Fault diagnosis method based on auxiliary generative network and domain adversarial graph network

Abstract

The invention discloses a fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network, belongs to the technical field of fault diagnosis, and provides a sample generation network ICGAN based on ACGAN for the problem of sample imbalance, so as to improve the generation quality of the generation network for a specified category of samples. Secondly, designing a fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution, enabling the generation network to pay more attention to space, channel and dimension information in a time-frequency graph by introducing a global attention mechanism (GAM), and further optimizing a generation network structure; and finally, aiming at the problem that the accuracy of a diagnosis model is reduced due to sensing data distribution difference in a variable working condition scene, a fault domain adaptive diagnosis method GIC-JDAGCN of a joint distribution domain adversarial network is provided by combining a graph convolutional network GCN, and a result shows that the GIC-JDAGCN can keep relatively high diagnosis accuracy in variable working condition application, so that the fault domain adaptive diagnosis method can be applied to the variable working condition scene. And the method has good working condition scene adaptability.

Description

Technical Field

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

[0002] In recent years, with the rapid development of society, rotating machinery has been widely used in energy conversion, industrial production, transportation, aerospace and national defense. Generally speaking, rotating machinery is composed of main structure, rotating parts, power transmission system and control system, which constitute a complex and tight system engineering. The damage of one part may affect other parts, resulting in abnormal operation of rotating machinery and equipment, causing certain economic losses [2]. The core components of rotating machinery are rotating parts, including shafts, bearings, rotors, etc. [3]. In harsh industrial production environments, long-term vibration and impact or high temperature and high pressure will cause wear and cracks on 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, the study of intelligent fault diagnosis methods is of great significance to industrial production. With the development of science and technology, the use of technologies such as intelligent manufacturing and the Internet of Things 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. At present, data-driven mechanical equipment fault diagnosis 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 are often carried out under the condition that there are sufficient training data samples for various fault states. However, in actual industrial production environments, the collection of various sensor data has been well solved, but effective data labeling is still difficult. In addition, since equipment is often in a relatively normal working state for a long time, sensor data under various fault states is difficult to obtain. Therefore, the data samples collected at industrial sites have problems of imbalanced state types and incomplete labeling. When using an unbalanced data set to train a fault diagnosis model, the classifier is more likely to lean towards the category with more samples, resulting in low recognition accuracy of the fault category by the fault diagnosis model and reduced generalization ability of the model.

[0004] Rotating machinery and equipment usually operate in different working environments and conditions, and the probability distribution of the equipment sensor data collected under different working conditions varies, resulting in poor diagnostic performance of the diagnosis model under variable working conditions. Therefore, when diagnosing mechanical equipment faults in actual industrial sites, it is necessary to consider how to make the extracted sensor data features less affected by working condition changes, which poses a challenge to traditional fault diagnosis methods. In recent years, transfer learning has solved this problem by using knowledge from related fields to the target field, mapping data with different distributions to a unified feature space, thereby solving the problem of inconsistent distribution between the source domain and the target domain, and improving fault diagnosis performance and adaptability to working conditions. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network in response to the shortcomings of the background technology; it effectively solves the problem of imbalance in training sample categories in mechanical equipment fault classification diagnosis tasks, and improves the adaptability of the diagnosis model to variable working conditions.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

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

[0008] Step 1: Aiming at the non-stationary and nonlinear characteristics of vibration signals, a vibration signal time-frequency analysis method based on wavelet analysis is used to pre-process 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 generation network and the diagnostic model, and an independent auxiliary classifier structure is designed for the ACGAN network. The weight adjustment mechanism of the discriminator and the auxiliary classifier is introduced to obtain an improved generation network model ICGAN, thereby improving the generation quality of the generation network for samples of the specified category;

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

[0010] Step 3. To address the problem of decreased accuracy of the diagnostic model caused by differences in sensor data distribution under variable operating conditions, the graph convolutional network (GCN) is combined with the backbone network structure GAMLRN based on the global attention residual network to extract the deep features of the time-frequency graph of the vibration signal samples. The adversarial transfer mechanism of the domain adversarial neural network is used to align the fault features under different operating conditions. At the same time, the marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified by introducing the joint maximum mean difference metric to optimize the backbone network parameters and further reduce the distribution difference between the deep features of the time-frequency graphs of samples under different operating conditions. At the same time, the graph convolutional network (GCN) embedding theory is used to construct the structural relationship between samples in the same / different fields to further reduce the distribution difference between the source domain and the target domain, and further reduce the distribution difference between the deep features of the time-frequency graphs of samples under different operating conditions.

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

[0012] 1. Aiming at the non-stationary and nonlinear characteristics of vibration signals, a vibration signal time-frequency analysis method based on wavelet analysis is designed. The one-dimensional original vibration signal is preprocessed 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 generation network and the diagnosis model, and an independent auxiliary classifier structure is designed for the ACGAN network. The weight adjustment mechanism of the discriminator and the auxiliary classifier is introduced to obtain an improved generation network model ICGAN, so as to improve the generation quality of the generation network for the specified category samples; the present invention improves ACGAN, designs an independent classifier structure, and introduces the weight adjustment mechanism of the discriminator and the auxiliary classifier to obtain an improved generation network model ICGAN;

[0013] 2. In order to further improve the generation network's attention to the fault time-frequency features in the vibration signal sample time-frequency graph, the present invention uses global attention and octave convolution to further optimize the generation network structure of ICGAN; designs a fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution, and introduces the global attention mechanism GAM to make the generation network pay more attention to the space, channel and dimensional information in the time-frequency graph; uses the octave convolution OctConv to replace the traditional two-dimensional convolution in the network to extract the high and low frequency component information in the time-frequency graph; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency graphs with key fault features, further improving the quality of generated samples; further improves ICGAN, and introduces the global attention mechanism GAM to make the generation network pay more attention to the space, channel and dimensional information in the time-frequency graph; uses the octave convolution OctConv to replace the traditional two-dimensional convolution in the network to extract the high and low frequency component information in the time-frequency graph, and obtains the improved generation network model GAMICGAN;

[0014] 3. Aiming at the problem of decreased accuracy of the diagnostic model caused by the difference in the distribution of sensor data under variable working conditions, the present invention proposes a mechanical equipment fault domain adaptive diagnosis method JDAGCN based on a joint distribution domain adversarial neural network in combination with a graph convolutional network (GCN); designs a backbone network structure GAMLRN based on a global attention residual network to extract the deep features of the time-frequency graph of vibration signal samples; uses the adversarial migration mechanism of the domain adversarial neural network to align the fault features under different working conditions; at the same time, by introducing the metric method of the joint maximum mean difference, the marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified to optimize the backbone network parameters, further reducing the distribution difference between the deep features of the sample time-frequency graph under different working conditions; at the same time, the graph convolutional network (GCN) embedding theory is used to construct the structural relationship between samples in the same / different fields, so as to further reduce the distribution difference between the source domain and the target domain, and further reduce the distribution difference between the deep features of the sample time-frequency graph under different working conditions; experimental results show that the mechanical equipment fault diagnosis method based on the joint distribution domain adversarial neural network can maintain a high diagnostic accuracy in variable working condition applications and has good adaptability to working condition scenarios;

[0015] Based on the expansion of training samples by the generative network, the present invention combines the deep transfer learning framework to establish a pre-adaptive diagnosis model for variable working condition scenarios. (1) The backbone network structure GAMLRN based on the global attention residual network is designed to extract the deep features of the time-frequency graph of the vibration signal sample; (2) The adversarial transfer mechanism is used to achieve feature alignment of fault samples under different working conditions; (3) The joint maximum mean difference metric is introduced to quantify the marginal distribution difference and conditional distribution difference between the source domain and the target domain; (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 domain and the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 It is a principle flow chart of a fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network of the present invention;

[0018] Figure 2 It 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 the present invention;

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

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

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

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

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

[0025] Fig. 9 It is a schematic diagram of the network structure generated by GAMICGAN of the present invention;

[0026] Fig.10 It is the GAMICGAN discriminator network structure of the present invention;

[0027] Fig.11 is the global attention network structure of the present invention;

[0028] Fig.12 It is the GAMICGAN generator network structure of the present invention;

[0029] Fig.13 It is the GAMLRN network structure of the present invention;

[0030] Fig.14 It is the JDAGCN mechanical equipment fault migration diagnosis model of the present invention;

[0031] Fig.15 This is the JDAGCN variable operating condition diagnosis flow chart of the present invention;

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

[0033] Fig.16 (b) is the iterative output of the discriminator of the present invention;

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

[0035] Fig.17 (b) is the iterative output of the discriminator of the present invention;

[0036] Fig.18 It is the time-frequency diagram of the fault samples generated by ICGAN of the present invention;

[0037] Fig.19 is the FID value in the iterative process of the present invention;

[0038] Fig. 20 It is a comparison of the classification results of different generation networks of the present invention under various imbalance ratios;

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

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

[0041] Fig. 22 It is a time-frequency diagram of fault samples generated by GAMICGAN of the present invention;

[0042] Fig.23 This is the comparison of the thermal maps of ICGAN and GAMICGAN of the present invention;

[0043] Fig.24 This is the accuracy comparison on the unbalanced data set of the present invention;

[0044] Fig.25 It is the accuracy curve of each comparison model under different working condition migration tasks of the present invention;

[0045] Fig.26 It is the t-SNE distribution diagram of each model of the ablation experiment of the present invention under working conditions 0 to 3;

[0046] Fig. 27 t-SNE diagram of each model in the comparative experiment of the present invention when migrating from working condition 0 to working condition 3;

[0047] Fig.28 It is the confusion matrix of each model in the comparative experiment of the present invention when migrating from working condition 0 to working condition 3. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The present invention is described in detail below based on the drawings and preferred embodiments, and the purpose and effect of the present invention will become clearer. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0050] A fault diagnosis method based on auxiliary generation network and domain adversarial graph network, such as Figure 1 As shown, the specific steps include the following:

[0051] Step 1. Aiming at the non-stationary and nonlinear characteristics of vibration signals, a vibration signal time-frequency analysis method based on wavelet analysis is used. The one-dimensional original vibration signal is preprocessed 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 generation network and the diagnostic model. An independent auxiliary classifier structure is designed for the ACGAN network, and a weight adjustment mechanism for the discriminator and the auxiliary classifier is introduced to obtain an improved generation network model ICGAN, thereby improving the generation quality of the generation network for samples of the specified category; ACGAN is improved, an independent classifier structure is designed, and a weight adjustment mechanism for the discriminator and the auxiliary classifier is introduced to obtain an improved generation network model ICGAN.

[0052] Step 2, using global attention and octave convolution to further optimize the generation network structure of ICGAN; based on the global attention and octave convolution fault sample time-frequency graph generation network GAMICGAN, by introducing the global attention mechanism GAM, the generation network pays more attention to the space, channel and dimensional information in the time-frequency graph; using the octave convolution OctConv to replace the traditional two-dimensional convolution in the network, extracting the high and low frequency component information in the time-frequency graph; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency graphs with key fault features, further improving the quality of generated samples; the present invention improves ICGAN, by introducing the global attention mechanism (GAM), the generation network pays more attention to the space, channel and dimensional information in the time-frequency graph; using the octave convolution (OctConv) to replace the traditional two-dimensional convolution in the network, extracting the high and low frequency component information in the time-frequency graph, and obtaining the improved generation network model GAMICGAN.

[0053] Step 3: To address the problem of decreased accuracy of the diagnostic model caused by differences in sensor data distribution under variable operating conditions, the deep features of the time-frequency graph of vibration signal samples are extracted by combining the graph convolutional network (GCN) and the backbone network structure GAMLRN based on the global attention residual network. The adversarial transfer mechanism of the domain adversarial neural network is used to align fault features under different operating conditions. At the same time, the marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified by introducing the joint maximum mean difference metric to optimize the backbone network parameters and further reduce the distribution difference between the deep features of the time-frequency graphs of samples under different operating conditions. At the same time, the structural relationship between samples in the same / different domains is constructed using the graph convolutional network (GCN) embedding theory to further reduce the distribution difference between the source domain and the target domain, and further reduce the distribution difference between the deep features of the time-frequency graphs of samples under different operating conditions. Based on the expansion of the training samples by the generative network, a pre-adaptive diagnostic model for variable operating conditions is established in combination with the deep transfer learning framework. (1) A backbone network structure GAMLRN based on the global attention residual network is designed to extract the deep features of the time-frequency graph of vibration signal samples; (2) The adversarial transfer mechanism is used to align the features of fault samples under different working conditions; (3) The joint maximum mean difference metric is introduced to quantify the marginal distribution difference and conditional distribution difference between the source domain and the target domain; (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 domain and the target domain.

[0054] Method for generating network fault samples based on ICGAN

[0055] When there is sufficient data and balanced samples, the data-driven artificial intelligence fault diagnosis method can accurately identify the health status of the equipment. However, in the actual operation of the equipment, the vibration signals accumulated by the field data acquisition system for a long time are mainly normal signals, and the acquisition of equipment fault signals is more difficult. There are problems such as insufficient fault samples, single fault category, and unbalanced samples, which affect the recognition accuracy of the equipment fault diagnosis method.

[0056] This section solves the sample imbalance problem from the perspective of data generation. On the basis of the ACGAN generation network, a fault sample time-frequency map expansion method ICGAN based on the improved ACGAN network is proposed. The one-dimensional vibration signal of the equipment is converted into a two-dimensional time-frequency image with rich time-frequency features by using the continuous wavelet transform method. A method based on the improved conditional generation network ICGAN is proposed to improve the ACGAN network structure and design an independent auxiliary classifier so that it has independent parameters and weights. The two-dimensional time-frequency map is input into the ICGAN network for data generation, and the quality of the generated samples is evaluated using the FID evaluation index. The qualified generated samples are added to the unbalanced dataset of the original samples to improve the classification performance of the subsequent fault diagnosis model.

[0057] Vibration signal construction time-frequency image:

[0058] 1. Vibration signal segmentation

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

[0060]

[0061] Where N represents the number of data points in each vibration signal sample, that is, the length of the sample. n is the number of revolutions per mechanical equipment vibration signal sample. s is the sampling frequency when collecting vibration signals, indicating the number of data points recorded per second. v is the rotation speed of the mechanical equipment, indicating the number of rotations of the mechanical equipment per minute.

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

[0063] The one-dimensional vibration signal generated when mechanical equipment fails contains a large amount of redundant information and noise interference that are irrelevant to the fault characteristics. This section uses continuous wavelet transform to perform pre-processing such as noise reduction and filtering on the original vibration signal, and converts the one-dimensional signal into a two-dimensional time-frequency diagram. In the process of signal analysis and processing, wavelet decomposition will help reduce the redundant information in the original time series vibration signal, so that the model can focus on the key fault characteristics of the generated vibration signal, thereby improving the accuracy and generalization ability of the model; on the other hand, the two-dimensional time-frequency diagram not only retains the time domain information in the original vibration signal, but also visualizes the energy distribution of the signal at different frequencies, and clearly shows the frequency domain characteristics of the vibration signal, so that the classification model can more easily learn the specific category feature representation in the time-frequency diagram to distinguish between different fault categories. In summary, the two-dimensional time-frequency diagram obtained by wavelet transform can better reflect the spectral characteristics of the vibration signal of mechanical equipment under different operating conditions, which helps to improve the sample generation quality and fault diagnosis classification accuracy.

[0064] It is crucial to select a suitable wavelet function to perform continuous wavelet transform on the original signal. This paper uses Bump wavelet to perform continuous wavelet transform on the original signal and converts the obtained wavelet coefficient matrix into a two-dimensional time-frequency feature map. The shape of the Bump wavelet function is similar to a band wave packet, which has the characteristics of finite time and finite frequency. It has certain advantages in focusing on the local characteristics and transient characteristics of the signal. It is finite and localized in the time domain, which is consistent with the characteristics of the impulse signal. Due to its localized characteristics in the frequency domain, the Bump wavelet can more effectively capture the transient and impulse components in the signal. The Bump wavelet function graph is shown in the figure below. Figure 3 As shown. The Bump wavelet function is expressed as:

[0065]

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

[0067] The horizontal axis represents time, the vertical axis represents frequency, and the color or brightness change in the time-frequency diagram represents the amplitude of the signal at different frequencies and times.

[0068] From the converted time-frequency diagram, it can be clearly observed that when a mechanical equipment fails, it has obvious periodicity on the time scale. The pulse signal caused by the mechanical equipment failure appears as a bright spot on the time-frequency diagram. At the same time, the noise signal appears blue in the time-frequency diagram because of its lower amplitude than the pulse signal. This shows that by converting it into a time-frequency diagram, compared with the original timing signal of the mechanical equipment, it can not only effectively shield the interference noise other than the mechanical equipment failure signal, but also provide a variety of feature representations with different frequencies and resolutions. These features can better reflect the dynamic characteristics of the signal and the changes in frequency components, which helps to more accurately identify different types of mechanical equipment failures.

[0069] Mechanical equipment fault sample expansion method based on ICGAN

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

[0071] The original ACGAN network architecture adds an auxiliary classifier to the discriminator based on CGAN, making the generated samples of higher quality. However, the classification of the auxiliary classifier is based on the architecture of the discriminator. The discriminator in ACGAN simultaneously undertakes the functions of classification and true and false discrimination, that is, the neural network corresponding to the classification and the neural network corresponding to the true and false discrimination have exactly the same hyperparameters and weights. For this reason, the auxiliary classifier in ACGAN is designed as an independent structure, and the classification performance is enhanced by giving the network an independent classifier, so that the quality of the samples of specific categories generated by the generator is higher.

[0072] The ICGAN network structure proposed in this paper is as follows Figure 5 As shown in Figure 1. On the one hand, the generator generates samples similar to the real sample time-frequency graph based on random noise and label vectors. On the other hand, the discriminator distinguishes the true and false between the generated sample time-frequency graph and the real sample time-frequency graph to optimize the quality of the time-frequency graph samples generated by the generator. By adopting the design structure of independent classifiers, the generated sample time-frequency graph and the real sample time-frequency graph in each category are more similar and more consistent with the distribution of the real sample time-frequency graph. Here, the generated sample evaluation index FID, loss curve and other visualization techniques are used to determine the optimal network model parameters.

[0073] The network structure of the generator in ICGAN is as follows Figure 6As shown in Figure 1, the input dimension is 110, of which 100 is the noise signal and 10 is the label information. The input noise is transformed through the fully connected layer and then passed through 5 layers of 2D deconvolution layers to obtain a 3-channel 128-size 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 in Figure 2, five 2D convolutional layers are used to convert the 3-channel 128-size feature map obtained by the generator into a 52-channel 13-size feature map, and then a fully connected layer is used to output 0 / 1 to represent the discriminant output of the generated sample. The specific network parameters of the discriminator are shown in Table 2.

[0077] Table 2

[0078]

[0079]

[0080] The network structure of the independent classifier is as follows Figure 8 As shown in Figure 3, the independent classifier classifies the feature map output by the generator and the image of the real sample, and back-propagates the classification results to further optimize the generation process of the generator and improve the generation quality of each type of generated samples. The network parameters of each layer of the independent classifier are shown in Table 3.

[0081] Table 3

[0082]

[0083] Mechanical equipment fault sample expansion steps based on ICGAN

[0084] (1) Data acquisition and processing. The 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. The one-dimensional signal is converted into a two-dimensional wavelet time-frequency graph using continuous wavelet transform, 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 sent to the ICGAN generative network for iterative training, and the optimal network model parameters are determined using visualization techniques such as the generated sample evaluation indicator FID and loss curve.

[0086] (3) Unbalanced sample expansion: Random noise and label vectors are input into the trained ICGAN generation network to generate time-frequency graphs of multiple fault types and expand the original unbalanced training sample set;

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

[0088] Here, ICGAN is further optimized, and a fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution is designed. By introducing the global attention mechanism (GAM), the generation network pays more attention to the spatial, channel and dimensional information in the time-frequency graph; the octave convolution (OctConv) is used to replace the traditional two-dimensional convolution in the network to extract the high and low frequency component information in the time-frequency graph; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency graphs with key fault features, further improving the quality of generated samples.

[0089] GAMICGAN Network Structure: In order to improve the quality of the two-dimensional time-frequency images generated by the ICGAN generation network and increase the attention of the generation network to the fault time-frequency features in the two-dimensional time-frequency graph of the vibration signal, a mechanical equipment fault diagnosis method GAMICGAN based on global attention and octave convolution is proposed. The schematic diagram of the GAMICGAN network structure is shown in the figure. Fig. 9 As shown in the figure, firstly, a global attention mechanism is embedded in the generator and discriminator of ICGAN, so that the generative network model can fully pay attention to the spatial, channel and dimensional information in the time-frequency diagram when generating the time-frequency diagram of fault samples; secondly, the ordinary transposed convolution and ordinary convolution of the generator and discriminator in ICGAN are replaced by 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 diagram. By introducing the global attention mechanism and octave convolution, the generation quality of the time-frequency diagram of fault samples is effectively improved, and the fault diagnosis accuracy of mechanical equipment under sample imbalance is further improved.

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

[0091] 1. Discriminator network: In the discriminator, the input image is first downsampled using traditional convolution, and the obtained feature map is passed through the global attention module GAMAttention to focus on the channel, space and dimension information in the feature map. Then, three OctConv modules are used to deeply extract the high and low frequency component features in the feature map. Finally, the traditional convolution and fully connected layers are used to distinguish the generated samples from true and false. The improved discriminator network structure is as follows: Fig.10 As shown, the GAMAttention module and OctConv module are described in detail below. The specific network structure parameters are shown in Table 4.

[0092] Table 4

[0093]

[0094] 2. GAMAttention global attention module: The global attention module is mainly composed of two parts: channel and spatial attention. For the discriminator network structure involved in this section, the specific parameters and data flows in the two attention modules are as follows: Fig.11 As shown:

[0095] Channel attention enables the model to focus on the key features in the input data by determining the weight of each channel. It focuses on the channels related to the task by learning the weights between each channel. By leveraging the complex dependencies between channels, it can effectively capture cross-channel interactions and improve the feature representation of specific faults.

[86] , so that the generative network generates high-quality sample images. The channel attention module retains channel information in the three dimensions of the image. First, the input feature map is transposed, and then the obtained feature map is passed through the MLP (multi-layer perceptron) encoding and decoding structure with a compression ratio of r. We set r to 4 to amplify the interaction between channels and spaces between dimensions. Finally, the feature map is transposed and activated by the Sigmoid function to obtain the feature map after channel attention processing.

[0096] The spatial attention mechanism considers the relationship between image pixels, so that the model can focus on important information related to the task in a targeted manner. It learns the weight of each position to represent the importance of each position in the image, thereby improving the model's ability to express key features of the image. First, the feature map processed by the channel attention mechanism is multiplied with the input feature map, and then the spatial information is fused through a 7×7 convolution layer to better capture the local structure of the image. Then, the same compression ratio r as the channel attention is used to further compress the dimension of the feature map. Next, the 7×7 convolution layer is used again to further extract spatial features. Finally, the feature map processed by the spatial attention is obtained through the Sigmoid activation function, and the feature map processed by the spatial attention is element-wise multiplied with the feature map processed by the channel feature map to obtain the final feature map output.

[0097] 3. OctConv octave convolution module: Fig.11 As shown, the input feature map is divided into high-frequency features through OctConv Figure X H and low frequency characteristics Figure X L , which represent the high and low frequency components of the input feature map respectively. OctConv uses α in and α out Controls the proportion of high and low frequency components in the input and output channels. The value of α ranges from [0,1]. In the OctConv convolution layer, except for the first and last layers, the value of α is set to α in is 0.5, α out The discriminator network structure uses a three-layer OctConv module to update high- and low-frequency information.

[0098] Y H→H Represents the information update output feature between high-frequency components, Y L→L Represents the information between low-frequency components to update the output features. H→L and Y L→H Represents the information interaction output characteristics between different frequency components. The implementation method is expressed as:

[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] Among them, pool(X,2) represents a pooling operation with a step size of 2, and upsample(X,2) represents an upsampling operation through nearest neighbor interpolation with a step size of 2.

[0104] The first layer of OctConv, α in Set to 0, α out Set to 0.5, that is, decompose the input feature map into two feature map components Y with high frequency and low frequency H and Y L , specifically expressed as:

[0105] Y H =Y H→H (3-7)

[0106] Y L =Y H→L(3-8)

[0107] The second layer of OctConv, α in and α out are all set to 0.5, and the high-frequency feature map Y H and low-frequency feature map Y L The new high and low frequency feature maps obtained by fusion are:

[0108] Y H =Y H→H +Y L→H (3-9)

[0109] Y L =Y H→L +Y L→L (3-10)

[0110] The third layer of OctConv, α out Set to 0, α in Set to 0.5, that is, the high-frequency feature map component Y obtained in the second layer H As the final feature map output, the output Y H for:

[0111] Y H =Y H→H +Y L→H (3-11)

[0112] 4. Generator network structure: In the generator, the one-dimensional vector obtained by the fully connected layer of random noise with label information is first subjected to traditional deconvolution to obtain a feature map, and then it is enlarged by three DeOctConv channels and sizes, and the feature map size of the same size is obtained by global attention. Finally, the final generator output is obtained by traditional convolution. The improved generator network model structure is as follows: Fig.12 The specific network structure parameters are shown in Table 5. Different from the discriminator OctConv module, DeOctConv is designed in the generator to replace the ordinary transposed convolution, where g(·) represents the ordinary transposed convolution.

[0113] Table 5

[0114]

[0115] Fault diagnosis of mechanical equipment with variable working conditions based on joint distribution domain adversarial graph convolutional neural network:

[0116] In actual industrial scenarios, rotating machinery and equipment usually operate under a variety of working conditions. Mechanical equipment with the same fault point under different working conditions will generate fault signals with different data distributions, resulting in problems such as weak generalization ability of the diagnostic model and decreased accuracy. In order to realize the fault diagnosis of electromechanical equipment under variable working conditions, a variable working condition mechanical equipment fault diagnosis method JDAGCN based on the joint distribution domain adaptive adversarial graph convolutional network is proposed on the basis of generating samples by GAMICGAN. Firstly, a lightweight feature extraction module GAMLRN based on the global attention residual network is designed to extract deep features in the time-frequency graph. Secondly, the adversarial transfer mechanism of the domain adversarial neural network is used to align the fault features under different working conditions. The marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified by introducing the metric of joint maximum mean difference. Finally, the graph convolutional network GCN is used to construct the structural relationship between samples in the same domain and between samples in different domains to further reduce the distribution difference between the source domain and the target domain. Experimental analysis was carried out on the CWRU bearing dataset. The experimental results show that the JDAGCN model has a higher 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 graph and prevent the network from being too complex and causing a decrease in the generalization ability of the model, this section designs a lightweight feature extractor based on the global attention residual network (GlobalAttention Mechanism Lightweight Residual neuralnetwork, GAMLRN).

[0118] In the GAMLRN network, a convolution kernel of size 3×3 is used, the convolution step 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 cancelled, and the deep features in the time-frequency graph are extracted by embedding global attention. The GAMLRN network structure is as follows: Fig.13 The specific network parameters are shown in Table 6.

[0119] Table 6

[0120]

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

[0122] JDAGCN consists of data preprocessing, feature extractor GAMLRN, graph embedding theory GCN, label classifier G y and domain discriminator G dThe data preprocessing module processes data from both the source domain and the target domain simultaneously. The feature extractor GAMLRN uses a lightweight global attention residual neural network model to extract deep features from both the source domain and the target domain simultaneously. The label classifier consists of a fully connected layer to classify the data in the source domain. The domain discriminator uses a fully connected layer to distinguish whether the data comes from the source domain or the target domain. In the domain adaptation stage, the JMMD combined with the GCN metric method is used to reduce the marginal distribution difference and conditional distribution difference between the source domain and the target domain.

[0123] The optimization objective of the JDAGCN model is the label classification loss L y , domain discrimination loss L d and JMMDGC distribution difference loss L JMMDGC The final target loss function is L JDANN :

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

[0125] Among them, λ1 and λ2 are both hyperparameters. λ1 is used to balance the label classification loss and domain discrimination loss. At the same time, the weight of the domain discrimination loss is negatively adjusted by using the inverted gradient method, so that the model can increase the domain discrimination loss as much as possible while reducing the label classification loss. Through this domain adversarial training method, the model can learn deeper feature representations and improve the generalization ability of the model in the target domain. λ2 is used to balance the importance between the label classification loss and the JMMD loss, so that the model pays attention to the difference in the joint feature distribution between the source domain and the target domain.

[0126] Variable operating condition fault diagnosis process: JDAGCN variable operating condition diagnosis process is as follows Fig.15 The JDAGCN variable operating condition diagnosis process is divided into the following four steps:

[0127] Step 1: Data preprocessing. The original vibration signals of the source domain and the target domain are obtained, and the vibration signals are processed by continuous wavelet transform to obtain the corresponding time-frequency graph samples. Then, the two-dimensional time-frequency images of different fault types are input into the fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution. By mixing the generated specific type of fault category sample data with the original sample set, the purpose of expanding the unbalanced data set is achieved to improve the fault diagnosis accuracy under unbalanced samples.

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

[0129] Step 3: Domain adaptation. Calculate the overall loss function and optimize the label classifier loss, domain discriminator loss, and joint maximum mean difference loss through the back-propagation process. And before the maximum number of rounds, cyclically extract the deep features of the source domain and the target domain through domain confrontation in order to obtain the same feature representation between the two domains.

[0130] Step 4: Diagnosis and classification. After obtaining the trained JDAGCN model, the fault sample time-frequency diagram of the target domain is input into the JDAGCN model to obtain the final target domain classification effect.

[0131] Experimental analysis: Experimental analysis based on ICGAN generation network:

[0132] In order to verify the data augmentation effect of the proposed ICGAN network in mechanical equipment fault diagnosis, the two-dimensional time-frequency graph samples are input into the ICGAN network to obtain the generated fault samples. The mixed samples change the unbalanced state of the original data set into a balanced sample state. By comparing the classification effect of the same CNN classifier before and after data augmentation on the CWRU bearing data set, the effectiveness and feasibility of the proposed network are verified. At the same time, by designing a comparative experiment with the augmentation performance of other generative networks, the high availability of the proposed network is verified. The experimental results show that the loss curve of the designed ICGAN generator can converge quickly, and the generated time-frequency graph can clearly express the characteristic information of the bearing under different health conditions; at the same time, when the imbalance ratio is 1:40, after ICGAN augments a small number of training samples, the model diagnosis accuracy is increased by about 15%, which is about 7% higher than other generative network models.

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

[0134] The two-dimensional time-frequency graph obtained on 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 in Figure 2. Fig.16 As shown in (a), the discriminator iterative output is Fig.16 (b) as shown.

[0135] In order to compare the improvement effect of the original ACGAN generation network, the hyperparameters of the ACGAN generation network are set to the same batch, number of iterations, learning rate and optimizer as those of the ICGAN generation network. The iterative loss curves of the generator and discriminator in the original ACGAN are shown in Figure 2. Fig.17 As shown in (a), the discriminator iterative output is Fig.17 (b) as shown.

[0136] By comparing the loss curves of ICGAN and the original ACGAN network, it can be seen that the loss of the generator of the original ACGAN network began to rise after 500 epochs, and the generator loss continued to rise in the subsequent iterations, indicating that the ACGAN network model did not converge and a mode collapse phenomenon occurred, that is, the generator generation quality of the ACGAN network was very poor, which made it easy for the discriminator to judge that the generated samples were "fake" samples. Under the action of the independent classifier, ICGAN constrained the generation effect of the generator, so that the loss curve continued to decline after 500 epochs. In the subsequent iterations, the losses of the generator and the discriminator tended to be stable. At the same time, 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 value for generated samples and real samples converges to the direction of 0.5, that is, the Nash equilibrium state.

[0137] Different types of fault generation samples obtained by ICGAN generation network, such as Fig.18 shown.

[0138] It can be seen that the generated two-dimensional time-frequency images of fault samples of the same type are similar, and different types of generated samples are distinguishable to a certain extent, that is, the data distribution of the generated fault samples is roughly consistent with the data distribution of the real samples, indicating that the ICGAN generation network has high stability in the generative adversarial training process. By designing the structure of the independent classifier, the generator can further optimize the generation process of the generator according to the classification results obtained by the classifier during the generation of fault samples, so that the quality of the generated sample two-dimensional time-frequency image is better.

[0139] In addition, in order to measure the quality and diversity of the generator output in the generative network, IS (Inception Score) and FID (Frechet Inception Distance Score) are often used as evaluation indicators. The IS indicator was proposed by Salimans et al. in 2016. This indicator uses the output image of the generator as the input of the Inception V3 model to obtain the probability that the image belongs to a certain category. The larger the probability value obtained, the clearer the image obtained, and the smaller the entropy of p(y|x), which is used to judge the generation quality of the generator. When the probability of the generated samples belonging to each category is the same, that is, the closer to the uniform distribution, the larger the entropy of p(y), and the more diverse the generated samples are. Using KL divergence to calculate the distance between two probability distributions, the expression of the IS indicator is:

[0140]

[0141] However, since the IS indicator is very sensitive to the weights of the network model, different model weights will cause large changes in the IS indicator. Secondly, the IS indicator cannot detect the mode collapse problem in the generative network, that is, although the generated images generate a certain number of images in each class, they may lack diversity, which means that it only considers the quality of the generated samples and does not consider the impact of the real data distribution.

[0142] To this end, Heusel et al. proposed the FID evaluation index in 2017. Compared with the IS index, the IS directly outputs the category when using the Inception V3 model, while the FID outputs the corresponding feature vector. FID uses the feature vector in the Inception model to compare the distribution difference between the real image and the generated image. The lower the FID, the closer the distribution of the generated image is to the real image, and the higher the quality of the generated sample. The calculation expression of FID is:

[0143]

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

[0145] This paper uses the FID indicator to evaluate the quality and diversity of the two-dimensional fault sample frequency graphs generated by the generator. Table 7 records the FID values ​​of the images generated by the ICGAN network during the iteration process. The changes of FID during the iteration process are shown in Table 7. Fig.19 shown.

[0146] Table 7

[0147]

[0148] Table 8

[0149]

[0150] Analysis of the above results shows that as the number of network iterations increases, the FID of the generated samples of each category gradually decreases and eventually stabilizes, indicating that the generator has learned the approximate data distribution of the real sample time-frequency graph and thus generated a generated sample time-frequency graph similar to the real sample time-frequency graph. The FID value decreases rapidly between the 300th and 600th rounds of iteration, which is consistent with the change law of the loss function curve of ICGAN, indicating that the network begins to converge gradually after 600 rounds. There are differences in the FID values ​​of different types of generated sample time-frequency graphs. For example, the FID of the generated sample time-frequency graphs of the inner circle fault 0.014 size and the outer circle fault 0.007 size is smaller than that of the generated samples of other types of faults, while the FID values ​​of the generated samples of the time-frequency graphs of the inner circle fault 0.021 size and the normal state are relatively large, indicating that the ICGAN network has diversity in generating two-dimensional time-frequency images of different fault types, making the generated two-dimensional time-frequency images of different fault types have certain distinguishability.

[0151] Fault diagnosis under unbalanced samples: To verify that the proposed ICGAN network can expand the data set in an unbalanced state into a data set in a balanced state, thereby improving the accuracy of bearing fault diagnosis under unbalanced samples. Set the unbalanced state data set of different proportions of CWRU, as shown in Table 9. Among them, the data set is divided into a training set and a test set, and the training set is divided into different unbalanced ratios. 1:N represents the quantitative ratio between the time-frequency diagram of a certain fault type state of the bearing and the time-frequency diagram of the normal state. The test set is tested with a ratio of 1:1.

[0152] Table 9

[0153]

[0154] In order to verify the data expansion effect of the ICGAN network, the following experiments are designed for verification:

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

[0156] (2) Experiment B: The unbalanced dataset of real samples is augmented through the GAN network to reach a balanced state and then classified.

[0157] (3) Experiment C: The unbalanced dataset of real samples is augmented through the ACGAN network to reach a balanced state and then classified.

[0158] (4) Experiment D: The unbalanced dataset of real samples is augmented through the ICGAN network to achieve a balanced state and then classified.

[0159] Different generative networks are used to transform the unbalanced data set into a balanced data set by data augmentation. The balanced data set is trained with the same CNN classifier. The model parameters of the CNN used are exactly the same as those of the independent classifier. The training is performed for 100 rounds with a batch size of 40. The Adam optimizer is used with a learning rate of 0.002. The test accuracy is shown in Table 10. The classification accuracy of different generative networks after augmentation at different imbalance ratios is shown in Table 10. Fig. 20 shown.

[0160] Table 10

[0161]

[0162] By comparing with other generative networks, it can be seen that the balanced state data set after expansion by the ICGAN network improves the final classification accuracy. The ICGAN generative network has a certain accuracy improvement under each imbalance ratio compared with the original ACGAN network through an independent classifier design structure, and the greater the imbalance ratio, the greater the improved accuracy value, indicating that the ICGAN network has learned a more accurate distribution of real sample time-frequency graph data than the original ACGAN network, and generated a two-dimensional time-frequency image with higher quality diversity and authenticity.

[0163] Experimental analysis of GAMICGAN network effect: Sample generation evaluation: The GAMICGAN network uses the CWRU unbalanced sample dataset constructed in Chapter 3 for training and testing. The number of training rounds, batch size, learning rate, and optimizer are the same as those of the ICGAN network. The loss curves of the GAMICGAN network discriminator and generator are shown in the figure below. Fig.21 As shown in (a), the output value of the discriminator is Fig.21 (b) as shown.

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

[0165] The time-frequency diagrams of various types of fault samples generated by the GAMICGAN network are shown in the figure below: Fig. 22 shown.

[0166] It can be seen from the figure that the generated fault sample time-frequency graph has less noise and better quality. The FID values ​​of the time-frequency samples generated by GAMICGAN under some iterations are shown in Table 11.

[0167] Table 11

[0168]

[0169] As can be seen from Table 11, when training to 300 rounds, the FID value of the GAMICGAN network has reached between 200 and 500. The FID value between the 300th and 600th rounds drops rapidly. When training to around 1000th round, the FID value has stabilized. Under the same number of iterations, the FID value of the GAMICGAN network is lower than that of the ICGAN network, indicating that the generated samples are more diverse and more similar to the real samples.

[0170] In order to verify the effectiveness of the GAMICGAN network in generating time-frequency diagrams of fault samples, in addition to the numerical analysis method of FID, this section also uses the heat map method to analyze the quality of the newly generated fault samples.

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

[88] It is a method for explaining the decision-making process of CNN models. It generates a class activation heat map to analyze the model's focus area for a certain category. Analyzing the heat map helps to understand the feature learning process within the network and evaluate 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, and then generates a class activation map by characterizing the importance of the last convolutional layer activation output for the model's decision classification. This process can be seen as a weighted operation that allows the model to represent the degree of attention paid to different features. By rendering the preprocessed class activation map as a thermal effect, the focus area is made more prominent. Finally, the original image is superimposed with the rendered thermal effect to obtain the final heat map. The brighter area in the heat map indicates that the model pays more attention to this area. The specific expression is:

[0172]

[0173] Among them, f R represents the ReLU activation function, Indicates A k The weight of , d represents the category, A represents the feature layer, k represents the kth channel in the feature layer A, A k 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 represents the score predicted by the network for category d, Represents the pixel value of feature layer A at the coordinate (i, j) in channel k.

[0177] The generated fault sample heat map is compared with Fig.23 By comparing the heat map of the fault sample with a ball fault damage size of 0.021 inch under 0HP generated by the ICGAN network and the heat map 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 graph during the generation process, and these high-frequency components are the characteristic representations of different bearing fault types in the time-frequency graph. By adding the global attention mechanism and OctConv to the ICGAN network, the network generates fault samples with key time-frequency features.

[0178] Fault diagnosis under unbalanced samples: In order to verify the fault classification accuracy after data expansion by GAMICGAN network under various unbalanced ratios, the classification accuracy after data expansion by ICGAN network is compared, and the fault classification accuracy results of various unbalanced ratios are obtained as shown in Table 12 and Fig.24 shown.

[0179] Table 12

[0180]

[0181] From the above results, it can be seen that as the imbalance ratio increases, the expansion effect of GAMICGAN is relatively obvious compared with the expansion effect of the ICGAN network, indicating that the GAMICGAN generation network further improves the attention to the fault time-frequency features in the vibration signal sample time-frequency diagram when generating the fault sample time-frequency diagram. By introducing the global attention mechanism on the basis of the ICGAN network, the generation network can fully pay attention to the space, channel and dimensional information in the time-frequency diagram. By replacing the traditional convolution with the octave convolution, the generation network can extract the high-frequency and low-frequency component information of the time-frequency diagram in the process of generating samples, thereby generating a higher quality two-dimensional fault sample time-frequency diagram.

[0182] Variable operating condition bearing fault diagnosis based on joint distribution domain adversarial neural network: In order to verify the performance of the variable operating condition bearing fault diagnosis method based on joint distribution domain adversarial neural network proposed in this section, this section conducts variable operating condition diagnosis experiments on the CWRU bearing dataset, designs ablation experiments to analyze the role of each component in the variable operating condition bearing fault diagnosis model, and verifies the superiority and effectiveness of the proposed method by comparing it with other variable operating condition diagnosis models.

[0183] Ablation experiment of fault diagnosis under variable working conditions: Based on the DANN model, a lightweight global attention residual neural network GAMLRN is used as the feature extraction network, and the graph convolution network GCN is input at the same time, and the JMMD metric method is used to reduce the distribution difference between the two domains. In order to verify the effectiveness of each module in improving the accuracy of fault diagnosis, 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: The feature extraction network in Model A is replaced by the feature extraction network GAMLRN in this paper.

[0186] (3) Model C: The joint maximum mean difference (JMMD) metric is used for migration, and the feature extraction network adopts GAMLRN.

[0187] (4) Model D: The variable operating 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] At the same time, 12 groups of domain migration tasks are designed, as shown in Table 13, where 0HP->1HP represents the migration diagnosis from the source domain to the target domain. Fig.25 The diagnostic process shown in the figure is carried out, the number of training and testing rounds is set to 300, the batch size is 50, the optimizer is Adam, the learning rate is 0.0002, and the hyperparameters λ1 and λ2 are both set to 1.

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

[0190] Table 14

[0191]

[0192] The iterative accuracy curves of each model in the migration diagnosis task under different working conditions are shown in the figure below: Fig. 27 shown.

[0193] From the above results we can see that:

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

[0195] (2) Compared with Model A, when Model B also uses the DANN model for two-domain migration, after replacing the CNN feature extraction network with GAMLRN, the model convergence stability and migration diagnosis accuracy are significantly improved, indicating that the deeper model improves the generalization ability.

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

[0197] (4) Model D combines the advantages of Model B and Model C. It uses the JMMD measurement method based on the DANN model and considers the differences in marginal distribution and conditional distribution between the two domains at the same time, thereby improving the performance of the feature extractor and making the model more stable, faster in convergence and more accurate in migration.

[0198] The t-SNE distribution results of each model under the migration from working condition 0 to working condition 3 are as follows: Fig.26 shown.

[0199] from Fig.26 It can be seen that when performing migration diagnosis tasks based on the DANN model, the feature representations shared by the source domain and the target domain are mapped to the same high-dimensional space through adversarial training, and there is a large category distance between different classes. This paper adds the JMMD measurement method on the basis of the DANN model, so that the model pays attention to the intra-class differences between the same class in the feature extraction stage, and achieves good intra-class aggregation while maintaining the difference distance between different categories.

[0200] 4.3.2 Comparative experiment on fault diagnosis under variable working conditions

[0201] In order to verify the superiority of the JDAGCN model proposed in this chapter in the variable working condition bearing fault diagnosis task, different diagnosis models are used to conduct comparative experiments on working condition migration tasks. By verifying on the CWRU bearing dataset, the migration diagnosis accuracy results under different working conditions are obtained. The fault diagnosis accuracy of each model is shown in Table 153. The details of each comparative experiment use the same settings as the ablation experiment. The selected comparative experiment models are as follows:

[0202] (1) CNN. Use the CNN classification model in Chapter 3 for variable operating condition diagnosis. Input the fault samples in the source domain into the CNN model for training, and use the trained CNN model to directly classify the fault samples in the target domain.

[0203] (2) MMD. The MMD measurement method is used to perform migration diagnosis tasks under different working conditions, where CNN is used as the feature extraction network.

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

[0205] The accuracy of the variable operating condition fault diagnosis results of the CWRU bearing data set comparison experiment is shown in Table 15.

[0206] Table 15

[0207]

[0208] By analyzing the fault classification accuracy of each diagnostic model under different working conditions, the following conclusions can be drawn:

[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 decreases. From Table 15, it can be seen that 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 the bearing fault. The model does not take into account the data distribution differences between different fields under different working conditions, resulting in overfitting of the model and poor generalization ability in the target domain.

[0210] (2) When the variable operating condition migration models such as MMD and DAN are used to classify bearing faults under different operating conditions, the fault classification accuracy is significantly improved. Compared with the CNN non-migration model, when performing fault migration diagnosis with large differences in operating conditions such as operating conditions 0 to 3, the fault classification accuracy of the target domain reaches more than 95%, indicating the effectiveness of the migration mechanism in variable operating condition fault diagnosis.

[0211] (3) The JDAGCN model further improves the accuracy of variable operating condition diagnosis by adding the JMMD joint distribution difference measurement method on the basis of the DANN migration model. The fault classification accuracy reaches 100% in most variable operating condition migration diagnosis tasks. This shows that the JDAGCN model has excellent robustness and generalization ability in variable operating condition bearing fault diagnosis tasks, can better capture the data distribution differences between different fields, and achieve better feature representation through adversarial training.

[0212] In the migration task from working condition 0 to working condition 3, by performing deep feature extraction on different comparison models and using the t-SNE algorithm to map high-dimensional features to two-dimensional space, the t-SNE distribution obtained is as follows: Fig. 27 shown.

[0213] from Fig. 27 It can be observed that compared with model A, the deep features of model B and model C with domain migration mechanism have larger inter-class distance and smaller intra-class distance in two-dimensional space, which means that model B and model C can better distinguish the features between different bearing states, thereby greatly improving the confusion problem between categories. In the JDAGCN model, it is found that the separability of the extracted deep features has been further improved, and the confusion between classes has been reduced, indicating that it has better ability to solve multi-domain fault diagnosis problems. In summary, compared with the other three groups of control models, the deep features extracted by the JDAGCN model have better cross-domain invariance, and can more accurately distinguish the different operating states of bearings when solving complex multi-domain fault diagnosis problems.

[0214] Fig.28 is the confusion matrix result of each comparison model under working condition 0 to working condition 3, from Fig.28 It can be seen that when the CNN model is used directly for transfer diagnosis, the accuracy of transfer diagnosis in the target domain is very poor. The feature transfer diagnosis method of MMD and DAN has significantly improved the performance of transfer diagnosis tasks under different working conditions, but there are still some misjudgments for certain types of bearing faults. The JDAGCN variable working condition bearing fault diagnosis method proposed in this paper has well realized the transfer diagnosis task under different working conditions.

[0215] (3) Fault diagnosis experiment under variable working conditions under unbalanced samples

[0216] In order to verify the variable working condition diagnosis performance under sample imbalance, this section first divides the CWRU bearing dataset into sample sets with different balance ratios, then uses the GAMICGAN generation network proposed in 3.2 to expand the datasets of each imbalance ratio and uses the JDAGCN model in this section to perform cross-working condition diagnosis experiments. The classification accuracy of the bearing variable working condition diagnosis is shown in Table 16.

[0217] Table 16

[0218]

[0219] It can be seen from Table 16 that when the sample imbalance ratio increases, the accuracy of bearing fault classification under variable working conditions decreases faster, which shows the importance of expanding the imbalance data set into a balanced state data set during variable working condition diagnosis. It can also be seen that under the same imbalance ratio, the fault classification accuracy under variable working conditions is not as good as that under the same working conditions, which shows the importance of data distribution differences during variable working condition migration.

Claims

1. A fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network, characterized in that: The specific steps include: Step 1: Aiming at the non-stationary and nonlinear characteristics of vibration signals, a vibration signal time-frequency analysis method based on wavelet analysis is used to pre-process 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 generation network and the diagnostic model, and an independent auxiliary classifier structure is designed for the ACGAN network. The weight adjustment mechanism of the discriminator and the auxiliary classifier is introduced to obtain an improved generation network model ICGAN, thereby improving the generation quality of the generation network for samples of the specified category; Step 2: Use global attention and octave convolution to further optimize the generative network structure of ICGAN; Based on global attention and octave convolution, the fault sample time-frequency graph generation network GAMICGAN introduces the global attention mechanism GAM to make the generation network pay more attention to the spatial, channel and dimensional information in the time-frequency graph; Use octave convolution OctConv to replace the traditional two-dimensional convolution in the network to extract the high and low frequency component information in the time-frequency graph; By combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency graphs with key fault features, further improving the quality of generated samples; Step 3: To address the problem of decreased accuracy of the diagnostic model caused by differences in sensor data distribution under variable working conditions, the deep features of the time-frequency graph of vibration signal samples are extracted by combining the graph convolutional network (GCN) and the backbone network structure GAMLRN based on the global attention residual network. The adversarial transfer mechanism of the domain adversarial neural network is used to align the fault features under different working conditions. At the same time, the marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified by introducing the joint maximum mean difference metric to optimize the backbone network parameters and further reduce the distribution difference between the deep features of the sample time-frequency graphs under different working conditions. At the same time, 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 domain and the target domain, and further reduce the distribution difference between the deep features of the sample time-frequency graphs under different working conditions.

2. The fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network according to claim 1, characterized in that: In step 1, the sample imbalance problem is solved from the perspective of data generation. Based on the ACGAN generation network, a fault sample time-frequency graph expansion method ICGAN based on the improved ACGAN network 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; an improved 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 index. The generated samples that pass the evaluation are added to the unbalanced dataset of the original samples to improve the classification performance of the subsequent fault diagnosis model. The specific steps include: Step 1.1, construct time-frequency image of vibration signal: Vibration signal segmentation: The one-dimensional vibration signal is segmented into multiple sample sets of the same sequence length using the sliding window method. In the sliding window value selection process, in order to ensure that the segmented vibration signal has the signal characteristics of at least one rotation period, the number of data points of each sequence sample must meet the following conditions: Where N represents the number of data points in each vibration signal sample, that is, the length of the sample; r n is the number of revolutions of each mechanical equipment vibration signal sample; f s is the sampling frequency when collecting vibration signals, indicating the number of data points recorded per second; v is the rotation speed of the mechanical equipment, indicating the number of rotations of the mechanical equipment per minute; Construction of two-dimensional time-frequency diagram: The one-dimensional vibration signal generated when mechanical equipment fails contains a large amount of redundant information and noise interference that are irrelevant to the fault characteristics. The original vibration signal is denoised and filtered using continuous wavelet transform, and the one-dimensional signal is converted into a two-dimensional time-frequency diagram. In the process of signal analysis and processing, wavelet decomposition will help reduce the redundant information in the original time series vibration signal, so that the model can focus on the key fault characteristics of the generated vibration signal, thereby improving the accuracy and generalization ability of the model. On the other hand, the two-dimensional time-frequency diagram not only retains the time domain information in the original vibration signal, but also visualizes the energy distribution of the signal at different frequencies, and clearly shows the frequency domain characteristics of the vibration signal, so that the classification model can more easily learn the specific category feature representation in the time-frequency diagram to distinguish between different fault categories. The two-dimensional time-frequency diagram obtained by wavelet transform can better reflect the spectral characteristics of the vibration signal of mechanical equipment under different operating conditions. Bump wavelet is used to perform continuous wavelet transform on the original signal, and the obtained wavelet coefficient matrix is ​​converted into a two-dimensional time-frequency feature map; the shape of the Bump wavelet function is similar to a strip wave packet, which has the characteristics of finite time and finite frequency. It has certain advantages in focusing on the local characteristics and transient characteristics of the signal. It is finite and localized in the time domain, which is consistent with the characteristics of the impulse signal. Due to its localized characteristics in the frequency domain, the Bump wavelet can more effectively capture the transient and impulse components in the signal; the Bump wavelet function is expressed as: The two-dimensional time-frequency image is obtained by converting the one-dimensional vibration signal using the Bump wavelet function; Step 1.2, mechanical equipment fault sample expansion method based on ICGAN: By improving the ACGAN network, a generative adversarial network ICGAN based on an independent classifier is proposed; by designing an independent auxiliary classifier structure for the ACGAN network and assigning different weights to the network parameters of the discriminator and the auxiliary classifier, the quality of the generated fault samples is improved; the auxiliary classifier in the ACGAN is designed as an independent structure, and the classification performance is enhanced by assigning an independent classifier to the network, so that the quality of the specific category samples generated by the generator is higher; The generator generates samples similar to the real sample time-frequency graphs based on random noise and label vectors; the discriminator distinguishes the generated sample time-frequency graphs from the real sample time-frequency graphs to optimize the quality of the time-frequency graph samples generated by the generator. By adopting the design structure of independent classifiers, the generated sample time-frequency graphs and the real sample time-frequency graphs in each category are more similar and more consistent with the distribution of the real sample time-frequency graphs; the optimal network model parameters are determined using visualization techniques such as the generated sample evaluation index FID and the loss curve; The network structure of the generator in ICGAN: The input dimension is 110 dimensions, of which 100 dimensions are noise signals and 10 dimensions are label information. The input noise is transformed into features 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 network structure of the discriminator in ICGAN: 5 2D convolutional layers are used to convert the 3-channel 128-size feature map obtained by the generator into a 52-channel 13-size feature map, and then a fully connected layer outputs 0 / 1 to represent the discriminant output of the generated sample; The network structure of the independent classifier: The independent classifier classifies the feature maps output by the generator and the images of the real samples, and back-propagates the classification results to further optimize the generation process of the generator and improve the generation quality of each type of generated samples; Step 1.3, mechanical equipment failure sample expansion steps based on ICGAN: Data acquisition and processing: Collect vibration signals of equipment under various working conditions and different fault types, segment the original vibration signals using sliding windows, and use continuous wavelet transform to convert the one-dimensional signals into two-dimensional wavelet time-frequency diagrams, thereby constructing unbalanced training sample sets and test sample sets; Generative network training and sample evaluation: A small amount of training samples are sent to the ICGAN generative network for iterative training, and the optimal network model parameters are determined using visualization techniques such as the generated sample evaluation indicator FID and loss curve; Unbalanced sample expansion: Random noise and label vectors are input into the trained ICGAN generation network to generate time-frequency graphs of multiple fault types and expand the original unbalanced training sample set.

3. The fault diagnosis method based on an auxiliary generation network and a 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 summary: ICGAN is further optimized, and the fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution is introduced. By introducing the global attention mechanism GAM, the generation network pays more attention to the spatial, channel and dimensional information in the time-frequency graph; the octave convolution OctConv is used to replace the traditional two-dimensional convolution in the network to extract the high and low frequency component information in the time-frequency graph; by combining the GAM module and OctConv, GAMICGAN can generate sample time-frequency graphs with key fault features, further improving the quality of generated samples; Step 2.2, GAMICGAN network structure: In order to improve the quality of the two-dimensional time-frequency images generated by the ICGAN generative network and increase the attention of the generative network to the fault time-frequency features in the two-dimensional time-frequency graph of the vibration signal, a mechanical equipment fault diagnosis method GAMICGAN based on global attention and octave convolution is proposed; A global attention mechanism is embedded in the generator and discriminator of ICGAN, so that the generative network model can fully pay attention to the spatial, channel and dimensional information in the time-frequency graph when generating the time-frequency graph of fault samples; The ordinary transposed convolution and ordinary convolution of the generator and discriminator in 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 components of the time-frequency graph. By introducing the global attention mechanism and octave convolution, the generation quality of the fault sample time-frequency graph is effectively improved, and the fault diagnosis accuracy of mechanical equipment under sample imbalance is further improved; Discriminator network: In the discriminator, the input image is downsampled using traditional convolution, and the obtained feature map is passed through the global attention module GAMAttention to focus on the channel, space and dimension information in the feature map. Three OctConv modules are used to deeply extract the high and low frequency component features in the feature map, and the generated samples are distinguished by traditional convolution and fully connected layers. GAMAttention global attention module: The global attention module is mainly composed of two parts: channel and spatial attention. Channel attention determines the weight of each channel so that the model can focus on the key features in the input data. It learns the weights between each channel to focus on the channels related to the task. By utilizing the complex dependencies between channels, it can effectively capture cross-channel interactions and improve the feature representation of specific faults, so that the generative network can generate high-quality sample images. The channel attention module retains channel information in three dimensions of the image. The input feature map is transposed and the obtained feature map is passed through the encoding and decoding structure of the multi-layer perceptron MLP with a compression ratio of r. r is set to 4 to amplify the interaction between channels and space between dimensions. The feature map is transposed and activated by the Sigmoid function to obtain the feature map after channel attention processing. The spatial attention mechanism considers the relationship between image pixels, so that the model can focus on important task-related information in a targeted manner. It learns the weight of each position to represent the importance of each position in the image, thereby improving the model's ability to express key image features. The feature map processed by the channel attention mechanism is multiplied with the input feature map, and then the spatial information is fused through a 7×7 convolution layer to better capture the local structure of the image. The same compression ratio r as the channel attention is used to further compress the dimension of the feature map. The 7×7 convolution layer is used again to further extract spatial features. The feature map processed by the spatial attention is obtained through the Sigmoid activation function, and the feature map processed by the spatial attention is element-wise multiplied with the feature map processed by the channel feature map to obtain the final feature map output. OctConv octave convolution module: OctConv divides the input feature map into high-frequency feature map X H And the low frequency feature map X L , which represent the high and low frequency components of the input feature map respectively; OctConv uses α in and α out Control the proportion of input and output channels of high and low frequency components. The value range of α is [0,1]. In the OctConv convolution layer, except for the first and last layers, the value of α is set to α in is 0.5, α out is 0.5 to facilitate the extraction of high- and low-frequency feature information; the discriminator network structure uses a three-layer OctConv module to achieve high- and low-frequency information updates; Y H→H Represents the information update output feature between high-frequency components, Y L→L Represents the information update output feature between low-frequency components; Y H→L and Y L→H Represents the information interaction output characteristics between different frequency components; the implementation method is expressed as: Y H→H =Conv(X H ) (3-3) Y L→L =Conv(X L ) (3-4) Y H→L =Conv(pool(X H ,2)) (3-5) Y L→H =upsample(Conv(X L ),2) (3-6) Among them, pool(X,2) represents a pooling operation with a step size of 2, and upsample(X,2) represents an upsampling operation through nearest neighbor interpolation with a step size of 2; The first layer of OctConv, α in Set to 0, α out Set to 0.5, that is, decompose the input feature map into two feature map components Y with high frequency and low frequency H and Y L , specifically expressed as: AND H =And H→H (3-7) AND L =And H→L (3-8) The second layer of OctConv, α in and α out are all set to 0.5, and the high-frequency feature map Y H and low-frequency feature map Y L The new high and low frequency feature maps obtained by fusion are: AND H =And H→H +Y L→H (3-9) AND L =And H→L +Y L→L (3-10) The third layer of OctConv, α out Set to 0, α in Set to 0.5, that is, the high-frequency feature map component Y obtained in the second layer H As the final feature map output, the output Y H for: AND H =And H→H +Y L→H (3-11) Generator network structure: In the generator, the one-dimensional vector obtained by the fully connected layer of random noise with label information is first subjected to traditional deconvolution to obtain a feature map, which is then amplified by three DeOctConvs for channel and size enlargement. The feature map size of the same size is obtained by global attention, and the final generator output is obtained by traditional convolution. Unlike the discriminator OctConv module, DeOctConv is designed in the generator to replace the ordinary transposed convolution, where g(·) represents the ordinary transposed convolution.

4. The fault diagnosis method based on an auxiliary generation network and a domain adversarial graph network according to claim 1, characterized in that: In step 3, the fault diagnosis of mechanical equipment under variable working conditions based on the joint distribution domain adversarial graph convolutional neural network is as follows: A lightweight feature extraction module GAMLRN based on the global attention residual network is used to extract deep features in the time-frequency graph. The adversarial transfer mechanism of the domain adversarial neural network is used to align fault features under different working conditions. The marginal distribution difference and conditional distribution difference between the source domain and the target domain are quantified by introducing the joint maximum mean difference metric. The graph convolutional network GCN is used to construct the structural relationship between samples in the same domain and between samples in different domains to further reduce the distribution difference between the source domain and the target domain. Step 3.1, GAMLRN feature extraction network design: A lightweight feature extractor GAMLRN based on global attention residual network; in the GAMLRN network, a convolution kernel of size 3×3 is used, the convolution step is set to 1, the padding value is set to 1 or 0, only three residual structures are used, the pooling operation of the middle layer is cancelled, and the deep features in the time-frequency graph are extracted by embedding global attention; Step 3.2, variable operating condition fault diagnosis network framework: The variable operating condition fault diagnosis model structure of the joint distribution domain adversarial neural network JDAGCN consists of data preprocessing, feature extractor GAMLRN, graph embedding theory GCN, label classifier G y and domain discriminator G d composition; Among them, the data preprocessing module processes data from the source domain and the target domain at the same time. The feature extractor GAMLRN uses a lightweight global attention residual neural network model to extract deep features of the source domain and the target domain at the same time. The label classifier is composed of a fully connected layer to classify the data in the source domain. The domain discriminator uses a fully connected layer to distinguish whether the data comes from the source domain or the target domain. In the domain adaptation stage, the JMMD combined with the GCN measurement method is used to reduce the marginal distribution difference and conditional distribution difference between the source domain and the target domain. The optimization objective of the JDAGCN model is the label classification loss L y , domain discrimination loss L d and JMMDGC distribution difference loss L JMMDGC The final target loss function is L JDANN : L JDANN =L y -λ1L d +λ2L JMMDGC (3-12) Among them, λ1 and λ2 are both hyperparameters; λ1 is used to balance the label classification loss and the domain discrimination loss, and the weight of the domain discrimination loss is negatively adjusted by using the inverted gradient method, so that the model can increase the domain discrimination loss as much as possible while reducing the label classification loss. Through this domain adversarial training method, the model can learn deeper feature representations and improve the generalization ability of the model in the target domain; λ2 is used to balance the importance between the label classification loss and the JMMD loss, so that the model pays attention to the difference in joint feature distribution between the source domain and the target domain; Step 3.3, variable operating condition fault diagnosis process: It is divided into the following four steps: Step 3.31, data preprocessing: obtain the original vibration signals of the source domain and the target domain, and perform continuous wavelet transform on the vibration signals to obtain the corresponding time-frequency graph samples, and then input the two-dimensional time-frequency images of different fault types into the fault sample time-frequency graph generation network GAMICGAN based on global attention and octave convolution; by mixing the generated specific type of fault category sample data with the original sample set, the purpose of expanding the unbalanced data set is achieved to improve the fault diagnosis accuracy under unbalanced samples; Step 3.32, feature extraction: After setting the model parameters of JDAGCN, the obtained source domain and target domain time-frequency map samples are input into GAMLRN to extract the deep features of the source domain and the target domain at the same time; 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 the back-propagation process; and before the maximum number of rounds, cyclically extract the deep features of the source domain and the target domain through domain confrontation, in order to obtain the same feature representation between the two domains; Step 3.34, diagnostic classification: After obtaining the trained JDAGCN model, the time-frequency diagram of the fault samples in the target domain is input into the JDAGCN model to obtain the final target domain classification effect.

Images