Decision boundary adaptive open set fault diagnosis method based on domain adversarial neural network

Abstract

The invention discloses a decision boundary adaptive open set fault diagnosis method based on a domain adversarial neural network, and the method comprises the steps: S1, obtaining a rolling bearing fault data set under multiple working conditions and multiple fault states, and dividing the rolling bearing fault data set into source domain data and target domain data; s2, establishing a bearing fault diagnosis model, wherein the bearing fault diagnosis model can determine a decision boundary of a known fault state; performing iterative training on the bearing fault diagnosis model, and when the total loss function is converged, obtaining a trained bearing fault diagnosis model; the total loss function comprises a Bhatcharyya coefficient loss function, and the Bhatcharyya coefficient loss function is utilized to reduce the feature overlapping degree of different types of faults; and S3, based on the trained bearing fault diagnosis model, carrying out fault diagnosis on unknown faults in the actually collected bearing operation signals. According to the method, cross-working-condition open-set fault diagnosis can be realized, and the fault diagnosis capability is remarkably improved by ensuring accurate identification of unknown faults.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a decision boundary adaptive open set fault diagnosis method based on a domain adversarial neural network. Background Art

[0002] Rolling bearings play a vital role in the operation of rotating machinery and are commonly used in most types of rotating machinery, such as transmissions, aircraft engines, wind turbines, and high-speed trains. Because rolling bearings often operate under challenging conditions, such as heavy and variable loads, they have a higher failure rate than other components. Bearing failures can cause machine or even entire system downtime, resulting in significant financial losses and production delays. Therefore, bearing fault diagnosis is a key research area in machinery health management and a key focus in actual production processes.

[0003] Existing bearing fault diagnosis technologies mainly include traditional signal processing (such as time-frequency analysis, envelope demodulation) and intelligent diagnostic methods (such as machine learning, deep learning). Traditional signal processing methods rely on manual feature extraction and are suitable for stable working conditions; deep learning methods based on CNN, LSTM or Transformer can automatically learn fault characteristics and are suitable for complex scenarios. Hybrid solutions (signal processing + AI) and emerging technologies (edge ​​computing, multi-sensor fusion) can further improve early fault detection capabilities and robustness. However, existing bearing fault diagnosis technologies still have several key problems, such as: traditional methods rely on expert experience and have limited detection capabilities for early weak faults and variable operating conditions; many existing deep learning studies usually assume that the training samples and test samples in the experiment are in the same feature space and follow the same feature distribution, but these assumptions do not hold true in actual bearing operations.

[0004] With the development of fault diagnosis technology, industrial needs have put forward higher requirements on fault diagnosis capabilities. Open set fault diagnosis and closed set fault diagnosis represent two fundamentally different fault identification and classification methods in machine learning. In the closed set identification framework, training and test data sets are extracted from predefined class sets. Assuming that the fault types in the training data set and the test data set are consistent, the training model recognizes known fault categories. Most current studies have achieved high recognition accuracy in this case. It is difficult to obtain a comprehensive data set covering all possible types of faults in a real production environment. Therefore, in comparison, open set fault diagnosis processing is suitable for more complex fault categories and more realistic industrial scenarios. In this case, not only known faults need to be classified, but also unknown bearing faults need to be detected, such as Figure 4 As shown, if unknown bearing faults cannot be identified without training, the faults cannot be discovered and repaired in time in practice. Summary of the Invention

[0005] The application provides a decision boundary adaptive open set fault diagnosis method based on a domain adversarial neural network to overcome the technical problem that existing models cannot identify unknown faults.

[0006] To achieve the above-mentioned purpose, the technical scheme of the application is:

[0007] A bearing open set fault diagnosis method based on a domain adversarial neural network and a decision boundary adaptation, comprising the following specific steps:

[0008] S1: Obtain a rolling bearing fault data set under multiple working conditions and multiple fault states, divide the rolling bearing fault data set into source domain data and target domain data according to different working conditions, wherein the source domain data includes a plurality of known fault state data, and the target domain data includes a plurality of known fault state data and a plurality of unknown fault state data; use a plurality of known fault state data in the source domain data and part of the known fault state data in the target domain data as a training set, and use the remaining known fault state data and the plurality of unknown fault state data in the target domain data as a test set;

[0009] S2: Establish a bearing fault diagnosis model based on a domain adversarial neural network, wherein the bearing fault diagnosis model can determine the decision boundary of the known fault state to realize open set fault diagnosis;

[0010] Set a total loss function of the bearing fault diagnosis model, and iteratively train the bearing fault diagnosis model based on the training set and the test set, and obtain the trained bearing fault diagnosis model when the total loss function converges;

[0011] The total loss function of the bearing fault diagnosis model includes a Bhattacharyya coefficient loss function, a label prediction loss function and a domain classification loss function;

[0012] S3: Perform fault diagnosis on unknown faults in an actual collected bearing operation signal based on the trained bearing fault diagnosis model.

[0013] Further, the bearing fault diagnosis model includes a feature extraction module, a label predictor, a domain classifier and an unknown fault diagnosis module:

[0014] The feature extraction module is used to extract vibration signal features in the training set and the test set and aggregate the same category fault features to obtain vibration signal aggregated features, and the vibration signal aggregated features are transmitted to the label predictor, the domain classifier and the unknown fault diagnosis module respectively;

[0015] The unknown fault diagnosis module uses the vibration signal aggregated features extracted based on the training set to obtain a decision boundary, and identifies unknown faults according to the decision boundary;

[0016] The label predictor is used to map the aggregated features of the vibration signal extracted based on the training set and the test set to the corresponding fault category; at the same time, the feature extraction module is driven by the reverse gradient propagation of the label prediction loss function to learn the discriminative features related to the health state to minimize the label prediction loss function;

[0017] The domain classifier is used to reduce the distribution difference between source domain data and target domain data through an adversarial training mechanism, and at the same time backpropagate the domain classification loss function through a gradient reversal layer to update the parameters of the feature extraction module, thereby maximizing the domain classification loss function.

[0018] Furthermore, the unknown fault diagnosis module obtains a decision boundary using aggregated features of vibration signals extracted based on the training set, and identifies unknown faults based on the decision boundary. The specific steps include:

[0019] Calculate the category center of each known fault state based on the aggregated features of the vibration signal extracted from the training set;

[0020] Calculate the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state;

[0021] A threshold is set according to the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state, and the threshold is used as the decision boundary of the known fault state;

[0022] The samples whose distance from the category centers of all known fault states in the aggregated features of the vibration signals extracted based on the test set exceeds the decision boundary are judged as unknown faults.

[0023] Furthermore, the feature extraction module includes a convolutional neural network (CNN), a graph generation layer (GGL), and a multi-view graph convolutional neural network (MFR-GCN);

[0024] The feature extraction module extracts vibration signal features from the training set and the test set and aggregates the same-category fault features to obtain the aggregated vibration signal features. The specific steps include:

[0025] Learning discriminative feature representations from the training set and the test set through the hierarchical convolution operation of the convolutional neural network (CNN) to generate features that can confuse the domain classifier;

[0026] The features that can confuse the domain classifier are converted into graph data through the graph generation layer GGL to capture the topological relationship between features and transmit them to the multi-view graph convolutional neural network MFR-GCN;

[0027] The multi-view graph convolutional neural network MFR-GCN aggregates and enhances the graph data transmitted by the graph generation layer GGL to obtain the vibration signal aggregation features.

[0028] Furthermore, the expression of the multi-view graph convolutional neural network MFR-GCN is:

[0029]

[0030] Among them, H represents the aggregated features of the vibration signal; M is the graph data transmitted by the graph generation layer GGL, l represents the total number of receptive fields; K1, K2 and K l Indicates the size of each receptive field, k1, k2 and k l Indicates the polynomial order, parameter and Corresponding to the weights of the three MFR-GConv layers in the multi-view graph convolutional neural network MFR-GCN, each layer is associated with a different receptive field size; is a Chebyshev polynomial of order k, is the rescaled Laplacian matrix.

[0031] Furthermore, the Bhattacharyya coefficient loss function is expressed as:

[0032]

[0033] in, is the Bhattacharyya coefficient, ranging from 0 to 1; P i (k) and Q j (k) are the probabilities of distribution P and Q in the i-th or j-th category respectively, Z represents the number of fault categories, k is a constant; ε is a constant.

[0034] Furthermore, the total loss function is expressed as:

[0035]

[0036] Among them, μ is the weight coefficient;

[0037]

[0038] Where n represents the number of source domain samples, n' represents the number of target domain samples, N is the total number of samples, and λ is the weight coefficient; G f (·;θ f ) indicates that the parameter is θ f The output result of the feature extraction module, G y (·;θ y ) indicates that the parameter is θ yThe output of the label predictor, x v ,y v Represent the input data and labels respectively, L y Represents the classification loss as the negative log probability of the correct label; G d (·;θ d ) indicates that the parameter is θ d The output of the domain classifier, d v Represents the binary variable of the vth input data, d v =0 or 1; Represents the operation of a gradient reversal layer.

[0039] Beneficial Effects: This invention provides source domain data consisting of several categories of known fault state data and target domain data consisting of several categories of known fault state data and several categories of unknown fault state data to train a bearing fault diagnosis model based on a domain adversarial neural network. The trained bearing fault diagnosis model can determine the decision boundary of known fault states to achieve open-set fault diagnosis across operating conditions. The Bhattacharyya coefficient loss function can reduce the overlap of features across different fault categories. By ensuring the accurate identification of unknown faults, this invention significantly improves fault diagnosis capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 This is a flow chart of a decision boundary adaptive open set fault diagnosis method based on domain adversarial neural network in the present invention;

[0042] Figure 2 1 is a flowchart of a bearing fault diagnosis model according to an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the operation of the multi-view graph convolution layer in an embodiment of the present invention;

[0044] Figure 4 Schematic diagram of traditional fault diagnosis and open set fault diagnosis in an embodiment of the present invention;

[0045] Figure 5 This is a confusion matrix diagram of the experimental results on the CWRU dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of 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. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] This embodiment provides a decision boundary adaptive open set fault diagnosis method based on domain adversarial neural network, such as Figure 1 and Figure 2 As shown, the specific steps include:

[0048] S1: Acquire a rolling bearing fault dataset under multiple working conditions and multiple fault states, and divide the rolling bearing fault dataset into source domain data and target domain data according to different working conditions, wherein the source domain data includes several types of known fault state data, and the target domain data includes several types of known fault state data and several types of unknown fault state data; use several types of known fault state data in the source domain data and part of the known fault state data in the target domain data as a training set, and use the remaining known fault state data and several types of unknown fault state data in the target domain data as a test set;

[0049] S2: establishing a bearing fault diagnosis model based on a domain adversarial neural network, setting a total loss function of the domain adversarial neural network, iteratively training the bearing fault diagnosis model based on the training set and the test set, and obtaining a trained bearing fault diagnosis model when the total loss function converges;

[0050] The total loss function of the bearing fault diagnosis model includes a Bhattacharyya coefficient loss function, a label prediction loss function, and a domain classification loss function;

[0051] S3: Based on the trained bearing fault diagnosis model, the known faults in the actually collected bearing operation signals are classified and unknown faults are identified.

[0052] In a specific embodiment, the bearing fault diagnosis model includes a feature extraction module, a label predictor (LabelPredictor), a domain classifier (DomainClassifier) ​​and an unknown fault diagnosis module (Unknown faultdiagnosis):

[0053] The feature extraction module is used to extract vibration signal features from the training set and the test set and aggregate fault features of the same category to obtain vibration signal aggregate features, and transmit them to the label predictor, domain classifier and unknown fault diagnosis module respectively;

[0054] The unknown fault diagnosis module obtains a decision boundary using aggregated features of vibration signals extracted based on a training set, and identifies unknown faults based on the decision boundary;

[0055] The label predictor is used to map the aggregated features of the vibration signal extracted based on the training set and the test set to the corresponding fault category; at the same time, the feature extraction module is driven by the reverse gradient propagation of the label prediction loss function to learn the discriminative features related to the health state to minimize the label prediction loss function;

[0056] The domain classifier is used to reduce the distribution difference between source domain data and target domain data through an adversarial training mechanism, and at the same time backpropagates the domain classification loss function through a gradient reversal layer (GRL) to update the parameters of the feature extraction module, thereby maximizing the domain classification loss function and ensuring the alignment of feature distributions between the two domains.

[0057] Specifically, the domain classifier loss gradient is reversed through the gradient reversal layer, and stochastic gradient descent is applied to uniformly optimize all parameters, dynamically balancing feature discriminability and domain invariance, thereby optimizing the bearing fault diagnosis model parameters.

[0058] Specifically, because the operating conditions of devices vary across different tasks, the feature distributions of training and test data may not be completely consistent. This discrepancy can reduce the generalization ability of trained diagnostic models. Domain adaptation techniques, however, can learn fault diagnosis and domain classification even when there is a shift in feature distribution between source and target domain data. They leverage labeled data from one or more related source domains to perform tasks in the target domain.

[0059] Specifically, the bearing fault diagnosis model proposed in this embodiment can effectively learn domain-invariant features and the ability to adapt to displacements in different domains, establish a mapping between source domain data and target domain data, and by combining the learned mapping, use labeled data from one or more related source domains to perform tasks in the target domain, reducing the distribution gap between the source domain and the target domain, so that the classifier trained in the source domain can be effectively applied to the target domain, ensuring that the trained bearing fault diagnosis model can be well generalized across different domains.

[0060] In a specific embodiment, the unknown fault diagnosis module obtains a decision boundary using aggregated features of vibration signals extracted based on a training set, and identifies unknown faults based on the decision boundary. The specific steps include:

[0061] Calculate the category center of each known fault state based on the aggregated features of the vibration signal extracted from the training set;

[0062] Specifically, calculating the category center is to calculate the arithmetic mean of all sample features of the same type of fault.

[0063] Calculate the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state;

[0064] A threshold is set according to the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state, and the threshold is used as the decision boundary of the known fault state;

[0065] The samples whose distance from the class centers of all known fault states in the aggregated features of vibration signals extracted based on the test set exceeds the decision boundary are judged as unknown faults to achieve open set fault diagnosis.

[0066] In a specific embodiment, the feature extraction module includes a convolutional neural network CNN, a graph generation layer GGL and a multi-view graph convolutional neural network MFR-GCN;

[0067] The feature extraction module extracts vibration signal features from the training set and the test set and aggregates the same-category fault features to obtain the aggregated vibration signal features. The specific steps include:

[0068] Learning discriminative feature representations from the training set and the test set through the hierarchical convolution operation of the convolutional neural network (CNN) to generate features that can confuse the domain classifier;

[0069] Specifically, features that can confuse domain classifiers are generated, that is, the feature distributions of similar fault signals in different domains (such as different load conditions) after passing through CNN should be as close as possible.

[0070] Specifically, CNN (Convolutional Neural Network) is a deep learning architecture specifically designed for processing structured grid data, such as images. The design of CNN is inspired by the organization of neurons in the biological brain and, in some respects, is similar to traditional neural networks. By combining structures such as convolutional and pooling layers, CNN demonstrates powerful capabilities in parameter learning and feature extraction. Furthermore, CNN can simultaneously learn feature extraction and classification layers, resulting in a model output that is not only well-structured but also strongly dependent on the learned features.

[0071] The features that can confuse the domain classifier (the features that can confuse the domain classifier are 1D sequence features) are converted into graph data through the graph generation layer GGL to capture the topological relationship between features and transmit them to the multi-view graph convolutional neural network MFR-GCN;

[0072] The graph data transmitted by the graph generation layer GGL is aggregated and enhanced by the multi-receptive field graph convolutional neural network MFR-GCN (Multireceptive Field Graph Convolutional Network) to obtain the vibration signal aggregation feature.

[0073] Specifically, the Graph Convolutional Network (GCN) extends the concept of CNN to graph data, enabling the aggregation of features of adjacent nodes to generate new node representations. However, the limitation of GCN is that it can only aggregate information from nodes that are very close to a given node, and the use of unweighted graphs may change the relationship between nodes. In order to address these limitations, this embodiment proposes a new MRF-GCN to enhance the robustness of node feature extraction. The role of MFR-GCN is to further aggregate and enhance the graph data transmitted by the graph generation layer GGL from the perspective of feature correlation and multi-scale topological structure, strengthen the feature aggregation between the same fault category, enhance the fault diagnosis effect, and thus improve the discriminability and domain invariance of fault features. This embodiment combines MRF-GCN with the Bhattacharyya coefficient to optimize the aggregation of various fault features and adjust the decision boundary range by adjusting the parameters of the bearing fault diagnosis model, so that the system can better adapt to different working conditions and minimize the overlap between different types of faults.

[0074] In a specific embodiment, the spectral domain graph filter can be expressed as a K-order polynomial, which means that the Laplacian operator has strict localization within K points. In the graph convolution operation, the parameter K is usually called the receptive field or kernel size. In order to overcome the limitation of the traditional graph convolution layer that cannot effectively capture multi-scale or multi-domain information, such as Figure 3 As shown in FIG, this embodiment aggregates information from different receptive fields through the multi-view graph convolution layer (MRF-GConv) in MRF-GCN.

[0075] Specifically, the core concept of the MRF-GConv layer is to collect information from different neighborhoods and integrate these features into a more robust and comprehensive representation. Structurally, it consists of multiple graph convolutional layers operating in parallel, each with a different receptive field, allowing the model to capture information from different neighborhoods. After extracting features from these different neighborhoods, the outputs are integrated and used as an enhanced node representation. Therefore, the expression of the multi-view graph convolutional neural network MFR-GCN is:

[0076]

[0077] Among them, H represents the aggregated features of the vibration signal; M is the graph data transmitted by the graph generation layer GGL, l represents the total number of receptive fields; K1, K2 and K l Indicates the size of each receptive field, k1, k2 and k l Indicates the polynomial order, parameter and Corresponding to the weights of the three MFR-GConv layers in the multi-view graph convolutional neural network MFR-GCN, each layer is associated with a different receptive field size; is a Chebyshev polynomial of order k, is the rescaled Laplacian matrix.

[0078] Specifically, the size of the receptive field depends on the value of K. When K = 1, the features of the node itself are aggregated. When K = 2, the aggregated features are the features of the adjacent nodes. When K = 3, the features of the secondary adjacent nodes are aggregated.

[0079] In a specific embodiment, the Bhattacharyya coefficient loss function is expressed as:

[0080]

[0081] in, is the Bhattacharyya coefficient, ranging from 0 to 1; P i (k) and Q j (k) are the probabilities of normalized discrete distributions P and Q in the i-th or j-th category, respectively. Z represents the number of fault categories. k is a constant. ε is a constant.

[0082] Specifically, the Bhattacharyya coefficient is a statistical measure used to quantify the similarity between two normalized probability distributions. It provides a way to assess the degree of overlap between distributions, making it valuable in fields such as pattern recognition, machine learning, and information theory.

[0083] Specifically, this embodiment uses the Bhattacharyya coefficient to measure the degree of feature overlap of different types of faults, and takes the Bhattacharyya coefficient loss as part of the overall loss, promotes the parameter update of the bearing fault diagnosis model through back propagation, and reduces feature overlap.

[0084] In a specific embodiment, the total loss function is expressed as:

[0085]

[0086] Among them, μ is the weight coefficient;

[0087]

[0088] Where n represents the number of source domain samples, n' represents the number of target domain samples, N is the total number of samples, and λ is the weight coefficient; G f (·;θ f ) indicates that the parameter is θ f The output result of the feature extraction module, G y (·;θ y ) indicates that the parameter is θ y The output of the label predictor, x v ,y v Represents input data and labels respectively, L y Represents the classification loss as the negative log probability of the correct label; G d (·;θ d ) indicates that the parameter is θ d The output of the domain classifier, d v Represents the binary variable of the vth input data, d v =0 or 1; Represents the operation of a gradient reversal layer.

[0089] In this embodiment, a comparative experiment is conducted based on a rolling bearing vibration signal dataset provided by the Case Western Reserve University (CWRU) Bearing Data Center, and the corresponding results are analyzed in detail to illustrate the performance improvement achieved by the method proposed in this embodiment.

[0090] Specifically, based on sampling rate and bearing fault location, the CWRU dataset is divided into four groups: normal baseline data (NB), 12k drive-end bearing fault data (12k DE), 48k drive-end bearing fault data (48k DE), and fan-end bearing fault data (FE). Each dataset is further divided into rolling element faults (RF), inner race faults (IF), and outer race faults (OF), where outer race faults include faults at the 6 o'clock, 3 o'clock, and 12 o'clock positions. This example uses samples collected under 1HP and 3HP loads as source domain data (operating condition 1 corresponds to a motor speed of 1730 rpm and a motor load of 3HP, or a motor speed of 1772 rpm and a motor load of 1HP) and samples collected under a 2HP load as target domain data (operating condition 2 corresponds to a motor speed of 1750 rpm and a motor load of 2HP). The training set includes data for nine fault states, and the test set includes data for all nine fault states and one or two unknown fault states.

[0091] To evaluate the effectiveness of the proposed Bhattacharyya coefficient in improving open-set fault diagnosis, a series of comparative experiments were conducted. The corresponding confusion matrix was generated to quantitatively evaluate the classification performance of the bearing fault diagnosis model. This example selected several representative distance measurement methods as benchmarks for comparison:

[0092] Euclidean distance (ED) represents the straight-line distance between two points in Euclidean space. However, when two points are equidistant from the cluster centroid, they are likely to be assigned to the same cluster regardless of their orientation, potentially ignoring potential structural differences.

[0093] Cosine distance (CD) is used to quantify the directional difference between two vectors. In machine learning, this metric is often used to compare sample vectors because it emphasizes the direction of the vectors rather than their magnitude or Euclidean distance.

[0094] Mahalanobis distance (MD) measures the distance between a point and a probability distribution. Unlike Euclidean distance, it accounts for the correlation between features, making it more suitable for assessing the similarity between multivariate samples.

[0095] Specifically, the confusion matrix of the experimental results on the CWRU dataset is as follows: Figure 5 The comparison results are summarized in Table 1. To ensure fair comparison between methods, all models are trained with the same common parameter settings, and the number of training times is fixed to 100 to ensure sufficient optimization. Experimental results on the CWRU dataset show that the proposed method is effective in classifying known faults and identifying unknown faults.

[0096] Specifically, comparative experiments with methods such as Euclidean distance show that the open-set fault diagnosis method combined with the Bhattacharyya coefficient outperforms other methods. Although the Euclidean distance-based method achieved higher accuracy in some test scenarios, its overall average performance was still slightly lower than that of the proposed method. Mahalanobis distance and cosine distance also performed slightly worse than the Bhattacharyya coefficient method in distinguishing between categorical features, affecting the model's diagnostic performance for unknown faults and resulting in lower fault identification accuracy.

[0097] Table 1 Comparative experimental results on the CWRU dataset

[0098]

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation, characterized by: The specific steps include: S1: Acquire a rolling bearing fault dataset under multiple working conditions and multiple fault states, and divide the rolling bearing fault dataset into source domain data and target domain data according to different working conditions, wherein the source domain data includes several types of known fault state data, and the target domain data includes several types of known fault state data and several types of unknown fault state data; use several types of known fault state data in the source domain data and part of the known fault state data in the target domain data as a training set, and use the remaining known fault state data and several types of unknown fault state data in the target domain data as a test set; S2: Establishing a bearing fault diagnosis model based on a domain adversarial neural network. The bearing fault diagnosis model can determine the decision boundary of known fault states to achieve open set fault diagnosis. Setting a total loss function of a bearing fault diagnosis model, iteratively training the bearing fault diagnosis model based on the training set and the test set, and obtaining a trained bearing fault diagnosis model when the total loss function converges; The total loss function of the bearing fault diagnosis model includes a Bhattacharyya coefficient loss function, a label prediction loss function, and a domain classification loss function; S3: Perform fault diagnosis on unknown faults in the actually collected bearing operation signal based on the trained bearing fault diagnosis model.

2. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 1 is characterized in that: The bearing fault diagnosis model includes a feature extraction module, a label predictor, a domain classifier and an unknown fault diagnosis module: The feature extraction module is used to extract vibration signal features from the training set and the test set and aggregate fault features of the same category to obtain vibration signal aggregate features, and transmit them to the label predictor, domain classifier and unknown fault diagnosis module respectively; The unknown fault diagnosis module obtains a decision boundary using aggregated features of vibration signals extracted based on a training set, and identifies unknown faults based on the decision boundary; The label predictor is used to map the aggregated features of the vibration signal extracted based on the training set and the test set to the corresponding fault category; at the same time, the feature extraction module is driven by the reverse gradient propagation of the label prediction loss function to learn the discriminative features related to the health state to minimize the label prediction loss function; The domain classifier is used to reduce the distribution difference between source domain data and target domain data through an adversarial training mechanism, and at the same time backpropagate the domain classification loss function through a gradient reversal layer to update the parameters of the feature extraction module, thereby maximizing the domain classification loss function.

3. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 2 is characterized in that: The unknown fault diagnosis module obtains a decision boundary using the aggregated features of the vibration signal extracted based on the training set, and identifies unknown faults based on the decision boundary. The specific steps include: Calculate the category center of each known fault state based on the aggregated features of the vibration signal extracted from the training set; Calculate the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state; A threshold is set according to the distance between the aggregated features of the vibration signal extracted based on the training set and the class center of each known fault state, and the threshold is used as the decision boundary of the known fault state; The samples whose distance from the category centers of all known fault states in the aggregated features of the vibration signals extracted based on the test set exceeds the decision boundary are judged as unknown faults.

4. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 3 is characterized in that: The feature extraction module includes a convolutional neural network (CNN), a graph generation layer (GGL), and a multi-view graph convolutional neural network (MFR-GCN); The feature extraction module extracts vibration signal features from the training set and the test set and aggregates the same-category fault features to obtain the aggregated vibration signal features. The specific steps include: Learning discriminative feature representations from the training set and the test set through the hierarchical convolution operation of the convolutional neural network (CNN) to generate features that can confuse the domain classifier; The features that can confuse the domain classifier are converted into graph data through the graph generation layer GGL to capture the topological relationship between features and transmit them to the multi-view graph convolutional neural network MFR-GCN; The multi-view graph convolutional neural network MFR-GCN aggregates and enhances the graph data transmitted by the graph generation layer GGL to obtain the vibration signal aggregation features.

5. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 4 is characterized in that: The expression of the multi-view graph convolutional neural network MFR-GCN is: Among them, H represents the aggregated features of the vibration signal; M is the graph data transmitted by the graph generation layer GGL, l represents the total number of receptive fields; K1, K2 and K l Indicates the size of each receptive field, k1, k2 and k l Indicates the polynomial order, parameter and Corresponding to the weights of the three MFR-GConv layers in the multi-view graph convolutional neural network MFR-GCN, each layer is associated with a different receptive field size; is a Chebyshev polynomial of order k, is the rescaled Laplacian matrix.

6. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 5 is characterized in that: The Bhattacharyya coefficient loss function is expressed as: in, is the Bhattacharyya coefficient, ranging from 0 to 1; P i (k) and Q j (k) are the probabilities of distribution P and Q in the i-th or j-th category respectively, Z represents the number of fault categories, k is a constant; ε is a constant.

7. The bearing open set fault diagnosis method based on domain adversarial neural network and decision boundary adaptation according to claim 6 is characterized in that: The total loss function is expressed as: Among them, μ is the weight coefficient; Where n represents the number of source domain samples, n' represents the number of target domain samples, N is the total number of samples, and λ is the weight coefficient; G f (·;θ f ) indicates that the parameter is θ f The output result of the feature extraction module, G y (·;θ y ) indicates that the parameter is θ y The output of the label predictor, x v ,y v Represent the input data and labels respectively, L y Represents the classification loss as the negative log probability of the correct label; G d (·;θ d ) indicates that the parameter is θ d The output of the domain classifier, d v Represents the binary variable of the vth input data, d v =0 or 1; Represents the operation of a gradient reversal layer.

Images