A fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network

Abstract

The application discloses a kind of based on auxiliary generation network and domain confrontation chart network's fault diagnosis method, belong to fault diagnosis technical field, for sample unbalance problem, it is proposed based on ACGAN sample generation network ICGAN, improve the generation quality of generation network to specified class sample;Then design based on global attention and octave convolution fault sample time-frequency chart generation network GAMICGAN, by introducing global attention mechanism (GAM) so that generation network pays more attention to spatial, channel and dimension information in time-frequency chart, further optimize generation network structure;Finally, for the problem that the diagnostic model accuracy decreases caused by the distribution difference of sensing data under variable working condition scene, combined with graph convolution network GCN, a kind of joint distribution domain confrontation network's fault field adaptive diagnosis method GIC-JDAGCN is proposed, and the result shows that GIC-JDAGCN can maintain higher diagnostic accuracy in variable working condition application, with good working condition scene adaptability.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis, and in particular relates to a fault diagnosis method based on auxiliary generation network and domain adversarial graph network. Background Technology

[0002] In recent years, with the rapid development of society, rotating machinery has been widely used in energy conversion, industrial production, transportation, aerospace and defense. Generally speaking, rotating machinery consists of a main structure, rotating parts, power transmission system and control system, etc., which together constitute a complex and close system engineering. Damage to one part may affect other parts, thereby causing abnormal operation of rotating machinery and causing certain economic losses [2]. The core component of rotating machinery is rotating parts, including shafts, bearings, rotors, etc. [3]. In harsh industrial production environments, long-term vibration and impact or high temperature and high pressure can cause wear and cracks in rotating parts. According to statistics, 70% of rotating machinery failures are caused by gears, bearings and rotors, of which more than 30% are caused by bearings. Therefore, research on intelligent fault diagnosis methods is of great significance to industrial production. With the development of science and technology, the use of intelligent manufacturing, Internet of Things and other technologies has made the production process more intelligent, and equipment fault diagnosis has entered the era of intelligence. Compared with traditional fault diagnosis methods, fault diagnosis methods based on deep learning can automatically learn the deep features of data and realize intelligent diagnosis of rotating machinery. Currently, data-driven fault diagnosis of mechanical equipment has become one of the research hotspots in the industrial field, and various deep learning algorithm models play an important role in fault diagnosis tasks.

[0003] Data-driven fault diagnosis methods typically rely on a sufficient amount of training data across various fault states. However, in real-world industrial production environments, while the acquisition of various sensor data has been largely resolved, effective data labeling remains challenging. Furthermore, because equipment often operates under relatively normal conditions for extended periods, acquiring sensor data under different fault states is difficult. Consequently, data samples collected in industrial settings suffer from imbalanced state types and incomplete labeling. When training fault diagnosis models using imbalanced datasets, classifiers tend to favor the class with more samples, leading to low accuracy in fault category identification and reduced generalization ability.

[0004] Rotating machinery typically operates under varying working environments and conditions. The probability distribution of sensor data collected under different operating conditions differs, impacting the diagnostic performance of models in varying operating scenarios. Therefore, in actual industrial field fault diagnosis of machinery, it is crucial to consider how to minimize the impact of changes in operating conditions on the extracted sensor data features, posing a challenge to traditional fault diagnosis methods. In recent years, transfer learning has addressed this issue by applying knowledge from related domains to the target domain, mapping data with different distributions to a unified feature space. This resolves the inconsistency between the source and target domain distributions, improving fault diagnosis performance and adaptability to different operating scenarios. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a fault diagnosis method based on auxiliary generative networks and domain adversarial graph networks to address the shortcomings of the prior art; it effectively solves the problem of class imbalance in training samples in mechanical equipment fault classification and diagnosis tasks, and improves the adaptability of the diagnostic model to changing working conditions.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A fault diagnosis method based on auxiliary generation network and domain adversarial graph network specifically includes the following steps;

[0008] Step 1: Targeting the non-stationary and nonlinear characteristics of vibration signals, a time-frequency analysis method based on wavelet analysis is used to preprocess the one-dimensional original vibration signal using continuous wavelet transform to obtain a two-dimensional wavelet time-frequency diagram that more effectively expresses the fault mechanism. The time-frequency diagram is used as the processing object of the generator network and diagnostic model. An independent auxiliary classifier structure is designed for the ACGAN network, and a weight adjustment mechanism for the discriminator and auxiliary classifier is introduced to obtain the improved generator network model ICGAN, thereby improving the generation quality of the generator network for samples of the specified category.

[0009] Step 2: Further optimize the ICGAN generative network structure using global attention and octave convolution; GAMICGAN, a fault sample time-frequency map generation network based on global attention and octave convolution, introduces the global attention mechanism GAM to make the generative network pay more attention to the spatial, channel, and dimensional information in the time-frequency map; replace the traditional two-dimensional convolution in the network with octave convolution OctConv to extract high and low frequency component information in the time-frequency map; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of generated samples;

[0010] Step 3 addresses the issue of decreased diagnostic model accuracy due to differences in sensor data distribution under varying operating conditions. It combines Graph Convolutional Networks (GCNs) with a backbone network structure, GAMLRN, based on a Global Attention Residual Network, to extract depth features from the time-frequency maps of vibration signal samples. The adversarial transfer mechanism of domain adversarial neural networks is used to align fault features under different operating conditions. Simultaneously, a joint maximum mean difference metric is introduced to quantify the edge and conditional distribution differences between the source and target domains, optimizing the backbone network parameters and further reducing the distribution differences between the depth features of the time-frequency maps of samples under different operating conditions. Furthermore, GCN embedding theory is used to construct structural relationships between samples in the same / different domains to further reduce the distribution differences between the source and target domains, and further reduce the distribution differences between the depth features of the time-frequency maps of samples under different operating conditions.

[0011] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0012] 1. To address the non-stationary and nonlinear characteristics of vibration signals, a time-frequency analysis method based on wavelet analysis is designed. Continuous wavelet transform is used to preprocess the one-dimensional original vibration signal, resulting in a two-dimensional wavelet time-frequency graph that more effectively expresses the fault mechanism. This time-frequency graph is then used as the processing object for the generator network and diagnostic model. An independent auxiliary classifier structure is designed for the ACGAN network, and a weight adjustment mechanism for the discriminator and auxiliary classifier is introduced, resulting in an improved generator network model, ICGAN, to enhance the generation quality of samples of a specified category. This invention improves ACGAN by designing an independent classifier structure and introducing a weight adjustment mechanism for the discriminator and auxiliary classifier, thus obtaining an improved generator network model, ICGAN.

[0013] 2. To further improve the attention paid by the generator network to the time-frequency features of faults in the time-frequency maps of vibration signal samples, this invention further optimizes the generator network structure of ICGAN using global attention and octave convolution. A fault sample time-frequency map generator network, GAMICGAN, based on global attention and octave convolution, is designed. By introducing the global attention mechanism GAM, the generator network pays more attention to spatial, channel, and dimensional information in the time-frequency map. OctConv octave convolution replaces the traditional two-dimensional convolution in the network to extract high- and low-frequency component information from the time-frequency map. By combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of generated samples. Furthermore, ICGAN is improved by introducing the global attention mechanism GAM to make the generator network pay more attention to spatial, channel, and dimensional information in the time-frequency map; OctConv octave convolution replaces the traditional two-dimensional convolution in the network to extract high- and low-frequency component information from the time-frequency map, resulting in the improved generator network model GAMICGAN.

[0014] 3. This invention addresses the issue of decreased diagnostic model accuracy due to variations in sensor data distribution under varying operating conditions. It proposes a mechanical equipment fault domain adaptation diagnostic method, JDAGCN, combining a Graph Convolutional Network (GCN) and a Joint Distributed Domain Adversarial Neural Network (DAGCN). A backbone network structure, GAMLRN, based on a Global Attention Residual Network, is designed to extract depth features from the time-frequency maps of vibration signal samples. The adversarial transfer mechanism of the domain adversarial neural network is used to align fault features under different operating conditions. Simultaneously, a joint maximum mean difference metric is introduced to quantify the edge and conditional distribution differences between the source and target domains, optimizing the backbone network parameters and further reducing the distribution differences between the depth features of the time-frequency maps of samples under different operating conditions. Furthermore, the embedding theory of the Graph Convolutional Network (GCN) is used to construct structural relationships between samples in the same / different domains to further reduce the distribution differences between the source and target domains, and further reduce the distribution differences between the depth features of the time-frequency maps of samples under different operating conditions. Experimental results show that the mechanical equipment fault diagnosis method based on the Joint Distributed Domain Adversarial Neural Network maintains high diagnostic accuracy in applications with varying operating conditions and exhibits good adaptability to different operating scenarios.

[0015] Based on the expansion of training samples by the generative network, this invention combines a deep transfer learning framework to establish a pre-adaptive diagnostic model for variable working conditions. (1) A backbone network structure GAMLRN based on a global attention residual network is designed to extract the deep features of the time-frequency map of vibration signal samples; (2) An adversarial transfer mechanism is used to align the features of fault samples under different working conditions; (3) A joint maximum mean difference metric is introduced to quantify the edge distribution difference and conditional distribution difference between the source and target domains; (4) The graph convolutional network GCN embedding theory is used to construct the structural relationship between samples in the same / different domains to further reduce the distribution difference between the source and target domains. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the principle of a fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network according to the present invention.

[0018] Figure 2 This is the vibration signal sample segmentation process of the present invention;

[0019] Figure 3 This is a schematic diagram of the Bump wavelet function of this invention;

[0020] Figure 4 This is a time-frequency image obtained by converting the vibration signal of the mechanical equipment according to the present invention;

[0021] Figure 5 This is a schematic diagram of the ICGAN network structure of the present invention;

[0022] Figure 6 This is the ICGAN generator network structure of the present invention;

[0023] Figure 7 This is the ICGAN discriminator network structure of the present invention;

[0024] Figure 8 This is the independent classifier network structure of the present invention;

[0025] Figure 9 This is a schematic diagram of the GAMICGAN generator network structure of the present invention;

[0026] Figure 10 This is the GAMICGAN discriminator network structure of the present invention;

[0027] Figure 11 This is the global attention network structure of the present invention;

[0028] Figure 12 This is the GAMICGAN generator network structure of the present invention;

[0029] Figure 13 This is the GAMLRN network structure of the present invention;

[0030] Figure 14 This invention is the JDAGCN mechanical equipment fault migration diagnosis model.

[0031] Figure 15 This is the flowchart of the JDAGCN variable operating condition diagnosis of the present invention;

[0032] Figure 16 (a) is the iterative loss curve of the discriminator of the present invention;

[0033] Figure 16 (b) is the iterative output of the discriminator of this invention;

[0034] Figure 17 (a) is the iterative loss curve of the discriminator of the present invention;

[0035] Figure 17 (b) is the iterative output of the discriminator of this invention;

[0036] Figure 18 This is a time-frequency diagram of fault samples generated by the ICGAN of this invention;

[0037] Figure 19 This is the FID value during the iteration process of this invention;

[0038] Figure 20 This is a comparison of the classification results of different generator networks of the present invention under various imbalance ratios;

[0039] Figure 21 (a) is the loss curve of the present invention;

[0040] Figure 21 (b) is the output value of the discriminator of the present invention;

[0041] Figure 22 This is a time-frequency diagram of fault samples generated by the GAMICGAN method of this invention;

[0042] Figure 23 This invention compares the heatmaps of ICGAN and GAMICGAN.

[0043] Figure 24 This invention compares the accuracy on imbalanced datasets.

[0044] Figure 25 These are the accuracy curves of the comparative models under different working conditions and migration tasks of this invention;

[0045] Figure 26 This is a t-SNE distribution diagram of each ablation experiment model of the present invention under operating conditions 0 to 3;

[0046] Figure 27 These are the t-SNE plots of each model in the comparative experiment of this invention, transitioning from condition 0 to condition 3;

[0047] Figure 28 This is the confusion matrix of each model in the comparative experiment of this invention, migrating from condition 0 to condition 3. Detailed Implementation

[0048] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0050] A fault diagnosis method based on auxiliary generative networks and domain adversarial graph networks, such as Figure 1 As shown, the specific steps are as follows;

[0051] Step 1: Addressing the non-stationary and nonlinear characteristics of vibration signals, a time-frequency analysis method based on wavelet analysis is employed. Continuous wavelet transform is used to preprocess the one-dimensional original vibration signal, yielding a two-dimensional wavelet time-frequency graph that more effectively expresses the fault mechanism. This time-frequency graph is then used as the processing object for the generative network and diagnostic model. An independent auxiliary classifier structure is designed for the ACGAN network, and a weight adjustment mechanism for the discriminator and auxiliary classifier is introduced, resulting in an improved generative network model, ICGAN. This enhances the generation quality of the generative network for samples of a specified category. Further improvements are made to ACGAN by designing an independent classifier structure and introducing a weight adjustment mechanism for the discriminator and auxiliary classifier, resulting in an improved generative network model, ICGAN.

[0052] Step 2: Further optimize the ICGAN generative network structure using global attention and octave convolution; GAMICGAN, a fault sample time-frequency map generation network based on global attention and octave convolution, introduces the global attention mechanism GAM to make the generative network pay more attention to the spatial, channel, and dimensional information in the time-frequency map; replace the traditional two-dimensional convolution in the network with octave convolution OctConv to extract the high and low frequency component information in the time-frequency map; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of generated samples; This invention improves ICGAN by introducing the global attention mechanism (GAM) to make the generative network pay more attention to the spatial, channel, and dimensional information in the time-frequency map; and replaces the traditional two-dimensional convolution in the network with octave convolution (OctConv) to extract the high and low frequency component information in the time-frequency map, resulting in the improved generative network model GAMICGAN.

[0053] Step 3 addresses the issue of decreased diagnostic model accuracy due to variations in sensor data distribution under varying operating conditions. It combines Graph Convolutional Networks (GCNs) with a backbone network structure, GAMLRN, based on a Global Attention Residual Network, to extract deep features from the time-frequency maps of vibration signal samples. The adversarial transfer mechanism of domain adversarial neural networks (GANNs) is used to align fault features under different operating conditions. Simultaneously, a joint maximum mean difference metric is introduced to quantify the edge and conditional distribution differences between the source and target domains, optimizing the backbone network parameters and further reducing the distribution differences between deep features in the time-frequency maps of samples under different operating conditions. Furthermore, GCN embedding theory is used to construct structural relationships between samples in the same / different domains to further reduce the distribution differences between the source and target domains, and further reduce the distribution differences between deep features in the time-frequency maps of samples under different operating conditions. Based on the expansion of training samples using a generative network, a pre-adaptive diagnostic model for varying operating conditions is established using a deep transfer learning framework. (1) A backbone network structure GAMLRN based on global attention residual network was designed to extract the depth features of the time-frequency map of vibration signal samples; (2) The adversarial transfer mechanism was used to align the features of fault samples under different working conditions; (3) The joint maximum mean difference metric was introduced to quantify the edge distribution difference and conditional distribution difference between the source domain and the target domain; (4) The graph convolutional network GCN embedding theory was used to construct the structural relationship between samples in the same / different domains to further reduce the distribution difference between the source domain and the target domain.

[0054] ICGAN-based method for generating fault samples in generative networks

[0055] When data is sufficient and the sample is balanced, data-driven AI-based fault diagnosis methods can accurately identify the health status of equipment. However, in actual equipment operation, the vibration signals accumulated by the field data acquisition system over a long period are mainly normal signals, making it difficult to collect equipment fault signals. This leads to problems such as insufficient fault samples, limited fault categories, and sample imbalance, which affect the identification accuracy of equipment fault diagnosis methods.

[0056] This section addresses the imbalanced sample problem from the perspective of data generation. Building upon the ACGAN generative network, it proposes an improved time-frequency image augmentation method, ICGAN, for fault samples. Continuous wavelet transform is used to convert the one-dimensional vibration signal of the equipment into a two-dimensional time-frequency image with rich time-frequency features. An improved conditional generative network, ICGAN, is proposed, modifying the ACGAN network structure and designing independent auxiliary classifiers with independent parameters and weights. The two-dimensional time-frequency image is input into the ICGAN network for data generation, and the FID (Fault-Induced Disorder) evaluation metric is used to assess the quality of the generated samples. Qualified generated samples are then added to the original imbalanced dataset to improve the classification performance of subsequent fault diagnosis models.

[0057] Vibration signal constructs time-frequency image:

[0058] 1. Vibration signal segmentation

[0059] The obtained one-dimensional vibration signal is divided into multiple sample sets of the same sequence length using a sliding window method. The sliding window value selection process is as follows: Figure 2 As shown, to ensure that the segmented vibration signal has signal characteristics of at least one rotation cycle, the number of data points in each sequence sample must meet the following condition:

[0060]

[0061] Where N represents the number of data points in each vibration signal sample, i.e., the length of the sample. n f is the number of rotations for each vibration signal sample from the mechanical equipment. s is the sampling frequency when collecting vibration signals, representing the number of data points recorded per second. v is the rotational speed of the mechanical equipment, representing the number of rotations per minute.

[0062] 2. Construction of two-dimensional time-frequency diagrams

[0063] One-dimensional vibration signals generated during mechanical equipment failure contain a large amount of redundant information and noise interference unrelated to the fault characteristics. This section utilizes continuous wavelet transform to preprocess the original vibration signal through noise reduction and filtering, transforming the one-dimensional signal into a two-dimensional time-frequency plot. During signal analysis and processing, wavelet decomposition helps reduce redundant information in the original time-series vibration signal, allowing the model to focus on the key fault characteristics that generate the vibration signal, thereby improving the model's accuracy and generalization ability. Furthermore, the two-dimensional time-frequency plot not only preserves the time-domain information of the original vibration signal but also visualizes the energy distribution of the signal at different frequencies, clearly demonstrating the frequency domain characteristics of the vibration signal. This makes it easier for the classification model to learn the specific category feature representations in the time-frequency plot, enabling differentiation between different fault categories. In summary, the two-dimensional time-frequency plot obtained through wavelet transform can better reflect the spectral characteristics of vibration signals from mechanical equipment under different operating conditions, which helps improve the quality of sample generation and the accuracy of fault diagnosis and classification.

[0064] Choosing a suitable wavelet function for continuous wavelet transform of the original signal is crucial. This paper employs the Bump wavelet to perform continuous wavelet transform on the original signal and converts the resulting wavelet coefficient matrix into a two-dimensional time-frequency feature map. The shape of the Bump wavelet function resembles a strip wave packet, possessing finite time and finite frequency characteristics. It has advantages in focusing on the local and instantaneous features of a signal. Its finite and localized nature in the time domain aligns with the characteristics of impulse signals. Furthermore, due to its localized nature in the frequency domain, the Bump wavelet can more effectively capture transient and impulse components in the signal. The graph of the Bump wavelet function is shown below. Figure 3 As shown. The Bump wavelet function is represented as:

[0065]

[0066] The two-dimensional time-frequency image obtained by converting a one-dimensional vibration signal using the Bump wavelet function, such as... Figure 4 As shown.

[0067] In this time-frequency graph, the horizontal axis represents time, the vertical axis represents frequency, and the color or brightness changes represent the amplitude of the signal at different frequencies and times.

[0068] The converted time-frequency graph clearly shows a distinct periodicity in the time scale when mechanical equipment malfunctions. The pulse signals caused by the mechanical equipment malfunction appear as bright spots on the time-frequency graph. Meanwhile, noise signals, due to their lower amplitude compared to the pulse signals, appear as blue on the time-frequency graph. This indicates that converting to a time-frequency graph, compared to the original time sequence signal of the mechanical equipment, not only effectively shields the interference noise outside the mechanical equipment malfunction signal, but also provides a variety of feature representations with different frequencies and resolutions. These features can better reflect the dynamic characteristics and frequency component changes of the signal, helping to more accurately identify different types of mechanical equipment malfunctions.

[0069] ICGAN-based method for expanding mechanical equipment fault samples

[0070] By improving the ACGAN network, a Generative Adversarial Network (ICGAN) based on Independent Classifier is proposed. By designing independent auxiliary classifier structures for the ACGAN network and assigning different weights to the network parameters of the discriminator and auxiliary classifiers, the quality of generated faulty samples is improved.

[0071] The original ACGAN network architecture improves the quality of generated samples by adding an auxiliary classifier to the discriminator on top of the CGAN network. However, the auxiliary classifier's classification is based on the discriminator's architecture. In ACGAN, the discriminator performs both classification and real / false discrimination functions; that is, the neural networks for classification and real / false discrimination have the exact same hyperparameters and weights. Therefore, the auxiliary classifier in ACGAN is designed as an independent structure. By giving the network an independent classifier, classification performance is enhanced, resulting in higher quality samples of specific categories generated by the generator.

[0072] The ICGAN network structure proposed in this paper is as follows: Figure 5 As shown, on one hand, the generator generates samples similar to the real sample time-frequency maps based on random noise and label vectors. On the other hand, the discriminator distinguishes between the generated sample time-frequency maps and the real sample time-frequency maps to optimize the quality of the time-frequency map samples generated by the generator. By adopting an independent classifier design structure, the generated sample time-frequency maps in each category are more similar to the real sample time-frequency maps and better match the distribution of the real sample time-frequency maps. Here, visualization techniques such as the generated sample evaluation index FID and loss curves are used to determine the optimal network model parameters.

[0073] The network structure of the generator in ICGAN is as follows: Figure 6As shown, the input dimension is 110, with 100 dimensions representing the noise signal and 10 dimensions representing the label information. After feature transformation of the input noise through a fully connected layer, it is passed through 5 layers of 2D deconvolution layers to obtain a 3-channel, 128-dimensional feature map. The specific network parameters of the ICGAN generator are shown in Table 1.

[0074] Table 1

[0075]

[0076] The network structure of the discriminator in ICGAN is as follows: Figure 7 As shown, the discriminator first uses five 2D convolutional layers to transform the 3-channel, 128-size feature map obtained by the generator into a 52-channel, 13-size feature map. Then, a fully connected layer outputs 0 / 1 to represent the discriminant output for the generated samples. The specific network parameters of the discriminator are shown in Table 2.

[0077] Table 2

[0078]

[0079]

[0080] The network structure of an independent classifier is as follows: Figure 8 As shown, the independent classifier classifies the feature map output by the generator and the image of the real sample, and backpropagates the classification result to further optimize the generator's generation process and improve the generation quality of each class of generated samples. The network parameters of each layer of the independent classifier are shown in Table 3.

[0081] Table 3

[0082]

[0083] Steps for Expanding Mechanical Equipment Fault Samples Based on ICGAN

[0084] (1) Data acquisition and processing. Vibration signals of the equipment under various working conditions and different fault types are acquired, and the original vibration signals are segmented using a sliding window. Continuous wavelet transform is used to convert the one-dimensional signal into a two-dimensional wavelet time-frequency diagram, thereby constructing an unbalanced training sample set and a test sample set.

[0085] (2) Generative network training and sample evaluation. A small number of training samples are fed into the ICGAN generative network for iterative training. The optimal network model parameters are determined using visualization techniques such as the generated sample evaluation index FID and loss curves.

[0086] (3) Imbalanced sample augmentation. Random noise and label vectors are input into the trained ICGAN generator network to generate time-frequency maps of multiple fault types, thereby augmenting the original imbalanced training sample set;

[0087] ICGAN Optimization Method Based on Global Attention and Octal Convolution: Model Overview

[0088] This section further optimizes ICGAN by designing GAMICGAN, a fault sample time-frequency map generation network based on global attention and octave convolution. By introducing a global attention mechanism (GAM), the generation network pays more attention to spatial, channel, and dimensional information in the time-frequency map. Octave convolution (OctConv) replaces the traditional two-dimensional convolution in the network to extract high- and low-frequency component information in the time-frequency map. By combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of generated samples.

[0089] GAMICGAN Network Structure: To improve the quality of two-dimensional time-frequency images generated by the ICGAN generator network and enhance the network's attention to the fault time-frequency features in the two-dimensional time-frequency image of the vibration signal, a mechanical equipment fault diagnosis method based on global attention and octave convolution, GAMICGAN, is proposed. A schematic diagram of the GAMICGAN network structure is shown below. Figure 9 As shown, firstly, a global attention mechanism is embedded in the generator and discriminator of ICGAN, enabling the generative network model to fully consider the spatial, channel, and dimensional information in the time-frequency map when generating fault sample time-frequency maps. Secondly, the ordinary transposed convolution and ordinary convolution in the generator and discriminator of ICGAN are replaced with octave transposed convolution DeOctConv and octave convolution OctConv, respectively, so that the network can pay attention to the high and low frequency component information of the time-frequency map. By introducing the global attention mechanism and octave convolution, the generation quality of fault sample time-frequency maps is effectively improved, and the accuracy of mechanical equipment fault diagnosis under sample imbalance is further improved.

[0090] The GAMICGAN network mainly improves the generator and discriminator structures of the ICGAN network. The structure of the independent classifier remains unchanged, and the octave convolution is designed to have the same feature map size as that obtained after traditional convolution operations.

[0091] 1. Discriminator Network: In the discriminator, the input image is first downsampled using traditional convolution, and the resulting feature map is then processed by the global attention module GAMATtention to focus on the channel, spatial, and dimensional information in the feature map. Next, three OctConv modules are used to deeply extract the high- and low-frequency component features from the feature map. Finally, traditional convolution and fully connected layers are used to distinguish between real and fake samples. The improved discriminator network structure is as follows: Figure 10 As shown in the table below, the GAMATtention module and the OctConv module will be described in detail. The specific network structure parameters are shown in Table 4.

[0092] Table 4

[0093]

[0094] 2. GAMATtention Global Attention Module: The global attention module mainly consists of two parts: channel attention and spatial attention. For the discriminator network structure discussed in this section, the specific parameters and data flow of the two attention modules are as follows: Figure 11 As shown:

[0095] Channel attention enables the model to focus on key features in the input data by determining the weights of each channel. It learns the weights between channels to focus on task-relevant channels and effectively captures cross-channel interactions by leveraging the complex dependencies between channels, thus improving the feature representation of specific faults.

[86] This allows the generative network to generate high-quality sample images. The channel attention module preserves channel information in the three dimensions of the image. First, the input feature map is transposed. Then, the resulting feature map is processed through the encoding and decoding structure of an MLP (Multilayer Perceptron) with a compression ratio of r. We set r to 4 to amplify the interaction between channels and space. Finally, the feature map is inverted and activated by the Sigmoid function, resulting in the feature map after channel attention processing.

[0096] Spatial attention mechanisms consider the relationships between image pixels, enabling the model to selectively focus on task-relevant information. It learns the weights of each location to represent the importance of each position in the image, thereby improving the model's ability to represent key image features. First, the feature map processed by channel attention is multiplied with the input feature map and then fused using a 7×7 convolutional layer to better capture the local structure of the image. Next, the same compression ratio *r* as channel attention is used to further compress the feature map's dimensionality. Then, another 7×7 convolutional layer is applied to further extract spatial features. Finally, a sigmoid activation function is used to obtain the spatially attention-processed feature map, which is then element-wise multiplied with the feature map processed by channel attention to obtain the final feature map output.

[0097] 3. OctConv octave convolution module: such as Figure 11 As shown, OctConv divides the input feature map into high-frequency features. Figure X H and low frequency characteristics Figure X L These represent the high-frequency and low-frequency components of the input feature map, respectively. OctConv uses α in and α out The input and output channel ratios of high and low frequency components are controlled, with α ranging from [0,1]. In the OctConv convolutional layer, except for the first and last layers, the value of α is set to α. in It is 0.5, α out The value is set to 0.5 to facilitate the extraction of high- and low-frequency feature information. The discriminator's network structure uses a three-layer OctConv module to update high- and low-frequency information.

[0098] Y H→H The output features representing the information update between high-frequency components, Y L→L This represents the information update output feature between low-frequency components. H→L and Y L→H This represents the information exchange output characteristics between different frequency components. The implementation is represented as follows:

[0099] Y H→H =Conv(X H (3-3)

[0100] Y L→L =Conv(X L (3-4)

[0101] Y H→L =Conv(pool(X) H ,2)) (3-5)

[0102] Y L→H =upsample(Conv(X) L ),2) (3-6)

[0103] Here, pool(X,2) represents a pooling operation with a step size of 2, and upsample(X,2) is an upsampling operation using nearest neighbor interpolation with a step size of 2.

[0104] The first layer of OctConv will have α in Set to 0, α out Setting it to 0.5 decomposes the input feature map into two feature map components, Y, which are high-frequency and low-frequency. H and Y L Specifically, it is expressed as:

[0105] Y H =Y H→H (3-7)

[0106] Y L =Y H→L(3-8)

[0107] OctConv second layer, α in and α out All values ​​are set to 0.5, and the high-frequency feature map Y obtained from the first layer is... H and low-frequency feature map Y L The fusion process yields the following new high- and low-frequency feature maps:

[0108] Y H =Y H→H +Y L→H (3-9)

[0109] Y L =Y H→L +Y L→L (3-10)

[0110] OctConv third layer, α out Set to 0, α in Setting it to 0.5 means that the high-frequency feature map component Y obtained from the second layer... H As the final feature map output, the output Y is obtained at this time. H for:

[0111] Y H =Y H→H +Y L→H (3-11)

[0112] 4. Generator Network Structure: In the generator, random noise with label information is passed through a fully connected layer to obtain a one-dimensional vector. This vector is first subjected to traditional deconvolution to obtain a feature map, then passed through three DeOctConv layers for channel and size scaling. Global attention is used to obtain a feature map of the same size. Finally, traditional convolution is used to obtain the final generator output. The improved generator network model structure is as follows: Figure 12 As shown in Table 5, the specific network structure parameters are as follows. Unlike the OctConv module of the discriminator, the generator is designed with DeOctConv instead of ordinary transposed convolution, where g(·) represents ordinary transposed convolution.

[0113] Table 5

[0114]

[0115] Fault diagnosis of mechanical equipment under varying operating conditions based on joint distributed domain adversarial graph convolutional neural networks:

[0116] In real-world industrial scenarios, rotating machinery typically operates under various conditions. The same fault point under different conditions can generate fault signals with varying data distributions, leading to problems such as weak generalization ability and decreased accuracy in diagnostic models. To achieve fault diagnosis of electromechanical equipment under varying operating conditions, this paper proposes JDAGCN, a fault diagnosis method for such equipment based on a joint distribution domain adaptive adversarial graph convolutional network, building upon the samples generated by GAMICGAN. First, a lightweight feature extraction module, GAMLRN, based on a global attention residual network, is designed to extract deep features from the time-frequency graph. Second, the adversarial transfer mechanism of a domain adversarial neural network is used to align fault features under different operating conditions. A joint maximum mean difference metric is introduced to quantify the differences in edge and conditional distributions between the source and target domains. Finally, a graph convolutional network (GCN) is used to construct structural relationships between samples within the same domain and between samples from different domains, further reducing the distribution differences between the source and target domains. Experimental analysis was conducted on the CWRU bearing dataset. The results show that the JDAGCN model has a high fault diagnosis accuracy in variable operating condition scenarios, and this method is superior to other variable operating condition fault diagnosis models.

[0117] GAMLRN Feature Extraction Network Design: In order to extract the deep features of the vibration signal time-frequency map, and to prevent the network from becoming too complex and causing a decrease in the model's generalization ability, this section designs a lightweight feature extractor based on a global attention residual network (GlobalAttention Mechanism Lightweight Residual neural network, GAMLRN).

[0118] In the GAMLRN network, a 3×3 convolutional kernel is used, with a stride of 1 and padding values ​​of 1 or 0. Only three residual structures are used, pooling operations in the intermediate layers are disabled, and deep features in the time-frequency map are extracted by embedding global attention. The GAMLRN network structure is as follows: Figure 13 As shown in Table 6.

[0119] Table 6

[0120]

[0121] Variable operating condition fault diagnosis network framework: The structure of the variable operating condition fault diagnosis model based on a joint distributed domain adversarial neural network is as follows: Figure 14 As shown.

[0122] JDAGCN consists of data preprocessing, the feature extractor GAMLRN, the graph embedding theory GCN, and the label classifier G... y Domain Discriminator G dThe system comprises several components. The data preprocessing module processes data from both the source and target domains. The feature extractor, GAMLRN, employs a lightweight global attention residual neural network model to extract deep features from both the source and target domains. The label classifier, composed of fully connected layers, classifies the data from the source domain. The domain discriminator uses fully connected layers to distinguish whether data originates from the source or target domain. In the domain adaptation stage, a metric combining JMMD and GCN is used to reduce marginal and conditional distribution differences between the source and target domains.

[0123] The optimization objective of the JDAGCN model is determined by the label classification loss L. y Domain discrimination loss L d The difference loss between the JMMDGC distribution and L JMMDGC The structure is as follows: The final target loss function is L. JDANN :

[0124] L JDANN =L y -λ1L d +λ2L JMMDGC (3-12)

[0125] Here, λ1 and λ2 are both hyperparameters. λ1 is used to balance the label classification loss and the domain discriminant loss, while using gradient inversion to negatively value the weights of the domain discriminant loss. This allows the model to reduce the label classification loss while maximizing the domain discriminant loss. This adversarial training method enables the model to learn deeper feature representations and improves its generalization ability in the target domain. λ2 is used to balance the importance between the label classification loss and the JMMD loss, making the model pay attention to the differences in the joint feature distribution between the source and target domains.

[0126] Variable operating condition fault diagnosis process: JDAGCN variable operating condition diagnosis process is as follows Figure 15 As shown. The JDAGCN variable operating condition diagnostic process consists of the following four steps:

[0127] Step 1: Data Preprocessing. The original vibration signals from the source and target domains are acquired, and continuous wavelet transform is performed on the vibration signals to obtain corresponding time-frequency image samples. Then, the two-dimensional time-frequency images of different fault types are input into the GAMICGAN fault sample time-frequency image generation network based on global attention and octave convolution. By mixing the generated fault category sample data with the original sample set, the imbalanced dataset is expanded, thereby improving the fault diagnosis accuracy under imbalanced samples.

[0128] Step 2: Feature Extraction. After setting the model parameters of JDAGCN, the obtained source and target domain time-frequency map samples are input into GAMLRN to extract deep features from both the source and target domains.

[0129] Step 3: Domain Adaptation. The overall loss function is calculated, and the label classifier loss, domain discriminator loss, and joint maximum mean difference loss are optimized through backpropagation. Before the maximum number of rounds, deep features from the source and target domains are extracted cyclically using a domain adversarial approach to obtain identical feature representations between the two domains.

[0130] Step 4: Diagnosis and Classification. After obtaining the trained JDAGCN model, the time-frequency map of the fault samples in the target domain is input into the JDAGCN model to obtain the final target domain classification result.

[0131] Experimental Analysis: Experimental analysis based on ICGAN generative networks:

[0132] To verify the data augmentation effect of the proposed ICGAN network in mechanical equipment fault diagnosis, two-dimensional time-frequency image samples were input into the ICGAN network to generate fault samples. The mixed samples transformed the imbalanced state of the original dataset into a balanced state. The effectiveness and feasibility of the proposed network were verified by comparing the classification performance of the same CNN classifier before and after data augmentation on the CWRU bearing dataset. Furthermore, comparative experiments were conducted to demonstrate the high availability of the proposed network by comparing its augmentation performance with other generative networks. Experimental results show that the loss curve of the designed ICGAN generator converges quickly, and the generated time-frequency images clearly express the feature information of bearings under different health conditions. Moreover, with an imbalance ratio of 1:40, after augmenting a small number of training samples with ICGAN, the model's diagnostic accuracy improved by approximately 15%, which is about 7% higher than other generative network models.

[0133] Sample generation evaluation: The hyperparameters of the ICGAN generator network are as follows: batch size is 40, number of iterations is 3000, learning rate of generator, discriminator and independent classifier is 0.0002, and Adam is used as the optimizer.

[0134] The two-dimensional time-frequency plot obtained from the CWRU bearing dataset is input into the ICGAN network to generate fault samples of a specific class. The iterative loss curves of the ICGAN network generator and discriminator are shown below. Figure 16 As shown in (a), the iterative output of the discriminator is as follows: Figure 16 As shown in (b).

[0135] To compare the improvement effect with the original ACGAN generator network, the hyperparameters of the ACGAN generator network were set to the same batch size, number of iterations, learning rate, and optimizer as the ICGAN generator network. The iterative loss curves of the generator and discriminator in the original ACGAN are shown below. Figure 17 As shown in (a), the iterative output of the discriminator is as follows: Figure 17 As shown in (b).

[0136] By comparing the loss curves of ICGAN and the original ACGAN network, it can be seen that the generator loss of the original ACGAN network begins to rise after 500 epochs, and continues to rise in subsequent iterations. This indicates that the ACGAN network model has not converged and has experienced mode collapse, meaning that the generator quality of the ACGAN network is very poor, causing the discriminator to easily identify the generated samples as 'fake' samples. In contrast, ICGAN, with its independent classifier, constrains the generator's generation performance, causing the loss curve to continue to decrease after 500 epochs. In subsequent iterations, the losses of both the generator and the discriminator tend to stabilize. Furthermore, by comparing the discriminator output values ​​of ICGAN and the original ACGAN network, it can be seen that by using an independent classifier structure, the discriminator's discrimination values ​​for generated and real samples converge towards 0.5, i.e., the Nash equilibrium state.

[0137] Different types of fault generation samples obtained through the ICGAN generation network, such as Figure 18 As shown.

[0138] It can be seen that the generated two-dimensional time-frequency images of the same type of fault samples are similar, and there is a certain degree of distinguishability between different types of generated samples. That is, the data distribution of the generated fault samples is roughly the same as that of the real samples. This indicates that the ICGAN generator network has high stability during the generative adversarial training process. By designing the structure of an independent classifier, the generator can further optimize the generation process based on the classification results obtained by the classifier during the generation of fault samples, resulting in better quality of the generated two-dimensional time-frequency images.

[0139] Furthermore, in generative networks, the Inception Score (IS) and Frechet Inception Distance Score (FID) are commonly used as evaluation metrics to measure the quality and diversity of generator output. The IS metric, proposed by Salimans et al. in 2016, uses the generator's output image as input to the Inception V3 model to obtain the probability that the image belongs to a certain class. A higher probability value indicates a clearer image and a smaller entropy of p(y|x), thus judging the generator's generation quality. When the probability of generated samples belonging to each class is more equal, i.e., closer to a uniform distribution, the larger the entropy of p(y), and the more diverse the generated samples. Using KL divergence to calculate the distance between two probability distributions, the IS metric expression is:

[0140]

[0141] However, since the IS index is very sensitive to the weights of the network model, different model weights can cause the IS index to change significantly. Secondly, the IS index cannot detect the pattern collapse problem in the generative network. That is, even if a certain number of images are generated in each class, there may be a lack of diversity. This indicates that it only considers the quality of the generated samples and does not take into account the influence of the distribution of real data.

[0142] To address this, Heusel et al. proposed the FID (Fixed Distribution Index) metric in 2017. Compared to the IS (Inception V3) metric, which directly outputs the category when using the Inception V3 model, FID outputs the corresponding feature vector. FID uses the feature vectors from the Inception model to compare the distribution differences between the real and generated images. A lower FID indicates a closer similarity between the distributions of the generated and real images, and thus a higher quality generated sample. The FID calculation expression is:

[0143]

[0144] In the formula, x g represents the real image, m represents the generated image, and g represents the generated image. x and m g These are the extracted feature vectors, Tr(·) is the trace of the matrix, and C and C are the feature vectors extracted. W They are m x and m g The covariance matrix.

[0145] This paper uses the FID metric to evaluate the quality and diversity of the two-dimensional fault sample time-frequency maps generated by the generator. Table 7 records the FID values ​​of the images generated by the ICGAN network during the iteration process. The changes in FID during the iteration process are as follows: Figure 19 As shown.

[0146] Table 7

[0147]

[0148] Table 8

[0149]

[0150] Analysis of the above results reveals that as the number of network iterations increases, the FID of generated samples for each category gradually decreases and eventually stabilizes. This indicates that the generator has learned the approximate data distribution of the real sample time-frequency image, thus generating a generated sample time-frequency image similar to the real sample time-frequency image. The FID value decreases rapidly between 300 and 600 iterations, consistent with the variation pattern of the ICGAN loss function curve, indicating that the network begins to converge gradually after 600 iterations. The FID values ​​of different types of generated sample time-frequency images vary. For example, the FID values ​​of generated sample time-frequency images with an inner circle fault size of 0.014 and an outer circle fault size of 0.007 are smaller than those of generated sample time-frequency images with other fault types, while the FID values ​​of generated sample time-frequency images with an inner circle fault size of 0.021 and the normal state are relatively larger. This indicates that the ICGAN network exhibits diversity in generating two-dimensional time-frequency images of different fault types, making the generated two-dimensional time-frequency images of different fault types somewhat distinguishable.

[0151] Fault Diagnosis under Imbalanced Samples: To verify that the proposed ICGAN network can expand an imbalanced dataset into a balanced dataset, thereby improving the accuracy of bearing fault diagnosis under imbalanced samples, imbalanced datasets with different proportions using CWRU were set up, as shown in Table 9. The dataset was divided into training and testing sets, with the training set further divided into different imbalance ratios. 1:N represents the ratio between the time-frequency graphs of a certain bearing fault type and the time-frequency graphs of the normal state. The testing set used a 1:1 ratio for testing.

[0152] Table 9

[0153]

[0154] To verify the data augmentation effect of the ICGAN network, the following experiment was designed for verification:

[0155] (1) Experiment A: Under each imbalance ratio, real samples are directly classified by the CNN classifier without data augmentation.

[0156] (2) Experiment B: The imbalanced dataset of real samples was augmented using a GAN network to achieve a balanced state, and then the dataset was classified.

[0157] (3) Experiment C: The imbalanced dataset of real samples was augmented using the ACGAN network to achieve a balanced state, and then it was classified.

[0158] (4) Experiment D: The imbalanced dataset of real samples was augmented using the ICGAN network to achieve a balanced state, and then it was classified.

[0159] Imbalanced datasets were transformed into balanced datasets using different generative networks through data augmentation. The balanced datasets were then trained with the same CNN classifier, whose model parameters were identical to those of the independent classifiers. Training lasted 100 epochs with a batch size of 40, using the Adam optimizer and a learning rate of 0.002. The test accuracies are shown in Table 10. The classification accuracies obtained after augmentation with different generative networks at various imbalance ratios are shown in Table 10. Figure 20 As shown.

[0160] Table 10

[0161]

[0162] By comparing with other generative networks, it can be seen that the balanced state dataset expanded by the ICGAN network improves the final classification accuracy. The ICGAN generative network, through its independent classifier design structure, improves the accuracy of the original ACGAN network at various imbalance ratios. Moreover, the larger the imbalance ratio, the greater the improvement in accuracy. This indicates that the ICGAN network learns a more accurate distribution of real sample time-frequency image data than the original ACGAN network, generating two-dimensional time-frequency images with higher quality diversity and realism.

[0163] Experimental Analysis of GAMICGAN Network Performance: Evaluation of Generated Samples: The GAMICGAN network was trained and tested using the CWRU imbalanced dataset constructed in Chapter 3. The number of training epochs, batch size, learning rate, and optimizer parameters were set the same as those of the ICGAN network. The resulting loss curves of the GAMICGAN network discriminator and generator are shown below. Figure 21 As shown in (a), the output value of the discriminator is as follows: Figure 21 As shown in (b).

[0164] It can be seen that by adding global attention and using octave convolution in the ICGAN generator network, the generator and discriminator loss decreases faster and converges more stably. From the discriminator output value, it can be seen that compared with the ICGAN generator network, the discriminator's discrimination value for the generated samples is close to 0.45, indicating that the quality of the generated faulty samples is better, further confusing the discriminator's output and making it unable to distinguish between real and fake generated samples.

[0165] Time-frequency diagrams of various types of fault samples generated by the GAMICGAN network are as follows: Figure 22 As shown.

[0166] As can be seen from the figure, the generated fault sample time-frequency plot has less noise and better generation quality. The FID values ​​of the time-frequency samples generated by GAMICGAN at some iteration numbers are shown in Table 11.

[0167] Table 11

[0168]

[0169] As shown in Table 11, after 300 training epochs, the FID value of the GAMICGAN network reached between 200 and 500. The FID value decreased rapidly between epochs 300 and 600, and stabilized around epoch 1000. Compared to the ICGAN network, the GAMICGAN network has a lower FID value at the same number of iterations, indicating better sample diversity and higher similarity to real samples.

[0170] To verify the effectiveness of the GAMICGAN network in generating time-frequency maps of fault samples, this section, in addition to employing the numerical analysis method of FID, also uses the heatmap method to analyze the quality of the newly generated fault samples.

[0171] Gradient-weighted Class Activation Mapping (Grad-CAM)

[88] Grad-CAM is a method for explaining the decision-making process of CNN models. It analyzes the model's focus regions for a particular class by generating class activation heatmaps. Analyzing these heatmaps helps understand the feature learning process within the network and assess whether the network has correctly learned the feature information. Specifically, Grad-CAM first obtains the output gradient of the last convolutional layer of the model. Then, it generates a class activation map by representing the importance of the activation output of the last convolutional layer to the model's classification decision. This process can be viewed as a weighted operation, allowing the model to represent the degree of attention to different features. By rendering the preprocessed class activation map as a heatmap, the focus regions are highlighted. Finally, the original image is overlaid with the rendered heatmap to obtain the final heatmap. Brighter areas in a heatmap indicate areas that the model pays more attention to; the specific expression is as follows:

[0172]

[0173] Among them, f R Represents the ReLU activation function. Indicates A k The weights are: d represents the class, A represents the feature layer, and k represents the k-th channel in feature layer A. k This represents the data of channel k in feature layer A.

[0174] a k d The calculation formula is:

[0175]

[0176] Where Z represents the number of pixels in the feature map. d This represents the score predicted by the network for category d. This represents the pixel value of feature layer A at coordinate (i,j) in channel k.

[0177] The generated fault sample heatmap is, for example Figure 23 As shown, by comparing the heatmap of a fault sample with a ball bearing failure damage size of 0.021 inch under 0HP generated by the ICGAN network with the heatmap of the same type of fault sample generated by the GAMICGAN network, it can be seen that the GAMICGAN network focuses on the high-frequency components in the time-frequency map during the generation process. These high-frequency components are the characteristic representations of different bearing fault types in the time-frequency map. By adding a global attention mechanism and OctConv to the ICGAN network, the network generates fault samples with key time-frequency features.

[0178] Fault diagnosis under imbalanced samples: To verify the fault classification accuracy after data augmentation using the GAMICGAN network at various imbalance ratios, the classification accuracy was compared with that after augmentation using the ICGAN network. The fault classification accuracy results for each imbalance ratio are shown in Table 12 and... Figure 24 As shown.

[0179] Table 12

[0180]

[0181] The results above show that as the imbalance ratio increases, the augmentation effect of GAMICGAN is significantly improved compared to that of ICGAN. This indicates that the GAMICGAN generator network further enhances its attention to the fault time-frequency features in the vibration signal sample time-frequency map when generating fault sample time-frequency maps. By introducing a global attention mechanism on the basis of the ICGAN network, the generator network can fully pay attention to the spatial, channel, and dimensional information in the time-frequency map. By replacing the traditional convolution with octave convolution, the generator network can extract the high-frequency and low-frequency component information of the time-frequency map during the sample generation process, thereby generating a higher quality two-dimensional fault sample time-frequency map.

[0182] Variable-condition bearing fault diagnosis based on joint distributed domain adversarial neural network: To verify the performance of the variable-condition bearing fault diagnosis method based on joint distributed domain adversarial neural network proposed in this section, variable-condition diagnosis experiments were conducted on the CWRU bearing dataset. Ablation experiments were designed to analyze the role of each component in the variable-condition bearing fault diagnosis model. The superiority and effectiveness of the proposed method were verified by comparing it with other variable-condition diagnosis models.

[0183] Ablation experiments for fault diagnosis under varying operating conditions: Based on the DANN model, a lightweight global attention residual neural network (GAMLRN) is used as the feature extraction network, while a graph convolutional network (GCN) is input. The JMMD metric is employed to reduce the distributional differences between the two domains. To verify the effectiveness of each module in improving fault diagnosis accuracy, the following ablation model, AD, is designed:

[0184] (1) Model A: Only the DANN model is used, where the feature extractor uses the CNN model in Chapter 3, that is, domain adaptation is achieved only through the adversarial training method between the two domains.

[0185] (2) Model B: Replace the feature extraction network in Model A with the feature extraction network GAMLRN in this paper.

[0186] (3) Model C: The transfer method is performed using the Joint Maximum Mean Difference (JMMD) metric, with the feature extraction network being GAMLRN.

[0187] (4) Model D: The variable working condition fault diagnosis model JDAGCN proposed in this paper uses JMMD to measure the distribution difference between the source domain and the target domain based on the DANN model, and adopts GAMLRN as the feature extraction network.

[0188] Twelve domain migration tasks were designed, as shown in Table 13, where 0HP->1HP represents the migration diagnosis from the source domain to the target domain. Ablation experiments were conducted using... Figure 25 The diagnostic process shown is performed with 300 training and testing rounds, a batch size of 50, an Adam optimizer, a learning rate of 0.0002, and hyperparameters λ1 and λ2 both set to 1.

[0189] The accuracy of the variable working condition fault diagnosis results of the CWRU bearing dataset ablation experiment is shown in Table 14.

[0190] Table 14

[0191]

[0192] The iterative accuracy curves of each model for transfer diagnostic tasks under different operating conditions are shown below. Figure 27 As shown.

[0193] The results above show that:

[0194] (1) When the working environment conditions change little, such as the transfer diagnosis task from working condition 0 to working condition 1, all models show good cross-working condition transfer diagnosis effect. When the working environment changes significantly, such as the transfer diagnosis task from working condition 0 to working condition 3, the transfer diagnosis accuracy of models A, B and C all decrease to a certain extent.

[0195] (2) Compared with Model A, when using the DANN model for two-domain transfer, Model B shows a significant improvement in convergence stability and transfer diagnosis accuracy after replacing the CNN feature extraction network with GAMLRN, indicating that the deeper model improves generalization ability.

[0196] (3) Compared with model B, model C converges faster and the model iteration process is more stable, but the transfer accuracy is lower, indicating that the adversarial training process of feature extraction network and domain discriminator is not stable when performing transfer diagnosis task based on DANN model.

[0197] (4) Model D combines the advantages of Model B and Model C. Based on the DANN model, it adopts the JMMD metric method to consider the difference between the marginal distribution and conditional distribution between the two domains, which improves the performance of the feature extractor and makes the model more stable, converges faster and has a higher transfer accuracy.

[0198] The t-SNE distribution results for the migration from operating condition 0 to operating condition 3 for each model are as follows: Figure 26 As shown.

[0199] from Figure 26 It can be seen that when performing transfer diagnostic tasks based on the DANN model, adversarial training maps the common feature representations of the source and target domains to the same high-dimensional space, resulting in a large class distance between different classes. This paper adds the JMMD metric method to the DANN model, enabling the model to notice intra-class differences between the same class during the feature extraction stage. While maintaining the distance between different classes, it achieves intra-class aggregation very well.

[0200] 4.3.2 Comparative Experiment on Fault Diagnosis under Variable Operating Conditions

[0201] To verify the superiority of the JDAGCN model proposed in this chapter on bearing fault diagnosis under varying operating conditions, comparative experiments were conducted using different diagnostic models to assess the accuracy of fault diagnosis under different operating conditions. Validation was performed on the CWRU bearing dataset, and the fault diagnosis accuracy of each model under different operating conditions was obtained. Table 153 shows the fault diagnosis accuracy of each model. The details of each comparative experiment used the same settings as the ablation experiment. The selected comparative experimental models are as follows:

[0202] (1) CNN. The CNN classification model in Chapter 3 is used for variable operating condition diagnosis. The fault samples in the source domain are input into the CNN model for training, and the trained CNN model is used to classify the fault samples in the target domain.

[0203] (2) MMD. The MMD metric method is used to perform transfer diagnostic tasks under different working conditions, where the feature extraction network adopts CNN.

[0204] (3) DAN. DAN is a commonly used deep transfer learning method that uses multi-core MMD and CNN for transfer diagnostic tasks.

[0205] The accuracy of fault diagnosis results under varying operating conditions in the CWRU bearing dataset comparison experiment is shown in Table 15.

[0206] Table 15

[0207]

[0208] Analysis of the fault classification accuracy of various diagnostic models under different operating conditions yields the following conclusions:

[0209] (1) Under the same migration conditions, the fault diagnosis accuracy of directly using CNN for variable working condition migration is the lowest. The greater the difference in working conditions, the more the fault diagnosis accuracy drops. As shown in Table 15, when migrating from working condition 0 to working condition 3, the fault diagnosis accuracy of the target domain is only 80.00%, indicating that the simple classification model cannot extract the deep features of bearing faults. Under different working conditions, the model does not take into account the differences in data distribution between different domains, resulting in overfitting and poor generalization ability in the target domain.

[0210] (2) When using MMD and DAN and other variable working condition transfer models to classify bearing faults under different working conditions, the fault classification accuracy was significantly improved. Compared with the CNN non-transfer model, when performing fault transfer diagnosis with large working condition differences such as working condition 0 to working condition 3, the fault classification accuracy of the target domain reached more than 95%, which shows the effectiveness of the transfer mechanism in performing variable working condition fault diagnosis.

[0211] (3) By adding the JMMD joint distribution difference measure method to the DANN transfer model, the JDAGCN model further improves the accuracy of variable working condition diagnosis. The fault classification accuracy reaches 100% on most variable working condition transfer diagnosis tasks, indicating that the JDAGCN model has excellent robustness and generalization ability in variable working condition bearing fault diagnosis tasks. It can better capture the data distribution differences between different domains and achieve better feature representation through adversarial training.

[0212] Under the transition task from operating condition 0 to operating condition 3, deep feature extraction is performed on different comparison models, and the t-SNE algorithm is used to map the high-dimensional features to a two-dimensional space. The resulting t-SNE distribution is as follows: Figure 27 As shown.

[0213] from Figure 27 It can be observed that, compared to model A, models B and C, which incorporate a domain transfer mechanism, exhibit larger inter-class distances and smaller intra-class distances in their deep features within two-dimensional space. This means that models B and C can better distinguish features between different bearing states, thus significantly improving the problem of inter-class confusion. In the JDAGCN model, the separability of its extracted deep features is further improved, and inter-class confusion is reduced, indicating its better ability to solve multi-domain fault diagnosis problems. In summary, compared to the other three control models, the deep features extracted by the JDAGCN model have better cross-domain invariance, enabling it to more accurately distinguish different bearing operating states when solving complex multi-domain fault diagnosis problems.

[0214] Figure 28 These are the confusion matrices of the various comparative models under operating conditions 0 to 3, from... Figure 28 It can be seen that the accuracy of transfer diagnosis in the target domain is very poor when directly using a CNN model for transfer diagnosis. While feature transfer diagnosis methods using MMD and DAN significantly improve the performance of transfer diagnosis tasks under different operating conditions, they still have some misjudgments for certain types of bearing faults. The JDAGCN variable operating condition bearing fault diagnosis method proposed in this paper effectively achieves transfer diagnosis tasks under different operating conditions.

[0215] (3) Fault diagnosis experiment under unbalanced sample and variable operating conditions

[0216] To verify the diagnostic performance under imbalanced conditions, this section first divides the CWRU bearing dataset into sample sets with different balance ratios. Then, the GAMICGAN generative network proposed in section 3.2 is used to augment the datasets with different imbalance ratios. The JDAGCN model in this section is used to conduct cross-condition diagnostic experiments. The obtained bearing variable condition diagnostic classification accuracy is shown in Table 16.

[0217] Table 16

[0218]

[0219] As can be seen from Table 16, the accuracy of bearing fault classification under variable operating conditions decreases faster as the proportion of imbalanced samples increases. This illustrates the importance of expanding the imbalanced dataset into a balanced dataset during the variable operating condition diagnosis process. Furthermore, it can be seen that under the same imbalance proportion, the fault classification accuracy under variable operating conditions is not as good as that under the same operating conditions, which illustrates the importance of data distribution differences during the variable operating condition migration process.

Claims

1. A fault diagnosis method based on auxiliary generation networks and domain adversarial graph networks, characterized in that: Specifically, it includes the following steps; Step 1: Targeting the non-stationary and nonlinear characteristics of vibration signals, a time-frequency analysis method based on wavelet analysis is used to preprocess the one-dimensional original vibration signal using continuous wavelet transform to obtain a two-dimensional wavelet time-frequency diagram that more effectively expresses the fault mechanism. The time-frequency diagram is used as the processing object of the generator network and diagnostic model. An independent auxiliary classifier structure is designed for the ACGAN network, and a weight adjustment mechanism for the discriminator and auxiliary classifier is introduced to obtain the improved generator network model ICGAN, thereby improving the generation quality of the generator network for samples of the specified category. Step 2: Further optimize the ICGAN generative network structure using global attention and octave convolution; GAMICGAN, a fault sample time-frequency map generation network based on global attention and octave convolution, introduces the global attention mechanism GAM to make the generative network pay more attention to the spatial, channel, and dimensional information in the time-frequency map; replace the traditional two-dimensional convolution in the network with octave convolution OctConv to extract high and low frequency component information in the time-frequency map; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of generated samples; Step 3: To address the issue of decreased diagnostic model accuracy caused by differences in sensor data distribution under varying operating conditions, we combine Graph Convolutional Network (GCN) with GAMLRN, a backbone network structure based on Global Attention Residual Network, to extract the depth features of the time-frequency map of vibration signal samples. This paper utilizes the adversarial transfer mechanism of domain adversarial neural networks to align fault features under different operating conditions. Simultaneously, it introduces a joint maximum mean difference metric to quantify the edge distribution differences and conditional distribution differences between the source and target domains, thereby optimizing the backbone network parameters and further reducing the distribution differences between the time-frequency map depth features of samples under different operating conditions. Furthermore, it employs graph convolutional network (GCN) embedding theory to construct structural relationships between samples in the same / different domains, further reducing the distribution differences between the source and target domains and thus further minimizing the distribution differences between the time-frequency map depth features of samples under different operating conditions.

2. The fault diagnosis method based on auxiliary generation network and domain adversarial graph network according to claim 1, characterized in that: In step 1, the problem of sample imbalance is addressed from the perspective of data generation. Based on the ACGAN generation network, a fault sample time-frequency map augmentation method, ICGAN, is proposed. The continuous wavelet transform method is used to convert the one-dimensional vibration signal of the equipment into a two-dimensional time-frequency image with rich time-frequency features. A modified conditional generative network ICGAN is proposed, the ACGAN network structure is improved, and an independent auxiliary classifier is designed to have independent parameters and weights. The two-dimensional time-frequency graph is input into the ICGAN network for data generation, and the quality of the generated samples is evaluated using the FID evaluation metric. The qualified generated samples are then added to the imbalanced dataset of the original samples to improve the classification performance of the subsequent fault diagnosis model. The specific steps include: Step 1.1, constructing a time-frequency image from the vibration signal: Vibration signal segmentation: The obtained one-dimensional vibration signal is segmented into multiple sample sets of the same sequence length using a sliding window method. During the sliding window selection process, to ensure that the segmented vibration signal has signal characteristics of at least one rotation cycle, the number of data points in each sequence sample must meet the following condition: （3-1） in, This indicates the number of data points in each vibration signal sample, i.e., the sample length; It represents the number of rotations of each vibration signal sample from the mechanical equipment. It is the sampling frequency when collecting vibration signals, representing the number of data points recorded per second; It is the rotational speed of mechanical equipment, representing the number of times the mechanical equipment rotates per minute; Two-dimensional time-frequency plot construction: The one-dimensional vibration signal generated when mechanical equipment fails contains a large amount of redundant information and noise interference unrelated to the fault characteristics. Continuous wavelet transform is used to denoise and filter the original vibration signal, transforming the one-dimensional signal into a two-dimensional time-frequency plot. During signal analysis and processing, wavelet decomposition helps reduce redundant information in the original time-series vibration signal, allowing the model to focus on the key fault characteristics that generate the vibration signal, thereby improving the model's accuracy and generalization ability. On the other hand, the two-dimensional time-frequency plot not only retains the time-domain information of the original vibration signal but also visualizes the energy distribution of the signal at different frequencies, clearly showing the frequency domain characteristics of the vibration signal. This makes it easier for the classification model to learn the specific category feature representation in the time-frequency plot, so as to distinguish between different fault categories. The two-dimensional time-frequency plot obtained through wavelet transform can better reflect the spectral characteristics of the vibration signal of mechanical equipment under different operating conditions. The original signal is subjected to continuous wavelet transform using the Bump wavelet, and the resulting wavelet coefficient matrix is ​​transformed into a two-dimensional time-frequency feature map. The shape of the Bump wavelet function is similar to a strip wave packet, possessing finite time and finite frequency characteristics. It has certain advantages in focusing on the local and instantaneous characteristics of a signal. It is finite and localized in the time domain, consistent with the characteristics of impulse signals. Furthermore, due to its localization in the frequency domain, the Bump wavelet can more effectively capture transient and impulse components in the signal. The Bump wavelet function is expressed as: （3-2） Two-dimensional time-frequency images obtained by converting one-dimensional vibration signals using the Bump wavelet function; Step 1.2, Mechanical Equipment Fault Sample Augmentation Method Based on ICGAN: By improving the ACGAN network, a generative adversarial network ICGAN based on independent classifiers is proposed; by designing independent auxiliary classifier structures for the ACGAN network and assigning different weights to the network parameters of the discriminator and auxiliary classifiers, the quality of generated fault samples is improved; by designing the auxiliary classifiers in ACGAN as independent structures and giving the network independent classifiers, the classification performance is enhanced, resulting in higher quality samples of specific categories generated by the generator; The generator generates samples similar to the real sample time-frequency maps based on random noise and label vectors. The discriminator distinguishes between the generated sample time-frequency maps and the real sample time-frequency maps to optimize the quality of the time-frequency map samples generated by the generator. By adopting an independent classifier design structure, the generated sample time-frequency maps in each category are more similar to the real sample time-frequency maps and better match the distribution of the real sample time-frequency maps. The optimal network model parameters are determined using the generated sample evaluation index FID and loss curve visualization technology. The network structure of the generator in ICGAN is as follows: The input dimension is 110, of which 100 dimensions are noise signals and 10 dimensions are label information. The input noise is transformed through a fully connected layer and then passed through 5 layers of 2D deconvolution layers to obtain a 3-channel 128-size feature map. The discriminator network structure in ICGAN is as follows: Five 2D convolutional layers are used to transform the 3-channel 128-size feature map obtained by the generator into a 52-channel 13-size feature map. Then, a fully connected layer outputs 0 / 1 to represent the discriminative output of the generated sample. The network structure of the independent classifier: The independent classifier classifies the feature map output by the generator and the image of the real sample, and backpropagates the classification result to further optimize the generator's generation process and improve the generation quality of each class of generated samples; Step 1.3, ICGAN-based expansion of mechanical equipment fault samples: Data acquisition and processing: Vibration signals of the equipment under various working conditions and different fault types are collected, and the original vibration signals are segmented using a sliding window. Continuous wavelet transform is used to convert the one-dimensional signal into a two-dimensional wavelet time-frequency diagram, thereby constructing an unbalanced training sample set and a test sample set. Generative Network Training and Sample Evaluation: A small number of training samples are fed into the ICGAN generative network for iterative training. The optimal network model parameters are determined using the generated sample evaluation index FID and loss curve visualization techniques. Imbalanced sample augmentation: Random noise and label vectors are input into the trained ICGAN generator network to generate time-frequency maps of multiple fault types, thus expanding the original imbalanced training sample set.

3. The fault diagnosis method based on auxiliary generation network and domain adversarial graph network according to claim 1, characterized in that: In step 2, the ICGAN optimization method based on global attention and octave convolution is as follows: Step 2.1, Model Overview: ICGAN is further optimized with the development of GAMICGAN, a fault sample time-frequency map generation network based on global attention and octave convolution. By introducing the global attention mechanism GAM, the generation network pays more attention to the spatial, channel, and dimensional information in the time-frequency map. OctConv, an octave convolution, replaces the traditional two-dimensional convolution in the network to extract high- and low-frequency component information from the time-frequency map. By combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency maps with key fault features, further improving the quality of the generated samples. Step 2.2, GAMICGAN network structure: To improve the quality of two-dimensional time-frequency images generated by the ICGAN generator network and increase the generator network's attention to the fault time-frequency features in the two-dimensional time-frequency image of the vibration signal, a mechanical equipment fault diagnosis method GAMICGAN based on global attention and octave convolution is proposed. By embedding a global attention mechanism in the generator and discriminator of ICGAN, the generative network model can fully consider the spatial, channel and dimensional information in the time-frequency map when generating time-frequency maps of fault samples. By replacing the ordinary transposed convolution and ordinary convolution of the generator and discriminator in ICGAN with octave transposed convolution DeOctConv and octave convolution OctConv respectively, the network can pay attention to the high and low frequency components of the time-frequency map. By introducing the global attention mechanism and octave convolution, the generation quality of the time-frequency map of the fault sample is effectively improved, and the accuracy of mechanical equipment fault diagnosis under sample imbalance is further improved. Discriminator Network: In the discriminator, the input image is downsampled using traditional convolution, and the resulting feature map is passed through the global attention module GAMATtention to focus on the channel, spatial and dimensional information in the feature map. Three OctConv modules are used to extract the high and low frequency component features in the feature map. Traditional convolution and fully connected layers are used to distinguish between real and fake samples. The GAMAttention global attention module consists of two main parts: channel attention and spatial attention. Channel attention determines the weights of each channel, enabling the model to focus on key features in the input data. It learns the weights between channels to focus on task-related channels. By leveraging the complex dependencies between channels, it effectively captures cross-channel interactions and improves the feature representation of specific faults, thus enabling the generative network to generate high-quality sample images. The channel attention module preserves channel information in all three dimensions of the image. It performs dimensionality transpose on the input feature map and then uses the encoding and decoding structure of a multilayer perceptron (MLP) to achieve a compression ratio of [insert compression ratio here]. ,Will The value is set to 4 to amplify the interaction between channels and space in the dimensions. The feature map is then dimensionally inverted and activated by the Sigmoid function to obtain the feature map after channel attention processing. Spatial attention mechanisms consider the relationships between image pixels, enabling the model to selectively focus on task-related information. It learns the weights of each location to represent the importance of each position in the image, thereby improving the model's ability to represent key image features. The feature map processed by channel attention is multiplied with the input feature map and then fused using a 7×7 convolutional layer to better capture the local structure of the image. The same compression ratio as channel attention is used. To further compress the dimension of the feature map; then pass it through a 7×7 convolutional layer to further extract spatial features; then pass it through a Sigmoid activation function to obtain a feature map after spatial attention processing; finally, multiply the feature map after spatial attention processing with the feature map after channel feature map processing element-wise to obtain the final feature map output. OctConv octave convolution module: Divides the input feature map into high-frequency feature maps using OctConv. and low-frequency feature maps These represent the high-frequency and low-frequency components of the input feature map, respectively; OctConv uses... and Control the proportion of high-frequency and low-frequency components in the input and output channels. The value range is [0,1]; in the OctConv convolutional layer, except for the first and last layers, The values ​​are all set to It is 0.

5. The value is 0.5 to facilitate the extraction of high and low frequency feature information; the discriminator's network structure uses a three-layer OctConv module to update high and low frequency information; The output features represent information updates between high-frequency components. This indicates the information update output features between low-frequency components; and This represents the information exchange output characteristics between different frequency components; the implementation method is as follows: （3-3） （3-4） （3-5） （3-6） in, This indicates a pooling operation with a step size of 2. Upsampling is performed using nearest neighbor interpolation with a step size of 2; The first layer of OctConv will Set to 0, Setting it to 0.5 decomposes the input feature map into two feature map components: high-frequency and low-frequency. and Specifically, it is expressed as: （3-7） （3-8） OctConv second layer, will and All values ​​are set to 0.5, and the high-frequency feature map obtained from the first layer is... and low-frequency feature maps The fusion yields the following new high- and low-frequency feature maps: （3-9） （3-10） OctConv third layer, will Set to 0, Setting it to 0.5 means that the high-frequency feature map components obtained from the second layer... The final feature map output is obtained at this point. for: （3-11） Generator Network Structure: In the generator, random noise with label information is passed through a fully connected layer to obtain a one-dimensional vector. This vector is first subjected to a traditional deconvolution to obtain a feature map. This feature map is then enlarged in channel and size by three DeOctConv layers. Global attention is used to obtain a feature map of the same size. Finally, a traditional convolution is used to obtain the generator output. Unlike the discriminator OctConv module, the generator incorporates DeOctConv instead of ordinary transposed convolution. This represents a regular transpose convolution.

4. The fault diagnosis method based on auxiliary generation network and domain adversarial graph network according to claim 1, characterized in that: In step 3, the fault diagnosis of mechanical equipment under varying operating conditions based on the joint distributed domain adversarial graph convolutional neural network is as follows: A lightweight feature extraction module, GAMLRN, based on a global attention residual network, is used to extract deep features from time-frequency maps. Adversarial transfer mechanisms of domain adversarial neural networks are utilized to align fault features under different operating conditions. A joint maximum mean difference metric is introduced to quantify the edge distribution differences and conditional distribution differences between the source and target domains. A graph convolutional network (GCN) is used to construct structural relationships between samples within the same domain and between samples from different domains, further reducing the distribution differences between the source and target domains. Step 3.1, GAMLRN feature extraction network design: A lightweight feature extractor based on a global attention residual network, GAMLRN, is used. In the GAMLRN network, a 3×3 convolutional kernel is used, the convolution stride is set to 1, the padding value is set to 1 or 0, only three residual structures are used, the pooling operation of the intermediate layer is canceled, and the deep features in the time-frequency map are extracted by embedding global attention. Step 3.2, Network framework for fault diagnosis under varying operating conditions: The JDAGCN fault diagnosis model structure, which combines distributed domain adversarial neural networks (DCNs), consists of data preprocessing, a feature extractor (GAMLRN), a graph embedding theory (GCN), and a label classifier. Domain discriminant composition; The data preprocessing module processes data from both the source and target domains. The feature extractor GAMLRN uses a lightweight global attention residual neural network model to extract deep features from both the source and target domains. The label classifier consists of fully connected layers to classify the data from the source domain. The domain discriminator uses fully connected layers to distinguish whether the data comes from the source or target domain. In the domain adaptation stage, the JMMD combined with GCN metric method is used to reduce the differences in marginal distribution and conditional distribution between the source and target domains. The optimization objective of the JDAGCN model is determined by the label classification loss. Domain-based loss discrimination and JMMDGC distribution difference loss The structure is as follows: The final target loss function is obtained. : （3-12） in, and These are all hyperparameters; This is used to balance the label classification loss and the domain discrimination loss. At the same time, the reverse gradient is used to negatively value the weight of the domain discrimination loss, so that the model can reduce the label classification loss while maximizing the domain discrimination loss. This domain adversarial training method enables the model to learn deeper feature representations and improve the model's generalization ability in the target domain. This is used to balance the importance of label classification loss and JMMD loss, so that the model pays attention to the differences in the distribution of joint features between the source and target domains; Step 3.3, Variable Operating Condition Fault Diagnosis Process: This consists of the following four steps: Step 3.31, Data Preprocessing: Obtain the original vibration signals from the source and target domains, and perform continuous wavelet transform on the vibration signals to obtain the corresponding time-frequency map samples. Then, input the two-dimensional time-frequency images of different fault types into the fault sample time-frequency map generation network GAMICGAN based on global attention and octave convolution. By mixing the generated fault category sample data of specific types with the original sample set, the purpose of expanding the imbalanced dataset is achieved, thereby improving the fault diagnosis accuracy under imbalanced samples. Step 3.32, Feature Extraction: After setting the model parameters of JDAGCN, the obtained source and target domain time-frequency map samples are input into GAMLRN to extract deep features of the source and target domains simultaneously. Step 3.33, Domain Adaptation: Calculate the overall loss function, optimize the label classifier loss, domain discriminator loss, and joint maximum mean difference loss through backpropagation; and before the maximum number of rounds, extract deep features of the source and target domains in a cyclical manner through domain adversarial methods in order to obtain the same feature representation between the two domains; Step 3.34, Diagnosis and Classification: After obtaining the trained JDAGCN model, input the time-frequency map of the fault samples in the target domain into the JDAGCN model to obtain the final target domain classification result.

Images