Single-source domain generalized bearing fault diagnosis method and system based on multi-Gaussian attention

Abstract

The present invention proposes a single-source domain generalized bearing fault diagnosis method and system based on multi-Gaussian attention, which relates to the technical field of bearing fault diagnosis. It includes obtaining the original data of the single-source domain of the bearing and inputting it into the single-source domain generalization model based on multi-Gaussian attention, using a feature extraction network to extract multi-scale convolution features, and using a multivariate Gaussian attention module to map the convolution features to the Gaussian feature space to obtain Gaussian features, normalizing the Gaussian features and weighting the convolution features to obtain weighted convolution features; and inputting the weighted convolution features into a classifier for classification. The present invention uses Gaussian functions and bimodal Gaussian functions to project multi-scale fault features into the Gaussian feature space, and allocates attention weights according to the degree of feature alignment of these functions, and uses covariance loss to ensure that the weighted features have different distribution parameters, thereby improving feature diversity and enabling the model to effectively generalize to unseen target domains.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular to a single-source domain generalized bearing fault diagnosis method and system based on multi-Gaussian attention. Background Art

[0002] The reliability and safety of industrial systems depend on effective fault diagnosis, especially for bearing fault diagnosis. Bearings provide support for rotating shafts and reduce friction. Therefore, timely and effective fault diagnosis of bearings is crucial to ensure the reliability of industrial systems. However, since bearings are prone to various faults, once a fault occurs, it will seriously disrupt normal industrial production activities. Therefore, accurate detection and diagnosis of bearing faults is crucial for predictive maintenance, reducing downtime and reducing maintenance costs.

[0003] Deep learning technology has been rapidly developed in intelligent fault diagnosis due to its ability to accurately learn complex patterns from large amounts of annotated data. Although some deep learning methods have achieved success, they often face challenges in real environments due to changes in operating conditions and environmental noise, leading to domain shift problems. Domain shift refers to the difference in statistical properties between training data (source domain) and test data (target domain). When a model trained on the source domain is applied to out-of-distribution data in the real world, it will lead to a decrease in the performance of the deep learning model.

[0004] To address the domain shift problem between training and test data, domain transfer techniques such as domain adaptation (DA) and domain generalization (DG) have been developed for intelligent fault diagnosis. DA methods in fault diagnosis use strategies such as maximum mean discrepancy (MMD) and adversarial learning to align the data distribution of the source and target domains. These methods improve the fault diagnosis performance of known source domains by learning features that are invariant across domains and effectively adapt to target tasks. However, DA methods often face challenges in practical deployment. For example, domain adaptation methods usually have a certain decline in effectiveness when generalizing to other target domains outside of their training.

[0005] Domain generalization is another transfer learning method. It uses multiple labeled source domain data to collaboratively construct a highly robust model without the involvement of target domain data. The main goal of fault diagnosis methods based on domain generalization is to identify and extract consistent and domain-independent features between different source domains. However, despite the many advantages of domain generalization methods, they still face some challenges in the field of fault diagnosis. For example, domain generalization methods require collecting different data samples from multiple source domains to participate in training to ensure the robustness of the trained model. However, since faults rarely occur, it is still difficult to collect a large amount of high-quality fault data under different working conditions in actual industrial applications.

[0006] In this context, single domain generalization (SDG) methods have been proposed and applied to the field of fault diagnosis. Most SDG methods focus on generating more data or features to improve the generalization ability of fault diagnosis models. Although data or feature enhancement is usually effective, these methods ignore the internal distribution relationship between fault data and fault types. Summary of the invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a single-source domain generalized bearing fault diagnosis method and system based on multi-Gaussian attention, which uses Gaussian functions and bimodal Gaussian functions to project multi-scale fault features into Gaussian feature space, and allocates attention weights according to the degree of feature alignment of these functions. At the same time, covariance loss is used to ensure that weighted features have different distribution parameters, thereby improving feature diversity and enabling the model to effectively generalize to unseen target domains.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect of the present invention provides a single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention.

[0010] The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention includes the following steps:

[0011] Obtain the original data of the single source domain of the bearing;

[0012] The original data of the single-source domain of the bearing is input into the single-source domain generalization model based on multi-Gaussian attention, and the multi-scale convolutional features are extracted using the feature extraction network. The convolutional features are mapped to the Gaussian feature space using the multi-Gaussian attention module to obtain Gaussian features. The Gaussian features are normalized and the convolutional features are weighted to obtain weighted convolutional features.

[0013] The weighted convolution features are input into the classifier for classification to complete the fault diagnosis of the bearing.

[0014] The second aspect of the present invention provides a single-source domain generalized bearing fault diagnosis system based on multi-Gaussian attention.

[0015] The single-source domain generalized bearing fault diagnosis system based on multi-Gaussian attention includes:

[0016] The data acquisition module is configured to: acquire the original data of the single source domain of the bearing;

[0017] The multi-Gaussian attention mechanism feature extraction module is configured to: input the original data of the single-source domain of the bearing into the single-source domain generalization model based on multi-Gaussian attention, use the feature extraction network to extract multi-scale convolution features, and use the multi-Gaussian attention module to map the convolution features into the Gaussian feature space to obtain Gaussian features, normalize the Gaussian features and weight the convolution features to obtain weighted convolution features;

[0018] The classification module is configured to: input the weighted convolution features into the classifier for classification to complete the fault diagnosis of the bearing.

[0019] One or more of the above technical solutions have the following beneficial effects:

[0020] In view of the problem that rolling bearings are prone to domain shift due to changes in operating conditions, the present invention provides a single-source domain generalized bearing fault diagnosis method and system based on multi-Gaussian attention. MGA-SDG (single-source domain generalization model based on multi-Gaussian attention) uses Gaussian functions and bimodal Gaussian functions to project multi-scale fault features into Gaussian feature space, and then uses normalized Gaussian features to weight convolution features, and assigns attention weights according to the degree of feature alignment of these functions. This process ensures that features that are more in line with a specific Gaussian distribution have a greater impact on fault classification. At the same time, the present invention uses covariance loss to ensure that weighted features have different distribution parameters, thereby improving feature diversity and enabling the model to effectively generalize to unseen target domains. Extensive experiments have been carried out on public rolling bearing datasets. The experimental results show that the method proposed in the present invention has higher performance and robustness, and has great potential in real-world industrial applications.

[0021] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0023] Figure 1This is a flow chart of the method of the first embodiment. DETAILED DESCRIPTION

[0024] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0025] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0026] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0027] Terminology explanation:

[0028] DA: Domain Adaptation;

[0029] DG: domain generalization;

[0030] MMD: Maximum mean difference;

[0031] SDG: Single Source Domain Generalization;

[0032] MGA-SDG: A single-source domain generalization model based on multi-Gaussian attention;

[0033] MGA: Multivariate Gaussian Attention Module;

[0034] MGAI module: Multivariate Gaussian Attention Module I;

[0035] MGAII module: Multivariate Gaussian Attention Module II.

[0036] Embodiment 1

[0037] As described in the previous background, most SDG methods focus on generating more data or features to improve the generalization ability of fault diagnosis models. Although data or feature enhancement is usually effective, these methods ignore the internal distribution relationship between fault data and fault types. If the generalizable distribution of most fault data can be learned in deep learning and described using these known distributions, the generalization ability of the model can be enhanced.

[0038] Gaussian and bimodal Gaussian functions can fit the convolution features of healthy bearings and describe the relationship between data and specific faults. Therefore, learning feature representations consistent with Gaussian and bimodal Gaussian functions can potentially improve the generalization ability of the model.

[0039] Inspired by this, the embodiment of the present invention proposes a single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention. Specifically, the present invention maps the convolution feature to a consistent distribution by using a Gaussian function and a bimodal Gaussian function, and then weights the convolution feature using the normalized Gaussian feature. In this process, it is ensured that the features that are more in line with a specific Gaussian distribution have a greater impact on fault classification.

[0040] In addition, the present invention uses covariance loss to differentiate weighted features to ensure that features with the same distribution function also have different distribution parameters, thereby increasing the diversity of features, generalizing the domain with a single source domain, and solving the domain shift problem in the bearing fault diagnosis process.

[0041] In the single-source domain generalization task, we have The source domain of samples used for training is defined as , where the data space is , Represents the i-th source domain data, where is the index of the data; the label space with the bearing health status is , task T is summarized as , including the label space and classifier , the above , , , , , The s in denoted the source domain. The invisible target domain is defined as ,in Represents the number of samples in the target domain. , , , In the t, the source is the target domain. In the single-source domain generalization task, due to different working conditions, the source domain and the target domain have different feature distributions, that is, The label space between the source domain and the target domain is consistent, that is, the fault types of all domains are consistent. The goal of the single-source domain generalization task is to use data from a single source domain to train and build a model that can adapt to new, unseen target domains.

[0042] Specifically, the single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention disclosed in this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0043] Obtain the original data of the single source domain of the bearing;

[0044] The original data of the single-source domain of the bearing is input into the single-source domain generalization model based on multi-Gaussian attention, and the multi-scale convolutional features are extracted using the feature extraction network. The convolutional features are mapped to the Gaussian feature space using the multi-Gaussian attention module to obtain Gaussian features. The Gaussian features are normalized and the convolutional features are weighted to obtain weighted convolutional features.

[0045] The weighted convolution features are input into the classifier for classification to complete the fault diagnosis of the bearing.

[0046] In some embodiments, the single-source domain raw data of the bearing is obtained, wherein the single-source domain raw data is raw data of a single source domain used for training, including Data and its corresponding The specific purpose is to obtain the label with A single source domain data with samples for training , where the data space is , the label space with the bearing health status is The feature extraction network includes a feature extraction network I and a feature extraction network II which are arranged in parallel, and the feature extraction network I has a larger convolution kernel than the feature extraction network II;

[0047] Convolutional feature extraction is performed using convolution kernels of different sizes in feature extraction network I and feature extraction network II.

[0048] In some embodiments, the method further includes using a segmentation attention module to fuse convolutional features extracted by feature extraction network I and feature extraction network II, specifically including:

[0049] The convolutional features extracted by feature extraction network I and feature extraction network II are divided into two equal parts along the channel dimension, and the corresponding parts of the convolutional features of the two feature extraction networks are fused to obtain two fused features;

[0050] A dynamic gating mechanism is used for the two fused features to determine the gating weights, and the gating weights are multiplied element-by-element with the two fused features to obtain two multi-scale convolutional features.

[0051] In some embodiments, the convolutional features are mapped to the Gaussian feature space using a multivariate Gaussian attention module to obtain Gaussian features, specifically including:

[0052] The multivariate Gaussian attention module includes a multivariate Gaussian attention module I and a multivariate Gaussian attention module II, wherein the multivariate Gaussian attention module I is used to map one of the convolutional features, and the multivariate Gaussian attention module II is used to map another convolutional feature to the Gaussian feature space, respectively, to obtain Gaussian features;

[0053] The multivariate Gaussian attention module I and the multivariate Gaussian attention module II both include ordinary Gaussian functions and bimodal Gaussian functions.

[0054] The process of this embodiment will be explained in detail below with reference to the accompanying drawings.

[0055] Figure 1 This is the overall flow chart of the present invention. The single-source domain generalization model based on multi-Gaussian attention in this embodiment includes two feature extractor modules (feature extraction network I and feature extraction network II), a segmentation attention module, two multi-Gaussian attention modules (multi-Gaussian attention module I and multi-Gaussian attention module II) and a classifier module.

[0056] In general, assuming Represents the i-th source domain data, where is the index of the data. In this embodiment, Figure 1 As shown, input single source domain original data, corresponding to Figure 1 The top box in the figure; the source domain features of all single-source domain original data are extracted by feature extraction network I with a larger convolution kernel (denoted as cnn1) and feature extraction network II with a smaller convolution kernel (denoted as cnn2). cnn1 and cnn2 use convolution kernels of different sizes to achieve multi-scale feature extraction. This process corresponds to Figure 1 "Feature extraction with multivariate Gaussian attention mechanism" in .

[0057] Afterwards, the split attention module is applied to cnn1 and cnn2 to fuse the features extracted by the two feature extraction networks. This process corresponds to Figure 1 "Divided attention" in .

[0058] Next, the multi-gaussian attention module (MGA) I and the multi-gaussian attention module II are used to map the convolutional features into the Gaussian feature space. Both the multi-gaussian attention module I and the multi-gaussian attention module II include ordinary Gaussian functions and bimodal Gaussian functions. The parameters in the Gaussian function are learned during the back propagation process. Then, the Gaussian features are normalized and used to weight the convolutional features, giving higher weights to those features that align with specific Gaussian functions. This process corresponds to Figure 1 "Multi-Gaussian Attention I" and "Multi-Gaussian Attention Module II" in .

[0059] In order to distinguish the features from MGAI and MGAII modules (i.e., Multivariate Gaussian Attention Module I and Multivariate Gaussian Attention Module II), the covariance loss is minimized. Finally, the classification network C is responsible for distinguishing the features of each class. This process corresponds to Figure 1 "Objective Function Optimization" in .

[0060] 1. Multivariate Gaussian Attention Mechanism

[0061] MGA-SDG uses a multivariate Gaussian attention module to enhance feature representation. This module allows the network to adaptively focus on the salient features of the input data, thereby improving the discriminability of feature representation and overall model performance. The convolution operation is defined as follows:

[0062]

[0063] Among them, cnn1 and cnn2 represent feature extraction network I and feature extraction network II respectively. and Represent the outputs of the two feature extraction networks cnn1 and cnn2 respectively, represents the i-th source domain data, i represents the index of the data, and The dimension is ,in The number of channels representing features, Represents the length of each feature vector.

[0064] In this embodiment, a split attention module is used to improve the representation ability of the model's extracted features. The attention module allows the network to adaptively fuse the features extracted by two different feature extraction networks cnn1 and cnn2, thereby further improving the feature representation ability and the performance of the overall model. Specifically, for each output and , divided into two equal parts along the channel dimension, denoted as , and then combine these parts, specifically:

[0065]

[0066] in represents the fused feature representation, and Concatenate represents the combination of the two parts. This fusion process enriches the feature representation by combining multi-scale features. It can be understood that the above for Divided into two equal parts along the channel dimension, for Divided into two equal parts along the channel dimension.

[0067] Then, the fusion features Adopt dynamic gating mechanism, gating weight Determined by the Softmax function:

[0068]

[0069] Add(·) represents the addition of two features element by element, and Softmax(·) represents the Softmax function. Then, the gate weights With fusion features Multiply element by element to get the transformed features , specifically:

[0070]

[0071] Where Multiply(·) represents an element-by-element multiplication operation.

[0072] The multivariate Gaussian attention module first converts the convolution features Transformed into Gaussian features. Two different types of Gaussian functions, ordinary Gaussian functions and bimodal Gaussian functions, are used to process feature distributions:

[0073]

[0074] in represent The corresponding trainable means in the Gaussian and bimodal Gaussian functions, represent The square of the trainable standard deviation in the corresponding Gaussian and bimodal Gaussian functions, represent The corresponding trainable means in the Gaussian and bimodal Gaussian functions, represent The square of the trainable standard deviation in the corresponding Gaussian and bimodal Gaussian functions, and Represent the ordinary Gaussian function and the bimodal Gaussian function respectively, and i represents the i-th source domain data Corresponding features, exp represents the exponential function.

[0075] The output of the multivariate Gaussian attention module is expressed as and ,in Represents the number of source domain samples. These Gaussian features are used to convolution features. Weighted, specifically:

[0076]

[0077] in Represents the denominator as a normal Gaussian feature right The weighted convolution features, Represents a bimodal Gaussian feature with a denominator right The weighted convolution features, Represents the denominator as a normal Gaussian feature right The weighted convolution features, Represents a bimodal Gaussian feature with a denominator right The weighted convolution features, The number of channels representing features. It is a minimum value to avoid the denominator being 0. Finally, the weighted convolution features are flattened and classified using Softmax.

[0078] 2. Optimization of the objective function of model training

[0079] The classification loss and the designed covariance loss are used to train the parameters of the model. The calculation formula is:

[0080]

[0081] in represents the batch size, represents the true label of the i-th data, Represents the classification prediction result of the i-th data.

[0082] The covariance loss function is used to ensure that the weighted features maintain a low correlation value, thereby controlling the degree of enhancement and ensuring that they are both related and different. Covariance loss between two multivariate Gaussian attention modules Defined as:

[0083]

[0084] in and is the output weighted convolution feature of the first MGA module (Multivariate Gaussian Attention Module I) for the i-th data, and Represents the output weighted convolution feature of the second MGA module (Multivariate Gaussian Attention Module II) for the i-th data, represents the batch size, Respectively represent The mean of .

[0085] Feature extraction networks cnn1 and cnn2, split attention module, MGA module and covariance loss are used to enhance the feature diversity and generalization ability of the model in fault intelligent diagnosis. These modules use two independent feature extractor paths and combine covariance loss to encourage feature diversity.

[0086] Finally, it is used to update the global loss of the entire model Defined as:

[0087]

[0088] In the present invention, Take 0.0001. These losses are used together to effectively use data from a single source domain to train a model with high robustness and strong generalization ability.

[0089] 3. Experiment

[0090] In order to verify the effectiveness of the embodiment of the present invention, the embodiment of the present invention carried out specific experiments on the Paderborn data set containing health and fault data. The data used by the method of this embodiment includes healthy bearing data, artificially damaged bearing data, and real damaged bearing data. The experiment uses bearing data in three different health states, namely healthy data (H), inner ring damage data (IR), and outer ring damage data (OR). The specific use of the data set is detailed in Table 1. The data of each health state was tested under four different working conditions, as shown in Table 2. In this embodiment, the original data is divided into 4096 sampling points of equal length and normalized before inputting into the trained network.

[0091] Table 1 Health status and codes of the Paderborn dataset

[0092]

[0093] Table 2 Four operating conditions of the Paderborn dataset

[0094]

[0095] In order to demonstrate the effectiveness of the proposed method, the MGA-SDG method is compared with existing SDG fault diagnosis methods, including a convolutional neural network (CNN) with two feature extractors and a classifier, a single domain generalization network based on adversarial mutual information (AMINet), and a multi-scale generative adversarial network (MSG-ACN). The specific experimental results are shown in Table 3.

[0096] Table 3 Comparison results with other methods on the Paderborn dataset

[0097]

[0098] The experimental results in Table 3 demonstrate that MGA-SDG achieves the highest classification accuracy in all tasks compared to other methods. For example, when trained on PD and tested on PB, MGA-SDG achieves a classification accuracy of 95.95%, significantly higher than all other comparison methods. This excellent performance is reflected in all tasks, proving that the MGA-SDG method has good robustness and excellent performance when facing fault diagnosis situations under different working conditions.

[0099] The excellent performance of the MGA-SDG method in the examples of the present invention in the face of highly challenging tasks demonstrates the potential of the present invention for real-world fault diagnosis. In general, the excellent performance of MGA-SDG in all tasks highlights its advanced design and its advantages in improving the accuracy of fault diagnosis.

[0100] Furthermore, this embodiment has multiple alternatives: remove the ordinary Gaussian module (denoted as G1), remove the bimodal Gaussian module (denoted as G2), remove the multivariate Gaussian attention mechanism (denoted as G3), remove the segmentation attention mechanism (denoted as G4), and remove the covariance loss (denoted as G5) from the single-source domain generalization model based on multi-Gaussian attention in this embodiment. The following examples will show and compare the implementation effects of different combinations. The experimental data set and evaluation criteria are the same as above. Table 4 shows the experimental results of the proposed method and alternatives on the Paderborn data set.

[0101] Table 4 Comparison results with alternatives on the Paderborn dataset

[0102]

[0103] It can be seen from Table 4 that the MGA-SDG of this embodiment can achieve the best result among all the alternatives.

[0104] Embodiment 2

[0105] This embodiment discloses a single-source domain generalized bearing fault diagnosis system based on multi-Gaussian attention.

[0106] The single-source domain generalized bearing fault diagnosis system based on multi-Gaussian attention includes:

[0107] The data acquisition module is configured to: acquire the original data of the single source domain of the bearing;

[0108] The multi-Gaussian attention mechanism feature extraction module is configured to: input the original data of the single-source domain of the bearing into the single-source domain generalization model based on multi-Gaussian attention, use the feature extraction network to extract multi-scale convolution features, and use the multi-Gaussian attention module to map the convolution features into the Gaussian feature space to obtain Gaussian features, normalize the Gaussian features and weight the convolution features to obtain weighted convolution features;

[0109] The classification module is configured to: input the weighted convolution features into the classifier for classification to complete the fault diagnosis of the bearing.

[0110] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0111] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention, characterized in that: The following steps are involved: Obtain the original data of the single source domain of the bearing; The original data of the single-source domain of the bearing is input into the single-source domain generalization model based on multi-Gaussian attention, and the multi-scale convolutional features are extracted using the feature extraction network. The convolutional features are mapped to the Gaussian feature space using the multi-Gaussian attention module to obtain Gaussian features. The Gaussian features are normalized and the convolutional features are weighted to obtain weighted convolutional features. The weighted convolution features are input into the classifier for classification to complete the fault diagnosis of the bearing; The single-source domain generalization model based on multi-Gaussian attention includes two feature extractor modules, a segmentation attention module, two multivariate Gaussian attention modules and a classifier module; The two feature extraction networks include a feature extraction network I and a feature extraction network II which are arranged in parallel, wherein the feature extraction network I has a larger convolution kernel than the feature extraction network II; and convolution feature extraction is performed using convolution kernels of different sizes of the feature extraction network I and the feature extraction network II; Use the split attention module to fuse the convolutional features extracted by feature extraction network I and feature extraction network II; The two multivariate Gaussian attention modules include a multivariate Gaussian attention module I and a multivariate Gaussian attention module II. The multivariate Gaussian attention module I is used to map one of the convolutional features, and the multivariate Gaussian attention module II is used to map the other convolutional feature to the Gaussian feature space, so as to obtain Gaussian features. The Gaussian features are normalized and used to weight the convolutional features, giving higher weights to features aligned with a specific Gaussian function; Covariance loss is minimized to distinguish features from Multivariate Gaussian Attention Module I and Multivariate Gaussian Attention Module II; The classifier module is responsible for distinguishing the features of each class.

2. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 1 is characterized in that: The split attention module is used to fuse the convolutional features extracted by feature extraction network I and feature extraction network II, including: The convolutional features extracted by feature extraction network I and feature extraction network II are divided into two equal parts along the channel dimension, and the corresponding parts of the convolutional features of the two feature extraction networks are fused to obtain two fused features; A dynamic gating mechanism is used for the two fused features to determine the gating weights, and the gating weights are multiplied element-by-element with the two fused features to obtain two multi-scale convolutional features.

3. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 1 is characterized in that: The multivariate Gaussian attention module I and the multivariate Gaussian attention module II both include ordinary Gaussian functions and bimodal Gaussian functions.

4. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 3 is characterized in that: The specific formula for obtaining Gaussian features is: in, are the two convolution features obtained; represent The corresponding trainable means in the Gaussian and bimodal Gaussian functions, represent The square of the trainable standard deviation in the corresponding Gaussian and bimodal Gaussian functions, represent The corresponding trainable means in the Gaussian and bimodal Gaussian functions, represent The square of the trainable standard deviation in the corresponding Gaussian and bimodal Gaussian functions, and They represent ordinary Gaussian function and bimodal Gaussian function respectively, i represents the i-th feature, and exp represents the exponential function.

5. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 4 is characterized in that: Normalize the Gaussian features and weight the convolution features to obtain weighted convolution features, including: in Represents the denominator as a normal Gaussian feature right The weighted convolution features, Represents a bimodal Gaussian feature with a denominator right The weighted convolution features, Represents the denominator as a normal Gaussian feature right The weighted convolution features, Represents a bimodal Gaussian feature with a denominator right The weighted convolution features, The number of channels representing features; A minimum value to avoid denominator being 0.

6. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 3 is characterized in that: Global loss for generalization model of single source domain based on multi-Gaussian attention Defined as: in, represents the classification loss; represents the covariance loss between the multivariate Gaussian attention module I and the multivariate Gaussian attention module II, is the weight parameter.

7. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 6 is characterized by: in represents the batch size, represents the true label of the i-th data, Represents the classification prediction result of the i-th data.

8. The single-source domain generalized bearing fault diagnosis method based on multi-Gaussian attention as claimed in claim 6 is characterized by: in and is the output weighted convolution feature of the multivariate Gaussian attention module I for the i-th data, and represents the output weighted convolution feature of the multivariate Gaussian attention module II for the i-th data, represents the batch size, Respectively represent The mean of .

9. Single-source domain generalized bearing fault diagnosis system based on multi-Gaussian attention, characterized by: include: The data acquisition module is configured to: acquire the original data of the single source domain of the bearing; The multi-Gaussian attention mechanism feature extraction module is configured to: input the original data of the single-source domain of the bearing into the single-source domain generalization model based on multi-Gaussian attention, use the feature extraction network to extract multi-scale convolution features, and use the multi-Gaussian attention module to map the convolution features into the Gaussian feature space to obtain Gaussian features, normalize the Gaussian features and weight the convolution features to obtain weighted convolution features; The classification module is configured to: input the weighted convolution features into the classifier for classification to complete the fault diagnosis of the bearing; The single-source domain generalization model based on multi-Gaussian attention includes two feature extractor modules, a segmentation attention module, two multivariate Gaussian attention modules and a classifier module; The two feature extraction networks include a feature extraction network I and a feature extraction network II which are arranged in parallel, wherein the feature extraction network I has a larger convolution kernel than the feature extraction network II; and convolution feature extraction is performed using convolution kernels of different sizes of the feature extraction network I and the feature extraction network II; Use the split attention module to fuse the convolutional features extracted by feature extraction network I and feature extraction network II; The two multivariate Gaussian attention modules include a multivariate Gaussian attention module I and a multivariate Gaussian attention module II. The multivariate Gaussian attention module I is used to map one of the convolutional features, and the multivariate Gaussian attention module II is used to map the other convolutional feature to the Gaussian feature space to obtain Gaussian features. The Gaussian features are normalized and used to weight the convolutional features, giving higher weights to features aligned with a specific Gaussian function; Covariance loss is minimized to distinguish features from Multivariate Gaussian Attention Module I and Multivariate Gaussian Attention Module II; The classifier module is responsible for distinguishing the features of each class.

Images