A gearbox fault diagnosis method based on graph convolution domain adaptive network

Abstract

The present invention discloses a gearbox fault diagnosis method based on a graph convolution domain adaptive network. The method comprises: constructing an unsupervised adaptive module to extract and integrate fault features from gearbox fault signals under target domain working conditions; constructing a dense connection module to perform feature propagation and dimensionality reduction based on the dense connection concept; constructing a graph convolution module to construct and model an instance graph based on the features output by the dense connection module; constructing a softmax classifier to classify the output of the graph convolution module to obtain a fault classification result for the signal; constructing a fault diagnosis model and training and optimizing the model; and inputting test set data into the trained fault diagnosis model to diagnose gearbox signal faults. The present invention can promptly detect gearbox faults and reliably identify fault signals, effectively predicting gearbox fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of gearbox fault detection, and in particular relates to a gearbox fault diagnosis method based on a graph convolution domain adaptive network. Background Art

[0002] Gearboxes are one of the most fundamental and crucial components of rotating machinery used in industry and helicopters. Their function is to transmit torque and power from one shaft to another. Typical applications include helicopters, automobiles, metal-cutting machine tools, and power systems. Failure of any gear in these machines, especially helicopters, during operation can have serious consequences. Therefore, accurate gearbox fault diagnosis is crucial to preventing mechanical system failures, which can lead to damage or even complete system downtime.

[0003] Gears can fail in a variety of ways, including wear, misalignment, cracks, fatigue, and chipping. Any of these failures, when severe, can lead to catastrophic failures in machines like helicopters. Gear wear is a slow process. When wear reaches a critical level, the gears can break, potentially causing machine failure.

[0004] Gear fault diagnosis has been extensively studied and investigated. In most factory systems, gear wear cannot be directly observed and can only be assessed using other measured data from the machine. Therefore, real-time monitoring and diagnosis of gearbox faults are crucial for the safe and reliable operation of machinery.

[0005] The inconsistent distribution of gearbox data under variable operating conditions means that networks trained on source operating condition data cannot be applied to target operating conditions. This is because the data distribution between the source domain used for network training and the gearbox fault data in the target domain used for actual diagnosis differs. Acquiring labeled data for various gearbox faults under different operating conditions is extremely time-consuming and energy-intensive. Furthermore, gearbox fault data under target operating conditions is often unlabeled. Therefore, there is no guarantee that a network that works well on labeled source operating condition data will have the same effect on unlabeled target operating condition data. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a gearbox fault diagnosis method based on a graph convolutional domain adaptive network. First, an attention mechanism is embedded to help the model better focus on the part of the input data related to the current task, thereby improving the performance and generalization ability of the model. Then, by adding a dense connection module, a dense connection is established between the front layer and the back layer, so that the network can pay attention to channel information and reuse the features of the front layer. Finally, by adding a graph convolution module, the model can retain the characteristics of the node itself and make full use of the global information of the entire graph, making the model more comprehensive in capturing the structural features of the graph. Therefore, compared with other algorithms that rely only on a single feature or a single local feature, the present invention can more effectively predict gearbox fault diagnosis, diagnose gearbox signal faults and reliably identify fault signals. It also has good stability, strong robustness to noise, and certain generalization ability. It is of great significance to ensure the safe and stable operation of the system and improve the technical level of motor gearboxes with higher accuracy and better adaptability.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0008] A gearbox fault diagnosis method based on a graph convolutional domain adaptive network, comprising:

[0009] 1) Construct an unsupervised adaptive module to extract and integrate fault features of gearbox fault signals under target domain working conditions;

[0010] 2) Construct a dense connection module to perform feature propagation and dimensionality reduction based on the dense connection idea;

[0011] 3) Construct a graph convolution module to construct and model an instance graph based on the features output by the dense connection module;

[0012] 4) Construct a softmax classifier to classify the output of the graph convolution module and obtain the fault classification result of the signal;

[0013] 5) Construct a fault diagnosis model using an unsupervised adaptive module, a dense connection module, a graph convolution module, and a softmax classifier and train and optimize the model;

[0014] 6) Input the test set data into the trained fault diagnosis model to diagnose the gearbox signal fault.

[0015] To optimize the above technical solutions, specific measures taken also include:

[0016] The unsupervised adaptive module constructed in 1) above consists of two convolution blocks with inconsistent kernel sizes and an attention mechanism using residual connections. The two convolution blocks with inconsistent kernel sizes are used to extract fault features from fault signals, and the attention mechanism is used to weight each feature to achieve fault feature integration.

[0017] The unsupervised adaptive module constructed in 1) above also includes batch normalization, linear correction and maximum pooling layers, which are used to perform batch normalization, linear correction and maximum pooling on the features integrated by the attention mechanism.

[0018] The dense connection module constructed in 2) above includes two dense connection blocks and a convolutional transformation layer. The convolutional transformation layer is inserted between the two dense connection blocks. The convolutional transformation layer includes a 1x1 convolution layer and an average pooling layer.

[0019] The densely connected block includes 1x1 and 3x3 convolutional layers, which are used to reduce the number of channels of the feature map and extract features respectively;

[0020] Each layer of the densely connected block is directly connected to all previous layers to reuse the features of the previous layers;

[0021] The convolution transformation layer is used for feature dimensionality reduction of two densely connected blocks.

[0022] The output of the above dense connection block is

[0023] X l =H L ([X0,X1,...,X l-1 ]) (1)

[0024] Among them, X l represents the output of the lth layer of the densely connected module, H L Represents the convolution and pooling nonlinear transformation function, X0, X1, ..., X l-1 represents the original input and the output of the previous l-1 layers concatenated in the channel direction.

[0025] The graph convolution module constructed in 3) above includes a graph generation layer and a two-layer residual connection graph convolution network. The instance graph is constructed through the graph generation layer, and the instance graph is modeled through the two-layer residual connection graph convolution network.

[0026] The graph generation layer above learns the data structure from the features output by the dense connection module, obtains the adjacency matrix and constructs the instance graph from the matrix;

[0027] The process of constructing an instance graph in the graph generation layer is as follows:

[0028] Each feature vector output by the dense connection module is regarded as a node, and its value is regarded as the node feature to form a feature matrix;

[0029] The feature matrix is input into the multi-layer perceptron MLP, and the matrix between the features output by the MLP and its transpose is multiplied to obtain the adjacency matrix;

[0030] According to the top-k sorting mechanism, the first k nearest neighbors of each node in the adjacency matrix are selected to obtain the sample structure characteristics;

[0031] Mining the relationship between sample structural features and constructing instance graphs;

[0032] The graph convolutional network models the structural information propagated along the weighted edges in the instance graph to embed data structure information into node features.

[0033] The softmax classifier constructed in 4) above is:

[0034]

[0035] Among them, Z i represents the score or feature of the i-th category, and K represents the total number of categories.

[0036] In the above 5) training optimization process, the overall loss function used is:

[0037]

[0038] in, Belong to A collection of

[0039] γ, κ are trade-off parameters;

[0040] Classification loss for:

[0041]

[0042] in is the i-th sample in the source domain, is the i-th label of the source domain, D s represents the source domain;

[0043] represents the prediction result of the label classifier, and E(·) represents the mathematical expectation.

[0044] Domain alignment lost for:

[0045]

[0046] in represents the mathematical expectation of the i-th sample in the source domain, Denotes the mathematical expectation of the jth sample in the target domain, D t represents the target domain;

[0047] and They represent the extracted features of the jth target sample and the ith source sample respectively. The value of D(·) takes 0 or 1, which can be used to distinguish which domain the sample belongs to;

[0048] Structural alignment loss for:

[0049]

[0050] in represents the mathematical expectation of the i-th sample in the source domain set, represents the mathematical expectation of the jth sample in the target domain set, Ω indicates that the distance is measured by embedding the extracted features into the RKHS;

[0051] Represents a nonlinear mapping function.

[0052] In the above overall loss function, θ F ,θ C and θ D Denote the parameters of the feature extractor, label classifier, and domain discriminator respectively. During the model training and optimization process, each parameter is updated through the back-propagation algorithm, which is expressed as

[0053]

[0054]

[0055]

[0056] Among them, L Total , L C , L DA They are overall loss, classification loss, and domain alignment loss;

[0057] η represents the learning rate.

[0058] The present invention has the following beneficial effects:

[0059] This paper constructs an unsupervised adaptive network module that automatically adjusts the network structure and parameters of the source and target domains through a mapping method. Furthermore, this module incorporates an attention mechanism to adapt to different data distributions and task requirements, ensuring that more important channel data has a greater impact on the results. This allows for a more accurate understanding and judgment of classification tasks.

[0060] The present invention constructs a densely connected structure module, which consists of two densely connected blocks and a convolutional transformation layer. The dense blocks can reuse the features of the previous layer to better utilize the feature information, and the convolutional transformation layer can achieve feature dimensionality reduction of the two dense blocks, alleviating the gradient vanishing problem, while also improving the expressiveness and accuracy of the model.

[0061] The present invention constructs a graph convolution module, which constructs an instance graph through a graph generation layer; models the instance graph through a two-layer residual connection graph convolutional network, and utilizes a parameter sharing mechanism so that the features of each node can be updated by the features of adjacent nodes, which is more comprehensive in capturing the graph structure features.

[0062] The present invention has faster computing speed and better performance, can adaptively extract feature information from source domain signals, and accurately and efficiently diagnose gearbox faults, providing an effective new approach for intelligent fault diagnosis of motors. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is the overall principle diagram of the method of the present invention;

[0064] Figure 2 It is the unsupervised adaptive module of the present invention;

[0065] Figure 3 It is a dense connection module of the present invention;

[0066] Figure 4 This is the graph generation layer of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0068] Although the steps in the present invention are arranged with numbers, they are not intended to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0069] The present invention provides a gearbox fault diagnosis method based on graph convolution domain adaptive network, such as Figure 1 Shown, including:

[0070] 1) Construct an unsupervised adaptive module to extract and integrate fault features of gearbox fault signals under target domain working conditions;

[0071] This part uses an unsupervised adaptive network structure to process data signals. First, the convolution operation in the structure is used to identify and extract the important features of the fault signal. The local feature information is extracted by the convolution operation with a convolution kernel size of 3. At the same time, the different feature information is integrated through the embedded attention mechanism, thereby enhancing the expressive power of the feature map. The purpose of extracting detailed features is achieved by the unsupervised adaptive convolutional neural network structure. At the same time, each feature is weighted by the embedded attention mechanism, so that the neural network pays more attention to those features that are more important for the classification task. The present invention specifically constructs an unsupervised adaptive network module, which automatically adjusts the network structure and parameters of the source domain and the target domain through the mapping method. The module combines the attention mechanism to adapt to different data distributions and task requirements, so that the more important channel data has a greater impact on the results. Thus, there is a more accurate understanding and judgment of the classification task. The embedded attention mechanism can help the model better focus on the part of the input data related to the current task, thereby improving the performance and generalization ability of the model. In the embodiment, the input data is first subjected to feature extraction through an unsupervised convolution operation, and then the output is processed through batch normalization, rectified linear unit (ReLu) and maximum pooling operations. The specific introduction is as follows:

[0072] In view of the characteristics that the gearbox fault data under the target domain working conditions in the gearbox fault diagnosis system is often unlabeled and obtaining various gearbox fault label data is very time-consuming and energy-consuming, the present invention designs an unsupervised domain adaptation module such as Figure 2 shown.

[0073] The module mainly consists of two convolution blocks with inconsistent kernel sizes and an attention mechanism using residual connections.

[0074] First, by using the adversarial training process and then performing convolution operations, the structure can automatically adjust the network structure and parameters according to the characteristics of the input data, adapt to different data distributions and task requirements, better generalize to unseen data, reduce the risk of overfitting, and improve the robustness and generalization ability of the model.

[0075] Secondly, in the design of the model, the residual connection method is adopted, which can effectively train the network and reduce the problems of gradient disappearance and gradient explosion in the training process of deep neural networks;

[0076] The importance of feature extraction was fully considered in the design, and technologies such as attention mechanism and convolution kernels of different scales were adopted to effectively improve the feature expression ability of the model.

[0077] Finally, the use of ReLU activation function in this module can enhance the nonlinear mapping ability of the model, avoid the problem of gradient disappearance, and help alleviate the training difficulties caused by gradient disappearance. During the training process, it can maintain the stability of the gradient and prevent gradient explosion, which helps to accelerate convergence and thus improve the efficiency and accuracy of the model.

[0078] 2) Construct a dense connection module to perform feature propagation and dimensionality reduction based on the dense connection idea;

[0079] This part uses a dense connection module, in which each layer is directly connected to all previous layers. This allows information to be transmitted more fully, and the features are input into the graph generation layer proposed by the graph convolution module to construct an instance graph by mining the relationship between the sample structure features; each layer of the dense block module receives the feature maps of all previous layers as input, and passes its own feature maps to all subsequent layers; the present invention specifically constructs a dense connection structure module, which consists of two dense connection blocks and a convolution conversion layer. The dense block can realize the reuse of the features of the previous layer to better utilize the feature information, and the convolution conversion layer can realize the dimensionality reduction of the features of the two dense blocks, alleviate the gradient disappearance problem, and also improve the expression ability and accuracy of the model. The dense connection module establishes dense connections between the front and back layers, so that the network not only pays attention to the channel information, but also can realize the reuse of the features of the front layer. Then, the dense connection block (dense block) is used to realize the reuse of the features of the front layer, and the features are input into the proposed graph generation layer through the graph convolution network to construct the instance graph by mining the relationship between the sample structure features.

[0080] The details are as follows:

[0081] In order to make the more important channel data have a greater impact on the results, the present invention designs a dense connection (Dense Connection Block) module, which consists of two dense connection blocks (dense blocks) and a convolution transformation layer with a kernel size of 1. In order to represent the features, a convolution feature module is used, which includes kernel=1 ordinary convolution and kernel=3 ordinary convolution. The convolution feature module is as follows Figure 3 shown.

[0082] In order to reduce the computational cost, a 1x1 convolution layer is first used inside each dense block to reduce the number of channels of the feature map, and then a 3x3 convolution layer is used for feature extraction. A convolution transition layer is inserted between dense blocks to control the dimension and size of the feature map. The convolution transition layer usually contains a 1x1 convolution layer and an average pooling layer to reduce the size of the feature map. By introducing the idea of dense connection, a dense connection is established between the front layer and the back layer, so that the network can pay attention to the channel information and reuse the features of the front layer. The output of the dense block with dense connection idea can be expressed as

[0083] X l =H L ([X0,X1,...,X l-1 ]) (1)

[0084] Among them, X l represents the output of layer l, H L Represents a nonlinear transformation function with a series of operations, including convolution and pooling, and X0, X1, ..., X l-1 Represents the original input and the output of the previous l-1 layers connected in the channel direction. Compared to traditional neural network structures, dense connections enable each layer to directly access the feature information of all previous layers, effectively promoting the transfer and reuse of features. This helps alleviate the vanishing gradient problem, accelerates the training process, and improves the model's generalization ability. It also more efficiently utilizes parameters, allowing features to be transferred more quickly and directly to subsequent layers, helping the model learn richer and more complex feature representations, improving the model's ability to understand the data and reducing the number of parameters required to learn.

[0085] 3) Construct a graph convolution module to construct and model an instance graph based on the features output by the dense connection module;

[0086] This part uses graph convolution (GCN) to model the instance graph constructed by the graph generation layer by mining the sample structure features; the graph convolution module uses the instance graph constructed by the graph generation layer of the feature input to model the instance graph; the present invention specifically constructs a graph convolution module, which constructs the instance graph through a graph generation layer; the instance graph is modeled through a two-layer residual connection graph convolution network, and a parameter sharing mechanism is used so that the features of each node can be updated through the features of the adjacent nodes, which is more comprehensive in capturing the graph structure features. The instance graph is modeled through the graph convolution network; the graph convolution module enables the model to retain the features of the node itself and make full use of the global information of the entire graph, making the model more comprehensive in capturing the graph structure features. The specific introduction is as follows:

[0087] Based on the problem that most methods ignore the modeling of data structure, which makes the information contained in the features extracted by deep networks incomplete, this paper designs a graph convolutional network (GCN) module. This module proposes a graph generation layer (GGL) to learn the data structure from the extracted CNN features. GGL obtains the adjacency matrix A and constructs an instance graph from the mini-batch input matrix. The process is as follows: Figure 4 shown.

[0088] First, the extracted feature matrix is fed into a multilayer perceptron (MLP). Then, the adjacency matrix is obtained by matrix multiplication between the MLP features and their transpose. Finally, the top k nearest neighbors of each node are selected according to a top-k ranking mechanism. Each feature vector is considered a node and its value as a node feature. A graph is automatically generated using the proposed GGL, and then an instance graph is constructed by mining the relationships between the sample structural features.

[0089] Finally, the obtained graph is fed into a GCN to embed data structure information into node features. A graph convolutional network (GCN) is applied to model the instance graph, which can model the structural information propagated along weighted edges in the graph. Finally, the obtained node features are used for fault classification and domain adversarial training. By employing graph convolution, the network can better leverage information from the graph structure, thereby better capturing the associations and dependencies between nodes. Furthermore, by utilizing a parameter sharing mechanism, the features of each node can be updated using the features of neighboring nodes, significantly reducing the model's parameter size and improving its generalization capabilities. It can also leverage global information, preserving the features of the node itself while fully leveraging the global information of the entire graph. This makes the model more comprehensive in capturing graph structural features, offering greater applicability and flexibility in practical applications, thereby improving its efficiency and accuracy.

[0090] 4) Construct a softmax classifier to classify the output of the graph convolution module and obtain the fault classification result of the signal;

[0091] This section uses a softmax-based classifier to calculate the probability of each fault type and classify the signal. The softmax classifier is used to classify signal faults. An overall objective function (including classification loss, domain alignment loss, and structural alignment loss) is used to estimate the structural difference between the predicted results and the actual values. Finally, the output features are passed through a fully connected layer to achieve classification for gearbox fault diagnosis. The details are as follows:

[0092] For the detection and identification of transient signal faults in power systems, this paper constructs a softmax classifier based on the softmax function. The softmax classifier first divides the input samples into multiple mutually exclusive categories and uses a fully connected layer to map the input features into a vector space. Then, these features are normalized by applying the softmax function to obtain the probability of each category for classification. The softmax function is defined as follows

[0093]

[0094] Among them, Z i represents the score or feature of the i-th category, and K represents the total number of categories. The softmax function converts the input score into a probability so that the sum of the probabilities of all categories is 1.

[0095] 5) Construct a fault diagnosis model using an unsupervised adaptive module, a dense connection module, a graph convolution module, and a softmax classifier and train and optimize the model;

[0096] This part uses the overall objective function (overall loss function) to measure the gap between the model's prediction results and the actual values, and the optimization algorithm adjusts the model parameters to minimize this gap; the details are as follows:

[0097] The proposed DAGCN consists of three parts, namely, feature extractor (F), domain discriminator (D), and label classifier (C). In order to learn transferable features and model the three types of important information mentioned above, the overall objective function contains three parts, including classification loss, domain alignment loss, and structural alignment loss.

[0098] Classification loss:

[0099] In order to ensure that the label classifier has promising prediction results, the classification loss between the true label and the predicted label is estimated by the cross entropy loss, which can be defined as:

[0100]

[0101] in is the i-th sample in the source domain, is the i-th label of the source domain, D s represents the source domain;

[0102] represents the prediction result of the label classifier, and E(·) represents the mathematical expectation.

[0103] Domain alignment lost:

[0104] Due to the domain covariance shift problem, the label classifier trained only with source domain data cannot work well with target domain data. To solve this problem, a domain discriminator (D) is used to judge whether the extracted features are from the target domain or the source domain, and the feature extractor is trained to deceive the domain discriminator. When the two-player minimax game reaches equilibrium, domain invariant features can be captured. Here, binary cross entropy loss is used as the domain alignment loss, which is expressed as

[0105]

[0106] in represents the mathematical expectation of the i-th sample in the source domain, Denotes the mathematical expectation of the jth sample in the target domain, D t represents the target domain;

[0107] and They represent the extracted features of the jth target sample and the ith source sample respectively. The value of D(·) takes 0 or 1, which can be used to distinguish which domain the sample belongs to.

[0108] Structural alignment loss:

[0109] In order to align the feature structures of the source and target domains, the MMD metric is used as the structural difference alignment loss, which is expressed as follows:

[0110]

[0111] in represents a nonlinear mapping function, which means that the distance is measured by embedding the extracted features into the RKHS.

[0112] Overall objective function:

[0113] Combining the three defined loss functions, the overall objective function for implementing UDA can be written as

[0114]

[0115] in Belong to A collection of

[0116] γ,κ are trade-off parameters.

[0117] The overall objective function is shown in (6), where θ F ,θ C and θ D Denote the parameters of the feature extractor, label classifier, and domain discriminator respectively. During the model training process, the parameters of each part of DAGCN can be updated through the back propagation (BP) algorithm, which is expressed as

[0118]

[0119]

[0120]

[0121] Where η represents the learning rate.

[0122] 6) Input the test set data into the trained fault diagnosis model to diagnose the gearbox signal fault.

[0123] When the model training is completed, the test set data is input into the model, and the gearbox signal fault diagnosis is completed based on the output results of the prediction classification.

[0124] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0125] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A gearbox fault diagnosis method based on graph convolution domain adaptive network, characterized in that: include: 1) Construct an unsupervised adaptive module to extract and integrate fault features of gearbox fault signals under target domain working conditions; 2) Construct a dense connection module to perform feature propagation and dimensionality reduction based on the idea of dense connection; 3) Construct a graph convolution module to construct and model an instance graph based on the features output by the dense connection module; 4) Construct a softmax classifier to classify the output of the graph convolution module and obtain the fault classification result of the signal; 5) Construct a fault diagnosis model using an unsupervised adaptive module, a dense connection module, a graph convolution module, and a softmax classifier, and train and optimize the model; 6) Input the test set data into the trained fault diagnosis model to diagnose the gearbox signal fault; The unsupervised adaptive module constructed in 1) is composed of two convolution blocks with different convolution kernel sizes and an attention mechanism using a residual connection. The two convolution blocks with different convolution kernel sizes are used to extract fault features from fault signals, and the attention mechanism is used to weight each feature to achieve fault feature integration. The unsupervised adaptive module constructed in 1) further includes batch normalization, linear correction, and maximum pooling layers, which are used to perform batch normalization, linear correction, and maximum pooling on the features integrated by the attention mechanism; The dense connection module constructed in 2) includes two dense connection blocks and a convolutional transformation layer. The convolutional transformation layer is inserted between the two dense connection blocks. The convolutional transformation layer includes a 1x1 convolutional layer and an average pooling layer. The densely connected block includes 1x1 and 3x3 convolutional layers, which are used to reduce the number of channels of the feature map and extract features respectively; Each layer of the densely connected block is directly connected to all previous layers to reuse the features of the previous layers; The convolution transformation layer is used for feature dimensionality reduction of two densely connected blocks; The output of the dense connection block is: （1） in, Represents the densely connected module l The output of the layer, represents the convolution and pooling nonlinear transformation function, Represents the original input and the preceding The output of the layer; The graph convolution module constructed in 3) includes a graph generation layer and a two-layer residual connection graph convolution network, which constructs an instance graph through the graph generation layer and models the instance graph through the two-layer residual connection graph convolution network; The graph generation layer learns the data structure from the features output by the dense connection module, obtains the adjacency matrix and constructs an instance graph from the matrix; The process of constructing an instance graph in the graph generation layer is as follows: Each feature vector output by the dense connection module is regarded as a node, and its value is regarded as the node feature to form a feature matrix; The feature matrix is input into the multi-layer perceptron MLP, and the matrix between the features output by the MLP and its transpose is multiplied to obtain the adjacency matrix; According to the top-k sorting mechanism, the first k nearest neighbors of each node in the adjacency matrix are selected to obtain the sample structure characteristics; Mining the relationship between sample structural features and constructing instance graphs; The graph convolutional network models the structural information propagated along the weighted edges in the instance graph to embed data structure information into node features.

2. The gearbox fault diagnosis method based on graph convolutional domain adaptive network according to claim 1, characterized in that: The softmax classifier constructed in 4) is: （2） Among them, Z i represents the score or feature of the i-th category, and K represents the total number of categories.

3. The gearbox fault diagnosis method based on graph convolutional domain adaptive network according to claim 1, characterized in that: In the above 5) training optimization process, the overall loss function used is: （6） in, Belong to 、 、 A collection of , is a trade-off parameter; Classification loss for: （3） in is the i-th sample in the source domain, is the i-th label of the source domain, represents the source domain; represents the prediction result of the label classifier, and E(·) represents the mathematical expectation; Domain alignment lost for: （4） in The source domain i The mathematical expectation of samples, The target domain j The mathematical expectation of samples, represents the target domain; and They represent the extracted features of the jth target sample and the i-th source sample respectively. The value of D(·) takes 0 or 1, which can be used to distinguish which domain the sample belongs to; Structural alignment loss for: （5） in Indicates the first i The mathematical expectation of samples, Indicates the first j The mathematical expectation of samples, The representation distance is measured by embedding the extracted features into RKHS; Represents a nonlinear mapping function.

Images