Vacuum dry pump bearing fault diagnosis method based on domain confrontation and attention transfer learning

Abstract

The invention discloses a vacuum dry pump bearing fault diagnosis method based on domain confrontation and attention transfer learning, and aims to solve the problem of performance reduction caused by data distribution difference of dry vacuum pump bearing faults in cross-domain fault diagnosis of a deep learning method. Domain adversarial training is carried out by introducing a gradient inversion layer, and feature weights are dynamically adjusted in combination with an attention mechanism. According to the method, the distribution difference between the bearing fault data source domain and the target domain can be effectively reduced, and the fault diagnosis precision of the model on the target domain is improved. The method specifically comprises the steps of constructing a deep transfer learning model combining a convolutional neural network, an attention mechanism and domain adversarial training, performing feature extraction and classification training on data of a source domain and a target domain, and optimizing the model by adopting cross entropy loss, domain adversarial loss and maximum mean difference loss. The method shows high diagnosis accuracy and good generalization ability under different working conditions, and is suitable for cross-domain fault detection and diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and specifically to a fault diagnosis method based on domain adversarial training and deep transfer learning of an attention mechanism. The method can effectively utilize the information of source domain data in dry vacuum pump bearing fault data and migrate it to target domain data to improve the accuracy of dry vacuum pump bearing fault diagnosis under different working conditions. The method is suitable for fault detection and diagnosis in the fields of rotating mechanical equipment, electronic equipment, etc. Background Art

[0002] In chip manufacturing and production, the stable operation of dry vacuum pump equipment is of great significance to ensure the continuity of the chip production process and product quality. Failure of the bearing of the dry vacuum pump may cause its operation to be unstable, and may cause shutdown alarms, resulting in interruption of individual chip batch processing, causing contamination of chambers or wafers, and may also cause safety accidents and economic losses. Therefore, chip manufacturers urgently need to ensure the reliable operation of dry vacuum pumps and minimize scheduling downtime to reduce the losses caused by downtime. Traditional fault diagnosis techniques often rely on specific domain knowledge and manual feature extraction, which limits the generalization ability in different equipment or working conditions. In recent years, with the development of artificial intelligence and big data technology, deep learning methods have gradually been applied to the field of rotating machinery fault diagnosis, providing a solution without manual feature extraction. However, when deep learning models are applied to different working conditions, they often face difficulties in cross-domain migration due to the difference in data distribution between the source domain and the target domain, resulting in a significant decrease in the diagnostic accuracy of the model in the target domain. This difference in the distribution of bearing fault data, the so-called "cross-domain problem", is particularly evident in actual industrial scenarios. For example, under different production conditions, such as different temperatures, speeds or loads, the vibration signal characteristics of the bearing fault of a dry vacuum pump may change significantly, making it difficult for the model trained in the source domain to achieve the expected fault diagnosis results in the target domain.

[0003] Therefore, to solve this problem, proposing an effective cross-domain transfer learning method is the key to improving the generalization ability of the diagnostic model. Transfer learning improves the performance of the model in the target domain by utilizing the knowledge of the source domain, reducing the dependence on a large amount of labeled data in the target domain. In order to reduce the distribution difference between the source domain and the target domain data, a domain adversarial training method is proposed, which guides the model to learn domain invariant features through an adversarial loss function. However, simple domain adversarial training still has certain limitations when facing highly complex working condition changes. The present invention proposes a deep transfer learning model that combines domain adversarial and attention mechanisms to better solve cross-domain problems and improve the fault diagnosis ability of the model. Summary of the invention

[0004] The purpose of the present invention is to provide a deep transfer learning fault diagnosis method based on domain adversarial training and attention mechanism to solve the performance degradation problem of current deep learning in cross-domain fault diagnosis. By introducing the attention mechanism and domain adversarial module into the network model, the present invention can effectively reduce the data distribution difference between the source domain and the target domain, realize cross-domain dry vacuum pump bearing fault diagnosis, and improve the generalization ability and diagnostic accuracy of the model. Specifically, the following steps are included:

[0005] (1) Constructing a bearing fault diagnosis network model for a dry vacuum pump, the present invention designs a deep transfer learning model that combines a convolutional neural network, a self-attention mechanism, a domain adversarial module, and a fault classifier. The network framework uses a convolutional layer as the main structure, and includes multiple convolutional layers and pooling layers, which are used to extract multi-level feature representations from the input signal. After each convolutional layer, an attention mechanism module is added to dynamically adjust the weights of different features. In addition, a gradient reversal layer is introduced for domain adversarial training to help the model learn domain invariant features. The domain classifier and the label classifier work in parallel, where the domain classifier is responsible for minimizing the distribution difference between the source domain and the target domain data, and the label classifier is used to classify the bearing fault type.

[0006] (2) Data preparation and preprocessing of faulty bearings. The bearing fault data in the source domain and the target domain are preprocessed, including data loading, sliding window segmentation, data standardization, and data set division. After data loading, the original bearing fault data is converted into a one-dimensional signal sequence and segmented into fault length samples using the sliding window technique. Each window contains 1024 data points. These samples are standardized. Finally, the processed bearing fault data is divided into a training set and a validation set in a ratio of 8:2.

[0007] (3) Feature extraction: Multiple convolutional layers and pooling layers are used to extract features from the time series data of the faulty bearing. An attention mechanism module is introduced after each convolutional layer to dynamically adjust its weight according to the different input features, so that the model pays more attention to the feature areas related to fault diagnosis.

[0008] (4) Domain adversarial training: Based on the feature extractor, a gradient reversal layer (GRL) and a domain classifier are introduced. GRL reverses the gradient direction during the back propagation process, guiding the feature extractor to generate features that are unfavorable to the domain classifier, thereby reducing the distribution difference of bearing fault data between the source domain and the target domain. The domain classifier performs binary classification on the input fault features, determines whether the input fault feature area comes from the source domain or the target domain, and uses the cross entropy loss to calculate the probability of domain classification error, thereby guiding the feature extractor to generate domain invariant features.

[0009] (5) Classifier training: The extracted features are input into the fault classifier, and the bearing fault data in the source domain and the target domain are used for joint training. The fault classifier predicts the fault category of each input sample through the cross entropy loss function.

[0010] (6) Design loss function model optimization. The present invention uses cross entropy loss, domain adversarial loss and loss based on maximum mean difference (MK-MMD) to optimize the model. Cross entropy loss is used for fault classification tasks to optimize classifier performance by minimizing the difference between the predicted results and the true labels. Domain adversarial loss guides the feature extractor to generate domain invariant features through GRL, reducing the distribution difference between the source domain and the target domain. MK-MMD loss is used to measure the distribution difference between the source domain and the target domain features and optimize the domain adaptability of the model. Model optimization continuously updates model parameters through the back propagation algorithm.

[0011] (7) Train the network model and use the processed source domain and target domain bearing fault data for network training. The training settings include batch size, optimizer selection, learning rate scheduling, etc. In each epoch, the source domain and target domain fault data are first forward propagated to calculate the classification loss and domain adversarial loss. Then the model parameters are updated through backpropagation. A dynamic learning rate scheduling strategy is adopted to gradually reduce the learning rate in the middle of training.

[0012] (8) Model verification: After training is completed, the model is applied to the test data of the target domain to evaluate its fault diagnosis performance under different working conditions.

[0013] Furthermore, the feature extraction network structure includes the following components: five convolution modules, each module includes a convolution layer, a batch normalization layer, a ReLU activation function and an attention mechanism module. Among them, the first convolution layer uses a convolution kernel of size 64, a stride of 16, and a padding of 24. Subsequent convolution layers use 3×3 convolution kernels, where the stride is 1. A batch normalization layer (BatchNorm) and a ReLU activation function are added after each convolution layer. An 'Attention' module is embedded in each convolution module, which generates 'query', 'key' and 'value' through the 'Conv1d' operation, and calculates the self-attention distribution of the feature. The 'Attention' module calculates the self-attention of the feature, outputs the weighted feature and linearly combines it with the original input to enhance the expressive ability of the key features.

[0014] Furthermore, the gradient reversal layer (GRL) implements domain adversarial training by dynamically adjusting the intensity of gradient reversal during the back-propagation process. Its detailed structure includes: an input feature mapping layer that receives and maintains the feature map output by the feature extractor; a reverse gradient operation module that reverses the gradient direction so that the feature extractor tends to generate features that are unfavorable to the domain classifier, thereby reducing the distribution difference between the source domain and the target domain; the calculation of the reversal coefficient is dynamically adjusted by the 'calc_coeff' function to meet the needs of different training stages. GRL multiplies the gradient of the input feature by a negative 'alpha' value to reverse the direction of the gradient. The calculation formula for the reversed gradient is:

[0015]

[0016] L is the loss function, x is the input feature, G(x) is the output of the gradient reversal layer, and alpha is the adjustment coefficient.

[0017] Furthermore, the structure of the domain classifier is composed of three layers of fully connected layers. The first fully connected layer maps the input features to the specified 'hidden_size' dimension (the default is 1024), and applies the ReLU activation function and Dropout to improve the model's expressiveness and prevent overfitting; the second fully connected layer maintains the 'hidden_size' dimension and further processes the features through ReLU activation and Dropout. The third fully connected layer maps the features to a binary output dimension, and finally maps the output value to the [0,1] range through the Sigmoid activation function as the probability value of the domain classification.

[0018] Furthermore, the structure of the fault classifier is different depending on whether a bottleneck layer is used. When a bottleneck layer is used, the fault classifier consists of a bottleneck layer and a fully connected layer. The bottleneck layer includes a linear layer, a ReLU activation function, and Dropout. The input 512-dimensional features are reduced to 256 dimensions through the bottleneck layer, and then mapped to the output space of the number of categories through the fully connected layer.

[0019] Furthermore, the domain adversarial loss function adopts a hybrid strategy combining adversarial loss and maximum mean difference (MK-MMD) loss, as follows:

[0020] Adversarial loss function: used for domain classifiers, it maximizes the ability of the feature extractor to generate domain inseparable features by minimizing the classification error rate of the source domain and the target domain in the domain classifier. Its formula is as follows:

[0021]

[0022] Where N is the batch size, d iis the domain label of sample i (source domain is 1, target domain is 0), q i The domain classifier predicts the probability that sample i belongs to the source domain.

[0023] MK-MMD loss function: used to measure the difference in feature distribution between the source domain and the target domain, and optimizes the domain adaptability of the model by calculating the maximum mean difference between the features of the two domains in the Reproducing Kernel Hilbert Space (RKHS). The formula is as follows:

[0024]

[0025] Where N s and N t are the number of samples in the source domain and the target domain respectively, and The samples φ(·) in the source domain and the target domain are respectively the mapping functions.

[0026] Overall loss function: By combining the adversarial loss and the MK-MMD loss, a total domain adversarial loss function is formed to better balance the classification task and the domain adaptation task. The total loss function formula is as follows:

[0027] L total =L CE +λ D L D +λ MMD L MMD (4)

[0028] Among them, L CE is the cross entropy loss, L D and L MMD are weight coefficients, which control the contribution of domain adversarial loss and MK-MMD loss respectively.

[0029] According to the technical solution proposed in this paper, we can use the bearing fault data in the source domain and the target domain to perform domain adversarial training and classification network training. The main goal of this network is to improve the model's fault diagnosis generalization ability in the target domain by learning the common features between the source domain and the target domain. Through this training method, we are able to build a model with generalization ability in cross-domain fault diagnosis. Under complex working conditions, this trained bearing fault diagnosis model can process input data from different distributions and predict its fault type. By extracting important features from the input signal, the model is able to achieve cross-domain fault diagnosis, thereby improving

[0030] Reliability and maintenance efficiency of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a fault diagnosis flow chart of the present invention;

[0032] Figure 2 It is a distribution diagram of the dataset of the present invention in the source domain and the target domain;

[0033] Figure 3 Schematic diagram of the structure of the attention mechanism module in the present invention;

[0034] Figure 4 This is a schematic diagram of the domain confrontation principle of the present invention;

[0035] Figure 5 The present invention designs a flow chart of loss function and model optimization;

[0036] Figure 6 Training flow chart for the present invention DETAILED DESCRIPTION

[0037] The present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention:

[0038] The fault diagnosis method of the present invention comprises the following parts: constructing a fault diagnosis network model, data preparation, feature extraction, domain adversarial training, classifier training, designing a loss function model optimization, training the network

[0039] Network model, model verification, the process of this method is as follows Figure 1 Shown

[0040] The main features of the deep transfer learning fault diagnosis method based on domain adversarial training and attention mechanism proposed in this paper are the construction of the fault diagnosis network and the two key links of training and migration between the source domain and the target domain. This method combines the characteristics of multiple network modules such as attention mechanism, domain adversarial principle, and classifier training.

[0041] The following are the detailed steps and features of the present invention:

[0042] Step 1: Build a fault diagnosis network model

[0043] The core of the present invention is to design a deep transfer learning model that combines convolutional neural network (CNN), attention mechanism and domain adversarial training. The main components of the model include:

[0044] Convolutional neural networks (CNNs) are used to extract multi-level feature representations from input signals. The CNN structure consists of multiple convolutional layers, pooling layers, and batch normalization layers. It extracts signal features layer by layer to capture information at different scales and levels. The attention mechanism is embedded after each convolution module to dynamically adjust the weight of the features and enhance the focus on the feature areas related to bearing fault diagnosis. The attention mechanism can guide the model to focus on key features and reduce noise interference by calculating the self-attention weights of the input features. The two parts of the network structure are combined into a feature extractor.

[0045] Gradient Reversal Layer (GRL): used for domain adversarial training. During the back-propagation process, GRL guides the feature extractor to generate features that are unfavorable to the domain classifier by reversing the loss gradient, reducing the distribution difference between the source domain and the target domain, and thus achieving the learning of domain-invariant features.

[0046] Domain classifier and fault classifier: The domain classifier is responsible for minimizing the distribution difference between the source domain and the target domain data, while the label classifier is used to classify the bearing fault types. The domain classifier and the fault classifier work in parallel to jointly optimize the generalization ability of the model.

[0047] Step 2: Dataset preparation and preprocessing

[0048] The bearing fault datasets of the source domain and the target domain are preprocessed separately. The specific operations include: loading the data, loading the original data into a one-dimensional signal sequence. Sliding window segmentation, segmenting the data into sliding windows, each window contains 1024 data points. Standardizing the data, standardizing the segmented signal samples to reduce the impact of the dimension difference of the data on the model training. Dividing the dataset, dividing the processed dataset into a training set and a validation set in a ratio of 8:2. Figure 2 As shown, this ensures the generalization ability during model training.

[0049] Step 3: Feature extraction and application of attention mechanism

[0050] Convolutional layers and pooling layers are used to extract features from time series data. After each convolutional layer, the attention mechanism module dynamically adjusts the weight of the feature so that the model can pay more attention to the feature area related to fault diagnosis. The attention mechanism module calculates the attention weight of the input feature to enhance the expression ability of the key features. Its structure diagram is shown in the figure below. Figure 3 As shown,.

[0051] Step 4: Domain Adversarial Training

[0052] Based on the feature extractor, the gradient reversal layer (GRL) and domain classifier are introduced to perform domain adversarial training. GRL guides the feature extractor to generate features that are difficult to distinguish between the source domain and the target domain by reversing the gradient direction, thereby reducing the distribution difference between them. The domain classifier performs binary classification on the input features to distinguish whether the features come from the source domain or the target domain, and uses the cross entropy loss to calculate the probability of domain classification error. The entire domain adversarial training process is as follows: Figure 4 shown.

[0053] Step 5: Classifier training

[0054] The extracted features are input into the fault classifier, and the data from the source domain and the target domain are used for joint training. The fault classifier predicts the fault category of each input sample through the cross entropy loss function. Since the model is trained on the source domain and the target domain at the same time, the fault classifier can learn more generalized features and ultimately achieve high-precision prediction of the target domain fault type. The training process is as follows: Figure 4 shown.

[0055] Step 6: Design loss function and model optimization

[0056] The present invention uses cross entropy loss, domain adversarial loss and maximum mean difference (MK-MMD) based loss to optimize the model. Specifically:

[0057] Cross entropy loss: used for fault classification tasks, optimizes classifier performance by minimizing the cross entropy loss between the predicted results and the true labels. Domain adversarial loss: guides the feature extractor to generate domain-invariant features through the gradient reversal layer, thereby reducing the distribution difference between the source domain and the target domain. MK-MMD loss: used to calculate the distance between the source domain and the target domain data features, further optimizing the domain adaptability of the model. Model optimization continuously updates the model parameters through the back-propagation algorithm to improve the generalization ability and diagnostic accuracy of the model. The loss function is designed as follows: Figure 5 shown.

[0058] Step 7: Train the network model

[0059] Use the source domain data and target domain data of the processed faulty bearings to train the network. The specific steps are as follows:

[0060] Training settings: The batch size is set to 64, and the Adam optimizer is used to optimize the feature extractor, fault classifier, and domain classifier respectively. Training process: The initial learning rate is set to 0.001. In each iteration, the source domain and target domain data are first forward propagated to calculate the classification loss and domain adversarial loss. Then the model parameters are updated through backpropagation. The training process lasts for 300 epochs, and the learning rate gradually decays after 150 epochs. Model saving: After the training is completed, the optimal model parameters are saved for subsequent testing and application. The training process is as follows Figure 6 shown.

[0061] Step 8: Model Validation and Performance Evaluation

[0062] After training, the model is applied to the test data of the target domain to evaluate its fault diagnosis performance under different working conditions. The generalization ability and diagnostic accuracy of the model are verified by calculating indicators such as classification accuracy and loss value.

[0063] Migration training under different speed conditions was performed on the Jiangnan University (JNU) dataset, as shown in Table 1. The domain adaptive model DANN performed relatively low in terms of fault recognition rate, while the model of the present invention still outperformed the DANN model in terms of fault recognition rate even when the attention mechanism was removed, but was slightly lower than the model including the attention mechanism. The method proposed in the present invention significantly improved the diagnostic accuracy in the target domain, with an average accuracy of 96.64%, which shows that the method of the present invention has obvious advantages and higher accuracy in cross-domain fault diagnosis, further verifying its effectiveness.

[0064] Table 1. Comparison of accuracy on the JNU dataset

[0065]

[0066] Note: [0][1][2] are the vibration signal fault data of different speed conditions in the JNU bearing data set, which are 600rpm, 800rpm, and 1000rpm respectively.

Claims

1. A vacuum dry pump bearing fault diagnosis method based on domain adversarial and attention transfer learning, characterized in that: The following steps are involved: (1) A bearing fault diagnosis network model for a dry vacuum pump is constructed; the fault diagnosis network model is a deep transfer learning model that combines a convolutional neural network, a self-attention mechanism, a domain adversarial module, and a fault classifier; the network framework of the fault diagnosis network model includes multiple convolutional layers and pooling layers for extracting multi-level feature representations from input signals; after each convolutional layer, an attention mechanism module is added for dynamically adjusting the weights of different features; a gradient reversal layer is introduced for domain adversarial training, and the fault diagnosis network model learns domain invariant features; The domain classifier and label classifier work in parallel, where the domain classifier is responsible for minimizing the distribution difference between the source domain and the target domain data, and the label classifier is used to classify the bearing fault type; (2) Data preparation and preprocessing of faulty bearings: The bearing fault data in the source domain and the target domain are preprocessed, including data loading, sliding window segmentation, data standardization, and data set division. After data loading, the original bearing fault data is converted into a one-dimensional signal sequence and segmented into fault length samples using the sliding window technology. Each window contains 1024 data points. The fault length samples are standardized; finally, the processed bearing fault data are divided into a training set and a validation set in a ratio of 8:2; (3) Feature extraction: Multiple convolutional layers and pooling layers are used to extract features from the time series data of the faulty bearing. An attention mechanism module is introduced after each convolutional layer to dynamically adjust its weight according to the different input features to focus on the feature areas related to fault diagnosis. (4) Domain adversarial training: Based on the feature extractor, a gradient reversal layer (GRL) and a domain classifier are introduced. GRL reverses the gradient direction during the back propagation process, guiding the feature extractor to generate features that are unfavorable to the domain classifier, thereby reducing the distribution difference of bearing fault data between the source domain and the target domain. The domain classifier classifies the input fault features into two categories, determines whether the input fault feature area comes from the source domain or the target domain, and uses the cross entropy loss to calculate the probability of domain classification error, thereby guiding the feature extractor to generate domain invariant features. (5) Fault classifier training: The extracted features are input into the fault classifier, and the bearing fault data in the source domain and the target domain are used for joint training. The fault classifier predicts the fault category of each input sample through the cross entropy loss function. (6) Design loss function model optimization, and use cross entropy loss, domain adversarial loss, and loss based on maximum mean difference MK-MMD to optimize the model; cross entropy loss is used for fault classification tasks to optimize classifier performance by minimizing the difference between the predicted results and the true labels; domain adversarial loss guides the feature extractor to generate domain invariant features through GRL, reducing the distribution difference between the source domain and the target domain; MK-MMD loss is used to measure the distribution difference between the source domain and the target domain features, and optimizes the domain adaptability of the model; model optimization updates model parameters through the back propagation algorithm; (7) Training the network model, using the processed source domain and target domain bearing fault data for network training; the training settings include batch size, optimizer selection, and learning rate scheduling; in each epoch, first forward propagation is performed on the source domain and target domain fault data to calculate the classification loss and domain adversarial loss; then the model parameters are updated through back propagation; a dynamic learning rate scheduling strategy is adopted to gradually reduce the learning rate in the middle of training; (8) Model verification: After training is completed, the model is applied to the test data in the target domain to evaluate the fault diagnosis performance under different working conditions.

2. The method for vacuum dry pump bearing fault diagnosis based on domain adversarial and attention transfer learning according to claim 1 is characterized in that: In the dry vacuum pump bearing fault diagnosis network model in step 1), the feature extraction network structure includes the following components: five convolution modules, each module includes a convolution layer, a batch normalization layer, a ReLU activation function and an attention mechanism module; wherein, the first convolution layer uses a convolution kernel of size 64, a stride of 16, and a padding of 24; and the subsequent convolution layers use a 3×3 convolution kernel, wherein the stride is 1; a batch normalization layer BatchNorm and a ReLU activation function are added after each convolution layer; an 'Attention' module is embedded in each convolution module, and the 'Attention' module generates 'query', 'key' and 'value' through the 'Conv1d' operation, and calculates the self-attention distribution of the feature, and the 'Attention' module calculates the self-attention of the feature, outputs the weighted feature and linearly combines it with the original input to enhance the ability to express key features.

3. The method for vacuum dry pump bearing fault diagnosis based on domain adversarial and attention transfer learning according to claim 1 is characterized in that: The gradient reversal layer GRL in step 4) realizes domain adversarial training by dynamically adjusting the intensity of gradient reversal during the back propagation process; the structure includes: an input feature mapping layer, which receives and maintains the feature map output by the feature extractor; a reverse gradient operation module, which makes the feature extractor tend to generate features that are unfavorable to the domain classifier by reversing the gradient direction, thereby reducing the distribution difference of bearing fault data between the source domain and the target domain; the calculation of the reversal coefficient is dynamically adjusted by the 'calc_coeff' function to meet the needs of different training stages; GRL multiplies the gradient of the input feature by a negative 'alpha' value to reverse the direction of the gradient; the calculation formula for the reverse gradient is: L is the loss function, x is the input feature, G(x) is the output of the gradient reversal layer, and alpha is the adjustment coefficient.

4. The method for vacuum dry pump bearing fault diagnosis based on domain adversarial and attention transfer learning according to claim 1 is characterized in that: The detailed structure of the domain classifier in step 4) includes: the structure of the domain classifier consists of three layers of fully connected layers; the first fully connected layer maps the input features to the specified 'hidden_size' dimension, and applies the ReLU activation function and Dropout to improve the expressiveness of the model and prevent overfitting; the second fully connected layer maintains the 'hidden_size' dimension, and further processes the features through ReLU activation and Dropout; the third fully connected layer maps the features to a binary output dimension, and finally maps the output value to the [0,1] range through the Sigmoid activation function as the probability value of the domain classification.

5. The method for vacuum dry pump bearing fault diagnosis based on domain adversarial and attention transfer learning according to claim 1 is characterized in that: The detailed structure of the fault classifier in step 5) includes: the structure of the fault classifier varies depending on whether a bottleneck layer is used; when a bottleneck layer is used, the fault classifier consists of a bottleneck layer and a fully connected layer; the bottleneck layer includes a linear layer, a ReLU activation function, and Dropout, and the input 512-dimensional features are reduced to 256 through the bottleneck layer, and then mapped to the output space of the number of categories through the fully connected layer.

6. The method for vacuum dry pump bearing fault diagnosis based on domain adversarial and attention transfer learning according to claim 1 is characterized in that: The detailed structure of the domain adversarial loss function in step 6) includes: the domain adversarial loss function adopts a hybrid strategy combining adversarial loss and maximum mean difference MK-MMD loss, as follows: Adversarial loss function: used for domain classifiers, it maximizes the ability of the feature extractor to generate domain inseparable features by minimizing the classification error rate of the source domain and the target domain in the domain classifier; the formula is as follows: Where N is the batch size, d i is the domain label of sample i, q i The domain classifier predicts the probability that sample i belongs to the source domain; MK-MMD loss function: used to measure the difference in fault feature distribution between the source domain and the target domain. It optimizes the domain adaptability of the model by calculating the maximum mean difference between the features of the two domains in the reproducing kernel Hilbert space RKHS. The formula is as follows: Where N s and N t are the number of samples in the source domain and the target domain respectively, and are samples of the source domain and the target domain, respectively; φ(·) is the mapping function; Overall loss function: By combining the adversarial loss and the MK-MMD loss, the overall domain adversarial loss function is formed to better balance the classification task and the domain adaptation task; the total loss function formula is as follows: L total =L CE +λ D L D +λ MMD L MMD (4) Among them, L CE is the cross entropy loss, L D and L MMD are weight coefficients, which control the contribution of domain adversarial loss and MK-MMD loss respectively.

Images