (1)

[0064]

[0065] In the formula, x represents the input, y represents its corresponding label, P r(x,y) represents the data distribution of the original image, P g represents the data distribution of the mixed generated image g mix , where g mix = α·g1 + (1 - α)·g2, g1 and g2 are the images generated by two generators respectively, and α is the dynamic weight.

[0066] In the supervised learning part, Softmax is used as the activation function of the discriminator, and the categorical cross-entropy loss L1 is used. In the unsupervised learning part, the discriminator is the same as the ordinary GAN. The only thing to judge is the authenticity of the output samples. The activation function used is Sigmoid, and the loss is L2. The generator and the discriminator can be trained by minimizing the loss function. The loss function of the generator is shown in formula (4):

[0067]

[0068] In the formula, represents the data distribution of the original image, and represent the data distributions of the images generated by two generators, g mix is the mixed generated image,, where g mix = α·g1 + (1 - α)·g2, g1 and g2 are the images generated by two generators respectively, α is the dynamic weight, λ is the regularization parameter, ‖θ‖1 represents the L1 norm of the model parameter θ, and λ‖θ‖1 represents the L1 loss;

[0069] The introduction of the AdaptiveMix module helps the generator better match the data distribution, which enables the AdaptiveMix method to generate higher-quality mixed images, thereby improving the performance of the GAN. Here, a soft loss is proposed, as shown in formula (5),

[0070]

[0071] In the formula, x i and x j represent the i-th and j-th training images respectively, Represents x i and x j The generated difficult samples, σ is the noise term sampled from Gaussian distribution to prevent overfitting, D v (·,·) refers to the metric used to evaluate the distance, such as the L1 norm, L2 norm, and λ is a hyperparameter sampled from the Beta distribution λ∈Beta(α,α).

[0072] Therefore, the total loss function of the improved GAN after applying the AdaptiveMix module is as follows:

[0073] L IGAN =L D +L G +L Ada (6)

[0074] Step S4: After the training of the improved generative adversarial network as a fault diagnosis model is completed, the network effect is verified using the test set samples.

[0075] To verify the effectiveness of the above method, the present invention is verified on the Case Western Reserve University (CWRU) data set. The data used in the experiment are vibration data under 0HP working conditions. The data used in the experiment include normal operation signals and 9 kinds of fault signals. Figure 2 、 Figure 3 and Figure 4 The feature extractor, generator, and discriminator shown in this embodiment used the training set wavelet scalograms and partial labels as inputs to the improved generative adversarial network for training. The network was then tested using the test set wavelet scalograms and partial labels under three label ratios: 1%, 5%, and 10%. Ten experiments were conducted for each ratio, and the results are shown in Table 4. The rolling bearing fault diagnosis method provided in this embodiment achieved an average diagnostic accuracy of 99.54% with 10% labeled samples. Figure 5 This is the optimal confusion matrix diagram after the discriminator classification. It can be seen from the figure that the discriminator has a strong classification ability, which verifies the effectiveness of the proposed method.

[0076] Table 4. Fault diagnosis experimental results under different label ratios

[0077]

[0078] The improved generative adversarial network in this paper combines the CBAM attention mechanism, spectral normalization module and residual module, and applies the AdaptiveMix mechanism. It can accurately distinguish in the case of few labels, improve the accuracy of system fault diagnosis, and has broad application prospects. The above describes the preferred specific embodiments of the present invention in detail.

[0079] The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions.

[0080] Computer-readable media includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM (compact disc read-only memory), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0081] It should also be noted that the term "comprises", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0082] The above embodiments should be understood as being only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the content recorded in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on an improved generative adversarial network, characterized in that: The following steps are involved: Step S1: Using sensors to collect rolling bearing fault vibration signals under different fault states to obtain signal samples; segmenting the signal samples piece by piece, and using continuous wavelet transform (CWT) to perform time-frequency image conversion on the segmented one-dimensional vibration signals to convert them into three-channel wavelet scaling images; Step S2: Build an improved generative adversarial network, which includes a feature extractor, two generators based on a residual module, a CBAM attention mechanism, and a spectral normalization module, and a discriminator; Step S3: The converted three-channel wavelet scalogram is used as the input of the improved generative adversarial network. The input of the generator is the feature information output by the feature extractor, and the input of the discriminator is the original wavelet scalogram and the image generated by the generator, as well as a small amount of label information. During the training process, the AdaptiveMix mechanism is applied to gradually reduce the loss of the generator and discriminator until a Nash equilibrium is reached. Step S4: After the training of the improved generative adversarial network as a fault diagnosis model is completed, the network effect is verified using the test set samples.

2. The rolling bearing fault diagnosis method based on improved generative adversarial network according to claim 1 is characterized in that: The step S1 uses a sensor to collect rolling bearing fault vibration signals under different fault states to obtain signal samples; segmenting the signal samples piece by piece, and using continuous wavelet transform (CWT) to perform time-frequency image conversion on the segmented one-dimensional vibration signals to convert them into three-channel wavelet scaling images, specifically including the following steps: Step S1.1: The sampling process is to segment the original vibration signal of length N into groups of m data points each, with an overlapping length of z; Step S1.2: Use cmor3-3 wavelet as the mother wavelet, and the number of segments is Decompose the original vibration signal and extract the corresponding wavelet scaling map; Step S1.3: Divide the data set into a training set and a test set according to the network training requirements for network training. The ratio of the training set to the test set is 3:

1.

3. The rolling bearing fault diagnosis method based on improved generative adversarial network according to claim 1 is characterized in that: In step S2, the specific structure of the improved generative adversarial network consisting of a feature extractor, two generators based on a residual module, a CBAM attention mechanism, and a spectral normalization module, and a discriminator is as follows: In the feature extractor, the input is the original image. The pre-trained ResNet50 network is used as the baseline network for feature extraction. The classification layer is removed, and the network output is processed through global average pooling and fully connected layers. The ReLU activation function is applied, and finally the feature map is compressed into a 128-dimensional feature representation. The output feature representation is used as the input of the generator. The generator contains four groups of residual structures, each of which consists of two convolutional layers, two normalization layers, and one ReLU activation function layer. The convolution kernel of each convolutional layer is a 3×3 convolution kernel. After being processed by the four groups of residual structures, the input is input to the CBAM attention mechanism for processing. The discriminator contains two sets of residual structures. Like the generator, each residual structure consists of two convolutional layers, two normalization layers, and one ReLU activation function layer. The convolution kernel of each convolutional layer is a 3×3 convolution kernel. After being processed by the two sets of residual structures, the data is input into the CBAM attention mechanism for processing. The outputs of the discriminator are category judgment and true or false judgment.

4. The rolling bearing fault diagnosis method based on improved generative adversarial network according to claim 1 is characterized in that: In step S3, the loss function of the improved generative adversarial network model is as follows: The wavelet scale map obtained by continuous wavelet transform is input into the improved generative adversarial network for training to generate high-quality generated images. The discriminator training involves two modes: supervised learning and unsupervised learning. The loss function of the discriminator consists of supervised learning loss L1 and unsupervised learning loss L2, as shown in formula (1), formula (2) and formula (3): L D =L1+L2 (1) In the formula, x represents the input, y represents its corresponding label, and P r(x,y) Represents the data distribution of the original image, P g Represents the mixed generated image g mix The data distribution of g mix =α·g1+(1-α)·g2, where g1 and g2 are the images generated by the two generators respectively, and α is the dynamic weight; In the supervised learning part, Softmax is used as the activation function of the discriminator, and the classification cross entropy loss L1 is used; in the unsupervised learning part, the discriminator is the same as the ordinary GAN. The only thing that needs to be judged is the authenticity of the output sample. The activation function used is Sigmoid and the loss is L2. The generator and discriminator can be trained by minimizing the loss function; the loss function of the generator is shown in formula (4): Where, represents the data distribution of the original image, and represents the data distribution of the images generated by the two generators, g mix To generate a mixed image, where g mix =α·g1+(1-α)·g2, where g1 and g2 are the images generated by the two generators respectively, α is the dynamic weight, λ is the regularization parameter, ‖θ‖1 represents the L1 norm of the model parameter θ, and λ‖θ‖1 represents the L1 loss; The introduction of the AdaptiveMix module proposes a soft loss, as shown in formula (5), Where x i and x j represent the i-th and j-th training images respectively, Represents x i and x j The generated difficult samples, σ is the noise term sampled from Gaussian distribution to prevent overfitting, D v (·,·) refers to the metric used to evaluate the distance, such as the L1 norm and the L2 norm, and λ is a hyperparameter sampled from the Beta distribution λ∈Beta(α,α); Therefore, the total loss function of the improved GAN after applying the AdaptiveMix module is as follows: L IGAN =L D +L G +L Ada (6) 5. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the rolling bearing fault diagnosis method based on the improved generative adversarial network as described in any one of claims 1 to 4.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the rolling bearing fault diagnosis method based on the improved generative adversarial network according to any one of claims 1 to 4 is implemented.

Images