GCN fluid dynamic pressure mechanical seal diagnosis method based on adjacent attention mechanism

Abstract

The invention belongs to the technical field of mechanical fault diagnosis, and relates to a GCN fluid dynamic pressure mechanical seal diagnosis method based on an adjacent attention mechanism. Frequency domain features of the original time domain signals are extracted through fast Fourier transform, and an initial graph is generated through a k-NN algorithm; then extracting signal samples representing nodes in the initial graph by using an adjacency attention mechanism, re-evaluating a neighbor relation of the signal samples, generating a new adjacency matrix and a new adjacency weight, and obtaining an optimized graph after iteration; inputting the optimized graph into a GCN module for feature extraction; and finally, directly connecting the feature map which is output by the GCN module and can be used for classification to a full connection layer output result to obtain a fault type. According to the method, the adjacent relation of the nodes is evaluated again through an adjacent attention mechanism, a new adjacent matrix and a new adjacent weight are generated, and an optimized iterative graph is obtained, so that local features in a sample are further learned from a feature graph of learned topological features, and the problem that the local features are ignored by a traditional GCN is solved.

Description

Technical Field

[0001] The present application belongs to the technical field of mechanical fault diagnosis, and more specifically, to a GCN fluid dynamic pressure mechanical seal diagnosis method based on an adjacent attention mechanism. Background Art

[0002] The main coolant pump of a million-kilowatt nuclear power plant reactor (referred to as the nuclear main pump) is a key equipment in the coolant system (primary loop system) of a pressurized water reactor nuclear power plant. It is used to transport the coolant circulation in the closed system composed of the reactor main pipeline and the steam generator, and transfer the heat generated by the reactor to the evaporator. Due to its particularity and importance, the shaft seal in the nuclear main pump structure has the characteristics of high technical difficulty, high cost, frequent preventive maintenance, and a large inventory of old machine seal spare parts. The realization of intelligent fault diagnosis is of great significance to improving the operating safety of nuclear power plants and the reliability of nuclear power equipment. However, the current domestic independent research on the development, manufacturing and maintenance of fluid dynamic pressure mechanical seals is relatively preliminary, and there is no effective practical experience in sealing ring fault monitoring, maintenance and reuse in China. With the development of sensors and monitoring methods, as well as the advancement of computer computing power and database technology, in order to cope with the increasingly high quantity and high dimensional characteristics of industrial equipment monitoring data, data-driven fault diagnosis methods have gradually attracted people's attention. Data-driven methods emphasize the direct use of a large amount of data generated by the system, and build diagnostic models through machine learning (ML) and even deep learning (DL) algorithms to automatically learn the system's working status and fault characteristics from the equipment's monitoring signals and give diagnostic results. This type of method usually does not require an in-depth understanding of the system's working principles, but learns the system's behavior patterns from data. Data-driven methods can avoid analyzing and exploring mechanisms for complex, nonlinear, and difficult-to-model systems, and have good adaptability to faults with strong uncertainty and randomness, and can discover potential associations and patterns that are difficult to identify manually. For large-scale complex systems, or for systems where the fault mechanism is not clear, data-driven methods can adapt to different working conditions and fault modes more flexibly, especially when dealing with large amounts of unstructured data or complex patterns. In addition, since such methods generally use computers to establish feature learning and recognition models to directly use signals for fault diagnosis, data-driven fault diagnosis methods are easier to integrate into automatic systems. Traditional fault diagnosis methods based on deep neural networks can effectively capture the hidden features of conventional data (such as images and time series), but most of them ignore the interdependence between data or various physical measurements of multiple sensors. In addition, in the real world, faults may change the operating state of the mechanical system and cause the dependencies between the samples obtained by the measurement signal to change. Based on this, many researchers have begun to try to describe fault data with graph-level data in order to explore potential features in the data space or jointly process multi-sensor data. The practice of converting fault signal data into graph data is essentially to convert data originally belonging to Euclidean space into data in non-Euclidean space, which makes operations that are easy to implement in Euclidean space (such as convolution, etc.) difficult to implement on data in non-Euclidean space, and it is impossible to use the above-mentioned various data-driven classification models for such data. GCN can learn the topological structure between samples and aggregate highly correlated features, and simulate higher-level features in the sample space by using graph-level expressions of data. However, the application of GCN in the fault diagnosis of fluid dynamic mechanical seals still has limitations: first, the original monitoring signals in fluid dynamic mechanical seals themselves contain inseparable noise and instrument-introduced errors. Therefore, the graph directly constructed from these signal samples contains a large number of errors and redundant edges. This has an adverse effect on the stability and accuracy of the GCN model in feature learning. Secondly, the operating state of fluid dynamic mechanical seals will change dynamically over time, so the discriminant features of fault signals are often reflected not only within a single node, but also in the interdependence between multiple nodes. However, the traditional GCN method can only learn feature representations from a single granularity and cannot effectively simulate the correlation between nodes. In addition, empirical evidence from signal processing-based fault diagnosis methods shows that there are local features in signal samples that are directly related to the fault category, and learning and extracting these features plays a vital role in improving the classification performance of the model. However, most traditional GCN models focus on learning and extracting global features in samples, ignoring the importance of local features. Summary of the invention

[0003] The main coolant pump of a million-kilowatt nuclear power plant reactor (referred to as the nuclear main pump) is a key equipment in the coolant system (primary loop system) of a pressurized water reactor nuclear power plant. It is used to transport the coolant circulation in the closed system composed of the reactor main pipeline and the steam generator, and transfer the heat generated by the reactor to the evaporator. Due to its particularity and importance, the shaft seal in the nuclear main pump structure has the characteristics of high technical difficulty, high cost, frequent preventive maintenance, and a large inventory of old machine seal spare parts. The realization of intelligent fault diagnosis is of great significance to improving the operating safety of nuclear power plants and the reliability of nuclear power equipment. However, the current domestic independent research on the development, manufacturing and maintenance of fluid dynamic pressure mechanical seals is relatively preliminary, and there is no effective practical experience in sealing ring fault monitoring, maintenance and reuse in China.

[0004] With the development of sensors and monitoring methods, as well as the advancement of computer computing power and database technology, in order to cope with the increasingly high quantity and high dimensional characteristics of industrial equipment monitoring data, data-driven fault diagnosis methods have gradually attracted people's attention. Data-driven methods emphasize the direct use of a large amount of data generated by the system, and build diagnostic models through machine learning (ML) and even deep learning (DL) algorithms to automatically learn the system's working status and fault characteristics from the equipment's monitoring signals and give diagnostic results. This type of method usually does not require an in-depth understanding of the system's working principles, but learns the system's behavior patterns from data. Data-driven methods can avoid analyzing and exploring the mechanisms of complex, nonlinear, and difficult-to-model systems, and have good adaptability to faults with strong uncertainty and randomness, and can discover potential associations and patterns that are difficult to identify manually. For large-scale complex systems, or for systems where the fault mechanism is not clear, data-driven methods can adapt to different working conditions and fault modes more flexibly, especially when dealing with large amounts of unstructured data or complex patterns. In addition, since such methods generally use computers to establish feature learning and recognition models to directly use signals for fault diagnosis, data-driven fault diagnosis methods are easier to integrate into automatic systems.

[0005] Traditional fault diagnosis methods based on deep neural networks can effectively capture the hidden features of conventional data (such as images and time series), but most of them ignore the interdependence between data or various physical measurements of multiple sensors. In addition, in the real world, faults may change the operating state of the mechanical system and cause the dependencies between the samples obtained by the measurement signal to change. Based on this, many researchers have begun to try to describe fault data with graph-level data in order to explore potential features in the data space or jointly process multi-sensor data. The practice of converting fault signal data into graph data is essentially to convert data originally belonging to Euclidean space into data in non-Euclidean space, which makes operations that are easy to implement in Euclidean space (such as convolution, etc.) difficult to implement on data in non-Euclidean space, and it is impossible to use the above-mentioned various data-driven classification models for such data. GCN can learn the topological structure between samples and aggregate highly correlated features, and simulate higher-level features in the sample space by using graph-level expressions of data. However, the application of GCN in the fault diagnosis of fluid dynamic mechanical seals still has limitations: first, the original monitoring signals in fluid dynamic mechanical seals themselves contain inseparable noise and instrument-introduced errors. Therefore, the graph directly constructed from these signal samples contains a large number of errors and redundant edges. This has an adverse effect on the stability and accuracy of the GCN model in feature learning. Secondly, the operating state of fluid dynamic mechanical seals will change dynamically over time, so the discriminant features of fault signals are often reflected not only within a single node, but also in the interdependence between multiple nodes. However, the traditional GCN method can only learn feature representations from a single granularity and cannot effectively simulate the correlation between nodes. In addition, empirical evidence from signal processing-based fault diagnosis methods shows that there are local features in signal samples that are directly related to the fault category, and learning and extracting these features plays a vital role in improving the classification performance of the model. However, most traditional GCN models focus on learning and extracting global features in samples, ignoring the importance of local features.

[0006] The present invention provides a GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism, which is intended to solve the technical problem that the application of GCN in fluid dynamic pressure mechanical seal fault diagnosis still has limitations.

[0007] The GCN fluid dynamic pressure mechanical seal diagnosis method based on the neighbor attention mechanism includes the following steps: Step 1: Divide the end face film thickness signals collected from multiple displacement sensors into several sub-samples, assign corresponding labels to each sub-sample, and then perform fast Fourier transform on each sub-sample to display the frequency domain characteristics of each sub-sample; Step 2: Consider each subsample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighboring nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weight to obtain a sparse initial graph; Step 3: Adopt the adjacency attention mechanism and use the sample matrix in the sparse initial graph to extract the signal samples representing the nodes in the sample matrix and re-evaluate their neighboring relationships, generate a new adjacency matrix and adjacency weights, and obtain the optimized iterative graph; Step 4: Input the matrix of the iteration graph into the GCN module for feature extraction to obtain a feature graph for classification; Step 5: Connect the feature map for classification output by the GCN module to the fully connected layer, use Softmax to distinguish categories, and obtain the fault type.

[0008] The present invention re-evaluates the neighboring relationship of nodes through the adjacency attention mechanism, generates a new adjacency matrix and adjacency weights, and obtains an optimized iterative graph, so that the local features in the sample can be further learned from the feature graph of the learned topological features, thus solving the problem of traditional GCN ignoring local features.

[0009] Preferably, the neighboring nodes of the node in step 2 are represented as follows: ; Where: Indicates the return node In the collection The first k nearest neighbor nodes in ; The parameter representing the number of neighboring nodes of the corresponding node in the K-NN algorithm; represents a subset consisting of m nodes; Representation Node All adjacent nodes of .

[0010] Preferably, the step of simplifying the adjacency relationship in the initial graph based on the adjacency matrix and the adjacency weight is as follows: The adjacency matrix defines nodes that are related to each other, and does not mark the relationship between two nodes that are essentially unrelated; The adjacency weight indicates the closeness of the adjacency relationship; The Gaussian kernel function is used to evaluate the weight of the relationship between nodes: ; Where: Representation Node and The adjacency weight between them; Represents the width of the Gaussian kernel, which is used to control the length scale parameter of the input space. Its value is related to the size of the input image. represents the base of natural logarithms; After obtaining the adjacency weights of all sub-samples, the top k samples with the highest similarity in each sub-sample are retained, and the other sub-samples are set to zero to obtain a sparse initial graph.

[0011] Preferably, step 3 comprises the following steps: The similarity vector of any two node samples x and y in the sparse initial graph is calculated through two convolutional layers: ; Where: and Respectively represent any two node samples in the sparse initial graph; They are two different convolutional layers, Indicates that the convolution sum size is 1; Then use the LeackReLU function to activate and obtain the similarity vector between the two nodes Then calculate the adjacency weight between nodes x and y : ; Where: Represents the set of neighboring nodes of node x; Respectively represent any three node samples in the sparse initial graph; After obtaining the adjacency weights of all nodes and other nodes, the K-NN algorithm is used to construct the graph and the nodes with the highest adjacency weight are selected. Nodes are regarded as nodes with neighboring relationships, and the adjacent attention score matrix is ​​obtained. Finally, the iterative adjacency matrix is ​​superimposed with the adjacency relationship matrix in the sparse initial graph to achieve adjacency attention and obtain the optimized iterative graph: ; Where: Represents the adjacency matrix of the sparse initial graph.

[0012] Preferably, the GCN module includes two layers of high-order graph convolutional layers; In the high-order graph convolutional layer, a single hidden layer is defined as: ; Where: Represents the features of the hidden nodes that have been learned; represents the dimension of the subgraph in the lth iteration; and Respectively indicated by The subgraph constructed by each node and its adjacent subgraphs; Representing a subgraph The set of all subgraphs in the local neighborhood of ; represents the learned features of the previous hidden node, Represents the weight matrix trained in the 1st and 2nd high-order graph convolutional layers; In the high-order graph convolution layer, an edge relationship graph pooling method EdgePool is used; the graph pooling method EdgePool adaptively merges nodes with edges between each other during the iteration process to establish new nodes, and the node features of the new nodes are calculated by the following function: ; Where: and Represents the node features of node a and node b; and yes Edgepool Two learned parameters in the layer; After establishing the new node, the features of the new node of each sample are aggregated through global average pooling to obtain a feature map for classification.

[0013] Preferably, the expression for class differentiation using Softmax is as follows: ; Where: For input The new feature representation of Indicates the number of target categories for the classification task, represents the number of hidden layers, is the feature representation, is a diagonal and satisfies The degree matrix of represents the weight matrix; For the Perform iterative optimization, where the loss function for iterative optimization is as follows: ; Where: represents the label set of the test set, represents the true value of the test sample label, Indicates the Among the test samples The beneficial effects of the present invention include: The present invention uses high-order graph convolution layers to extract global features of data, and has stronger hierarchical representation learning capabilities compared to other GCNs. The residual connection convolution module introduced by the present invention can further learn local features in samples from feature maps that have learned topological features, solving the problem that traditional GCNs ignore local features. The two convolution layers used in the adjacent attention module calculate and use the similarity vectors between nodes. The module constructs a graph to better represent the dependencies between nodes. As a graph optimization method, it suppresses the redundant edges of the graph structure caused by noise and ensures the robustness of graph representation learning. It is better than traditional GCN in suppressing noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0015] Figure 1 An overall step block diagram provided for an embodiment of the present invention.

[0016] Figure 2 This is a specific model structure diagram of the present invention. DETAILED DESCRIPTION

[0017] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] See also Figure 1 and Figure 2 As shown, the best embodiment of the present invention is further described; See also Figure 1 and Figure 2 As shown, the GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism includes the following steps: Step 1: Divide the end face film thickness signals collected from multiple displacement sensors into several sub-samples, assign corresponding labels to each sub-sample, and then perform fast Fourier transform on each sub-sample to display the frequency domain characteristics of each sub-sample; Step 2: Consider each subsample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighboring nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weight to obtain a sparse initial graph; The representation of the neighboring nodes of the node in step 2 is as follows: ; Where: Indicates the return node In the collection The first k nearest neighbor nodes in ; express The parameter of the number of neighboring nodes corresponding to the node in the algorithm; Indicated by A subset of nodes; Representation Node All adjacent nodes of .

[0019] The steps of simplifying the adjacency relationship in the initial graph based on the adjacency matrix and the adjacency weight are as follows: The adjacency matrix defines nodes that are related to each other, and does not mark the relationship between two nodes that are essentially unrelated; The adjacency weight indicates the closeness of the adjacency relationship; The Gaussian kernel function is used to evaluate the weight of the relationship between nodes: ; Where: Representation Node and The adjacency weight between them; Represents the width of the Gaussian kernel, which is used to control the length scale parameter of the input space. Its value is related to the size of the input image. represents the base of natural logarithms; After obtaining the adjacency weights of all sub-samples, the top k samples with the highest similarity in each sub-sample are retained, and the other sub-samples are set to zero to obtain a sparse initial graph.

[0020] Step 3: Adopt the adjacency attention mechanism and use the sample matrix in the sparse initial graph to extract the signal samples representing the nodes in the sample matrix and re-evaluate their neighboring relationships, generate a new adjacency matrix and adjacency weights, and obtain the optimized iterative graph; The step 3 comprises the following steps: The similarity vector of any two node samples x and y in the sparse initial graph is calculated through two convolutional layers: Where: and Respectively represent any two node samples in the sparse initial graph; and They are two different convolutional layers, Indicates that the convolution sum size is 1; Then use the LeackReLU function to activate and obtain the similarity vector between the two nodes ; Then calculate the adjacency weight between nodes x and y : Where: Represents the set of neighboring nodes of node x; Represent any three node samples in the sparse initial graph respectively; after obtaining the adjacency weights of all nodes and other nodes, use The algorithm constructs the graph and selects the one with the highest adjacency weight Nodes are regarded as nodes with neighboring relationships, and the adjacent attention score matrix is ​​obtained. , and finally superimposed with the adjacency matrix in the sparse initial graph to obtain the iterative adjacency matrix, realize the adjacency attention, and obtain the optimized iterative graph: ; Where: Represents the adjacency matrix of the sparse initial graph.

[0021] Step 4: Input the matrix of the iteration graph into the GCN module for feature extraction to obtain a feature graph for classification; The GCN module includes two layers of high-order graph convolutional layers; In the high-order graph convolutional layer, a single hidden layer is defined as: Where: Represents the features of the hidden nodes that have been learned; represents the dimension of the subgraph in the lth iteration; and Respectively indicated by The subgraph constructed by each node and its adjacent subgraphs; represents the set of all subgraphs in the local neighborhood of subgraph S; represents the learned features of the previous hidden node, Indicates the weight matrix that needs to be trained in the first and second high-order convolutional layers; Indicates the weight matrix that needs to be trained in the first and second high-order convolutional layers; In the high-order graph convolution layer, an edge relationship graph pooling method EdgePool is used; the graph pooling method EdgePool adaptively merges nodes with edges between each other during the iteration process to establish new nodes, and the node features of the new nodes are calculated by the following function: ; Where: and Represents the node features of node a and node b; and yes Edgepool Two learned parameters in the layer; After establishing the new node, the features of the new node of each sample are aggregated through global average pooling to obtain a feature map for classification.

[0022] Step 5: Connect the feature map for classification output by the GCN module to the fully connected layer, use Softmax to distinguish categories, and obtain the fault type.

[0023] The expression for using Softmax to distinguish categories is as follows: ; Where: For input A new feature representation, n represents the number of input data, Indicates the number of target categories in the classification task, for example Indicates The data is The probability of type, represents the number of hidden layers, is the feature representation, is a diagonal and satisfies The degree matrix of represents the weight matrix; For the Perform iterative optimization, where the loss function for iterative optimization is as follows: Where: represents the label set of the test set, represents the true value of the test sample label, Indicates the Among the test samples The invention uses high-order graph convolution layers to extract global features of data, and has stronger hierarchical representation learning capabilities compared to other GCNs. The residual connection convolution module introduced by the invention can further learn local features in samples from feature maps that have learned topological features, solving the problem of traditional GCN ignoring local features. The two convolution layers used by the adjacent attention module calculate and use the similarity vector between nodes. The module constructs a graph to better represent the dependencies between nodes. As a graph optimization method, it suppresses the redundant edges of the graph structure caused by noise and ensures the robustness of graph representation learning. It is better than traditional GCN in suppressing noise.

[0024] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism, characterized in that: The following steps are involved: Step 1: Divide the end face film thickness signals collected from multiple displacement sensors into several sub-samples, assign corresponding labels to each sub-sample, and then perform fast Fourier transform on each sub-sample to display the frequency domain characteristics of each sub-sample; Step 2: Consider each subsample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighboring nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weight to obtain a sparse initial graph; Step 3: Adopt the adjacency attention mechanism and use the sample matrix in the sparse initial graph to extract the signal samples representing the nodes in the sample matrix and re-evaluate their neighboring relationships, generate a new adjacency matrix and adjacency weights, and obtain the optimized iterative graph; Step 4: Input the matrix of the iteration graph into the GCN module for feature extraction to obtain a feature graph for classification; Step 5: Connect the feature map for classification output by the GCN module to the fully connected layer, use Softmax to distinguish categories, and obtain the fault type.

2. The GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1 is characterized in that: The representation of the neighboring nodes of the node in step 2 is as follows: ; Where: Indicates the return node In the collection The first k nearest neighbor nodes in ; The parameter representing the number of neighboring nodes of the corresponding node in the K-NN algorithm; represents a subset consisting of m nodes; Representation Node All adjacent nodes of .

3. The GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1 is characterized in that: The steps of simplifying the adjacency relationship in the initial graph based on the adjacency matrix and the adjacency weight are as follows: The adjacency matrix defines nodes that are related to each other, and does not mark the relationship between two nodes that are essentially unrelated; The adjacency weight indicates the closeness of the adjacency relationship; The Gaussian kernel function is used to evaluate the weight of the relationship between nodes: Representation Node and The adjacency weight between them; Represents the width of the Gaussian kernel, which is used to control the length scale parameter of the input space. Its value is related to the size of the input image. Represents the base of the natural logarithm. After obtaining the adjacency weights of all sub-samples, the top k samples with the highest similarity in each sub-sample are retained, and the other sub-samples are set to zero to obtain a sparse initial graph.

4. The GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1 is characterized in that: The step 3 comprises the following steps: The similarity vector of any two node samples x and y in the sparse initial graph is calculated through two convolutional layers: ; Where: and Respectively represent any two node samples in the sparse initial graph; and They are two different convolutional layers, Indicates that the convolution sum size is 1; then use the LeackReLU function to activate and obtain the similarity vector between the two nodes ; Then calculate the adjacency weight between nodes x and y : ; Where: Represents the set of neighboring nodes of node x; Represent any three node samples in the sparse initial graph respectively; after obtaining the adjacency weights of all nodes and other nodes, the K-NN algorithm is used to construct the graph and the node with the highest adjacency weight is selected Nodes are regarded as nodes with neighboring relationships, and the adjacent attention score matrix is ​​obtained. , and finally superimposed with the adjacency matrix in the sparse initial graph to obtain the iterative adjacency matrix, realize the adjacency attention, and obtain the optimized iterative graph: Where: Represents the adjacency matrix of the sparse initial graph.

5. The GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1 is characterized in that: The GCN module includes two layers of high-order graph convolutional layers; In the high-order graph convolutional layer, a single hidden layer is defined as: ; Where: Represents the features of the hidden nodes that have been learned; represents the dimension of the subgraph in the lth iteration; and Respectively indicated by The subgraph constructed by each node and its adjacent subgraphs; Representing a subgraph The set of all subgraphs in the local neighborhood of ; represents the learned features of the previous hidden node, Represents the weight matrix trained in the 1st and 2nd high-order graph convolutional layers; In the high-order graph convolution layer, an edge relationship graph pooling method EdgePool is used; the graph pooling method EdgePool adaptively merges nodes with edges between each other during the iteration process to establish new nodes, and the node features of the new nodes are calculated by the following function: ; Where: and Represents the node features of node a and node b; and yes Edgepool Two learned parameters in the layer; After establishing the new node, the features of the new node of each sample are aggregated through global average pooling to obtain a feature map for classification.

6. The GCN fluid dynamic pressure mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1 is characterized in that: The expression for using Softmax to distinguish categories is as follows: ; Where: For input The new feature representation of Indicates the number of target categories for the classification task, represents the number of hidden layers, H is the feature representation, is a diagonal and satisfies The degree matrix of represents the weight matrix; For the Perform iterative optimization, where the loss function for iterative optimization is as follows: ； Where: represents the label set of the test set, represents the true value of the test sample label, Indicates the Among the test samples The predicted label value for each class.

Images