GCN Hydrodynamic Mechanical Seal Diagnosis Method Based on Adjacent Attention Mechanism

Abstract

This application belongs to the technical field of mechanical fault diagnosis, and relates to a GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism; the frequency domain features of the original time domain signal are extracted using the fast Fourier transform, and an initial graph is generated through the k-NN algorithm; then the signal samples representing nodes in the initial graph are extracted using the adjacency attention mechanism and their neighbor relationships are re-evaluated, and a new adjacency matrix and adjacency weights are generated to obtain an optimized graph after iteration; then the optimized graph is input into the GCN module for feature extraction; finally, the feature map that can be used for classification output by the GCN module is directly connected to a fully connected layer to output the result, obtaining the fault type; by re-evaluating the neighbor relationships of the nodes using the adjacency attention mechanism, a new adjacency matrix and adjacency weights are generated, and an optimized iterative graph is obtained, enabling further learning of the local features in the samples from the feature map of the learned topological features, solving the problem of the traditional GCN's neglect of local features.

Description

Technical Field

[0001] This application belongs to the technical field of mechanical fault diagnosis. More specifically, it relates to a GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism. Background Technique

[0002] The main coolant pump of a million-kilowatt nuclear power plant reactor (referred to as the nuclear main pump) is a key device 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 mechanical seal spare parts. Achieving intelligent fault diagnosis for it is of great significance for improving the operation safety of nuclear power plants and the reliability of nuclear power equipment. However, at present, the domestic independent research on the development, manufacturing, and maintenance of hydrodynamic mechanical seals is relatively preliminary, and there is no effective practical experience in the fault monitoring, repair, and reuse of sealing rings in China.

[0003] With the development of sensors and monitoring means, as well as the progress of computer computing power and database technology, in order to cope with the high quantity and high dimension characteristics of industrial equipment monitoring data that are increasingly generated, data-driven fault diagnosis methods have gradually attracted people's attention. Data-driven methods emphasize directly using a large amount of data generated by the system, and constructing a diagnostic model through algorithms such as machine learning (ML) and even deep learning (DL). They automatically learn the working state and fault characteristics of the system from the monitoring signals of the equipment and give a diagnostic result. Such methods usually do not require in-depth understanding of the system working principle, but learn the behavior pattern of the system from the data. Data-driven methods can avoid analyzing and researching complex, non-linear, and difficult-to-model systems and exploring their mechanisms, and have good adaptability to faults with strong uncertainty and randomness. They can discover potential associations and patterns that are difficult to identify manually. For large-scale complex systems, or for systems with unclear fault mechanisms, data-driven methods can more flexibly adapt to different working conditions and fault modes, especially performing well when dealing with a large amount of unstructured data or scenarios with complex patterns. In addition, since such methods generally establish a feature learning and recognition model through a computer to directly use signals for fault diagnosis, the data-driven fault diagnosis method is more easily integrated into an automatic system to play a role.

[0004] 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 interdependencies between data or various physical measurements of multiple sensors. In addition, in the real world, faults may change the operating state of a mechanical system and cause changes in the dependencies between samples obtained from measurement signals. Based on this, many researchers have begun to attempt to describe fault data in the form of graph-level data in order to explore potential features in the data space or jointly process multi-sensor data. And this practice of transforming fault signal data into graph data essentially changes the 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 thus various data-driven classification models mentioned above cannot be used for such data. GCN can learn the topological structure between samples and aggregate highly relevant features, and simulate higher-level features in the sample space by using the graph-level expression of data. However, the application of GCN in the fault diagnosis of hydrodynamic mechanical seals still has limitations: First, the original monitoring signals in hydrodynamic mechanical seals themselves contain inseparable noise and errors introduced by instruments. Therefore, the graphs directly constructed from these signal samples contain 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. Second, the operating state of hydrodynamic mechanical seals changes dynamically over time, so the discriminative features of fault signals often lie not only within a single node but also in the interdependencies between multiple nodes. However, traditional GCN methods 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 fault categories, and learning and extracting these features play a crucial role in improving the classification performance of the model. However, most traditional GCN models focus on learning and extracting global features in samples and ignore the importance of local features. Summary of the Invention

[0005] The reactor coolant pump of a million-kilowatt nuclear power plant (referred to as the nuclear main pump) is a key piece of 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 structure of the nuclear main pump is characterized by high technical difficulty, high cost, frequent preventive maintenance, and a large inventory of old mechanical seal spare parts. Achieving the intelligent diagnosis of its faults is of great significance for improving the operation safety of nuclear power plants and the reliability of nuclear power equipment. However, at present, the domestic independent research on the development, manufacturing, and maintenance of hydrodynamic mechanical seals is relatively preliminary, and there is no effective practical experience in the fault monitoring, repair, and reuse of seal rings in China.

[0006] With the development of sensors and monitoring means, as well as the progress of computer computing power and database technology, in order to cope with the high volume and high dimensionality characteristics of industrial equipment monitoring data that are increasingly generated, data-driven fault diagnosis methods have gradually attracted people's attention. Data-driven methods emphasize directly using the large amount of data generated by the system, and constructing a diagnostic model through algorithms such as machine learning (ML) and even deep learning (DL). Automatically learn the working state and fault characteristics of the system from the monitoring signals of the equipment and give the diagnostic results. This type of method usually does not require an in-depth understanding of the system working principle, but learns the behavior pattern of the system from the data. The data-driven method can avoid the analysis and research of complex, non-linear, and difficult-to-model systems and the exploration of mechanisms, has 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 systems with unclear fault mechanisms, the data-driven method can more flexibly adapt to different working conditions and fault modes, especially performing well in dealing with a large amount of unstructured data or scenarios with complex patterns. In addition, since such methods generally establish a feature learning and recognition model through a computer to directly use the signal for fault diagnosis, the data-driven fault diagnosis method is easier to integrate into an automatic system to play a role.

[0007] 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 interdependencies between data or various physical measurements of multiple sensors. In addition, in the real world, faults may change the operating state of a mechanical system and cause changes in the dependencies between samples obtained from measurement signals. Based on this, many researchers have begun to attempt to describe fault data in the form of graph-level data in order to explore potential features in the data space or jointly process multi-sensor data. And this practice of transforming fault signal data into graph data essentially changes the data originally belonging to the Euclidean space into data in the non-Euclidean space, which makes operations that are easy to implement in the Euclidean space (such as convolution, etc.) difficult to implement on data in the non-Euclidean space, and thus various data-driven classification models mentioned above cannot be used for such data. GCN can learn the topological structure between samples and aggregate highly relevant features, and simulate higher-level features in the sample space by using the graph-level expression of data. However, the application of GCN in the fault diagnosis of hydrodynamic mechanical seals still has limitations: First, the original monitoring signals in hydrodynamic mechanical seals themselves contain inseparable noise and errors introduced by instruments. Therefore, the graphs directly constructed from these signal samples contain a large number of errors and redundant edges. This has an adverse impact on the stability and accuracy of the GCN model in feature learning. Second, the operating state of hydrodynamic mechanical seals changes dynamically over time. Therefore, the discriminant features of fault signals are often not only reflected within a single node but also in the interdependencies between multiple nodes. However, traditional GCN methods 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 fault categories, and learning and extracting these features play a crucial role in improving the classification performance of the model. However, most traditional GCN models focus on learning and extracting global features in samples and ignore the importance of local features.

[0008] The present invention provides a GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism, aiming to solve the technical problem that the application of GCN in the fault diagnosis of hydrodynamic mechanical seals still has limitations.

[0009] The GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism includes the following steps:

[0010] 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 a fast Fourier transform on each sub-sample to display the frequency domain features of each sub-sample;

[0011] Step 2: Regarding each sub-sample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighbor nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights to obtain a sparse initial graph;

[0012] Step 3: Adopt the adjacency attention mechanism to use the sample matrix in the sparse initial graph, extract the signal samples representing the nodes in the sample matrix and re-evaluate their proximity relationships to generate a new adjacency matrix and adjacency weights, and obtain an optimized iterative graph;

[0013] Step 4: Input the matrix of the iterative graph into the GCN module for feature extraction to obtain a feature graph for classification;

[0014] Step 5: Connect the feature graph for classification output by the GCN module with a fully connected layer, and use Softmax for class discrimination to obtain the fault type.

[0015] In the present invention, the proximity relationship of nodes is re-evaluated through the adjacency attention mechanism to generate a new adjacency matrix and adjacency weights, and an optimized iterative graph is obtained, so as to further learn the local features in the samples from the feature graph of the already learned topological features, solving the problem of the traditional GCN's neglect of local features.

[0016] Preferably, the representation of the neighbor nodes of the nodes in Step 2 is as follows:

[0017] ; where: represents the returned node in the set the first k nearest neighbor nodes; represents the parameter of the number of neighbor nodes corresponding to the node in the K-NN algorithm; represents a subset composed of m nodes; represents the node all adjacent nodes.

[0018] Preferably, the step of simplifying the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights is as follows:

[0019] Among them, the adjacency matrix defines the nodes with a mutual relationship and does not mark the relationship between two nodes that are substantially unrelated;

[0020] The adjacency weight represents the degree of closeness of the adjacency relationship;

[0021] The Gaussian kernel function is used to evaluate the weight of the relationship between nodes: ; where: represents the node and the adjacency weight between; denotes the width of the Gaussian kernel, which is a parameter for controlling the length ratio in the input space, and its value is related to the size of the input graph; denotes the base of the natural logarithm;

[0022] After obtaining the adjacency weights of all subsamples, retain the top k samples with the highest similarity in each subsample and set the other subsamples to zero to obtain a sparse initial graph.

[0023] Preferably, step 3 includes the following steps:

[0024] Calculate the similarity vector for any two node samples x and y in the sparse initial graph through two convolutional layers: ; where: and respectively represent any two node samples in the sparse initial graph; are two different convolutional layers, where denotes that the convolution sum size is 1;

[0025] Then activate using the LeackReLU function to obtain the similarity vector between the two nodes Then calculate the adjacency weight between nodes x and y : ; where: denotes the set composed of the neighborhood nodes of node x; respectively represent any three node samples in the sparse initial graph;

[0026] After obtaining the adjacency weights of all nodes with other nodes, use the K-NN algorithm for graph construction, and select the nodes with the highest adjacency weights as the nodes with near-neighbor relationships to obtain the adjacency attention score matrix Finally, superimpose it with the adjacency relationship matrix in the sparse initial graph to obtain the iterative adjacency matrix, realize adjacency attention, and obtain the optimized iterative graph: ; where: denotes the adjacency relationship matrix of the sparse initial graph.

[0027] Preferably, the GCN module includes two layers of high-order graph convolutional layers;

[0028] In the high-order graph convolutional layer, the single-layer hidden layer is defined as: ; where: denotes the learned features of the hidden nodes; denotes the dimension of the subgraph in the l-th iteration; and respectively represent the subgraph constructed by nodes and its neighboring subgraphs; denotes the subgraph The set of all subgraphs within the local neighborhood of; Represents the features of the previously learned hidden node, Represents the weight matrices trained in the 1st and 2nd high-order graph convolutional layers;

[0029] In the high-order graph convolutional layer, the graph pooling method EdgePool for edge relationships is adopted; through the graph pooling method EdgePool, nodes with edges between each other are adaptively merged during the iteration process to establish new nodes, and the node features of the new nodes are calculated by the following function: ; In the formula: and Represent the node features of node a and node b; With Is Edgepool Two learning parameters in the layer;

[0030] After establishing the new nodes, the features of the new nodes of each sample are aggregated through global average pooling to obtain a feature map for classification.

[0031] Preferably, the expression for using Softmax for class discrimination is as follows: ; In the formula: Is the input The brand-new feature representation of, Represents the number of target categories for the classification task, Represents the number of hidden layers, Is the feature representation, Is a diagonal and satisfies Degree matrix of, Represents the weight matrix;

[0032] For the Iterative optimization is performed, and the loss function for iterative optimization is as follows: ; In the formula: Represents the label set of the test set, Represents the true label of the test sample, Represents the In the Value of the predicted label of the

[0033] In the present invention, the high-order graph convolutional layer is used to extract the global features of the data. Compared with other GCNs, it has a stronger hierarchical representation learning ability; and the introduced residual connection convolutional module is used to further learn the local features in the samples from the feature map that has already learned the topological features, solving the problem of traditional GCNs ignoring local features; the two convolutional layers used in the adjacency attention module calculate and utilize the similarity vector between nodes Construct a module diagram to better represent the dependencies between nodes. As a graph optimization method, it plays a role in suppressing the redundant edges of the graph structure caused by noise and ensuring the robustness of graph representation learning. Moreover, its performance in suppressing noise is superior to that of traditional GCNs. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0035] Figure 1 It is the overall step block diagram provided by the embodiments of the present invention.

[0036] Figure 2 It is the specific model structure diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] In order to make the technical problems, technical solutions, and beneficial effects to be solved by the present application more clearly understood, the following further details the present application in conjunction with the 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.

[0038] See Figure 1 and Figure 2 as shown, for a further description of the optimal embodiment of the present invention;

[0039] See Figure 1 and Figure 2 as shown, a GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism includes the following steps:

[0040] 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 a fast Fourier transform on each sub-sample to display the frequency domain characteristics of each sub-sample;

[0041] Step 2: Regard each sub-sample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighbor nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights to obtain a sparse initial graph;

[0042] The representation of the neighbor nodes of the nodes in Step 2 is as follows:

[0043] ; where: represents returning the node The first k nearest neighbor nodes in the set ; Denote The parameter representing the number of neighboring nodes corresponding to the node in the algorithm; Denote A subset consisting of nodes; Denote all adjacent nodes of node

[0044] The steps for simplifying the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights are as follows:

[0045] Among them, the adjacency matrix defines the nodes with relationships to each other and does not mark the relationship between two nodes that are essentially unrelated;

[0046] The adjacency weight represents the closeness of the adjacency relationship;

[0047] Use the Gaussian kernel function to evaluate the weight of the relationship between nodes: ; In the formula: Denote and The adjacency weight between nodes; Denote the width of the Gaussian kernel, which is a parameter for controlling the length ratio of the input space, and its value is related to the size of the input graph; Denote the base of the natural logarithm;

[0048] After obtaining the adjacency weights of all subsamples, retain the top k samples with the highest similarity in each subsample and set the other subsamples to zero to obtain a sparse initial graph.

[0049] Step 3: Use the adjacency attention mechanism to utilize the sample matrix in the sparse initial graph, 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 an optimized iterative graph;

[0050] The said step 3 includes the following steps:

[0051] Calculate the similarity vector of any two node samples x and y in the sparse initial graph through two convolutional layers: In the formula: and Respectively denote any two node samples in the sparse initial graph; and Are two different convolutional layers, where Denote the convolution sum size as 1;

[0052] Then activate using the LeackReLU function to obtain the similarity vector between the two nodes; then calculate the adjacency weight between nodes x and y : In the formula: represents the set composed of the neighboring nodes of node x;

[0053] respectively represent any three node samples in the sparse initial graph; after obtaining the adjacency weights between all nodes and other nodes, the algorithm is used for graph construction, and the nodes with the highest adjacency weights are selected as the nodes with neighbor relationships to obtain the adjacency attention score matrix , and finally it is superimposed with the adjacency relationship matrix in the sparse initial graph to obtain the iterative adjacency matrix, realizing adjacency attention and obtaining the optimized iterative graph: ; in the formula: represents the adjacency relationship matrix of the sparse initial graph.

[0054] Step 4: Input the matrix of the iterative graph into the GCN module for feature extraction to obtain the feature map for classification;

[0055] The GCN module includes two layers of high-order graph convolutional layers;

[0056] In the high-order graph convolutional layer, a single hidden layer is defined as: In the formula: represents the feature of the learned hidden node; represents the dimension of the subgraph in the l-th iteration; and respectively represent the subgraph constructed by nodes and its neighboring subgraphs; represents the set of all subgraphs within the local neighborhood of subgraph S; represents the feature of the previous learned hidden node, represents the weight matrix to be trained in the first and second high-order convolutional layers; represents the weight matrix to be trained in the first and second high-order convolutional layers;

[0057] In the high-order graph convolutional layer, the graph pooling method EdgePool for edge relationships is adopted; through the graph pooling method EdgePool, nodes with edges between each other are adaptively merged during the iteration to establish new nodes, and the node features of the new nodes are calculated by the following function: ; in the formula: and represent the node features of node a and node b; and are Edgepool two learning parameters in the

[0058] After establishing the new nodes, the features of the new nodes of each sample are aggregated through global average pooling to obtain a feature map for classification.

[0059] Step 5: Connect the feature map for classification output by the GCN module to the fully connected layer, and use Softmax for class discrimination to obtain the fault type.

[0060] The expression for using Softmax for class discrimination is as follows: ; In the formula: is the input of the new feature representation, n represents the number of input data, represents the number of target categories for the classification task. For example represents the th data is the probability of the type, represents the number of hidden layers, is the feature representation, is a diagonal and satisfies degree matrix, represents the weight matrix;

[0061] For the perform iterative optimization, and the loss function for iterative optimization is as follows: In the formula: represents the label set of the test set, represents the true label value of the test sample, represents the value of the predicted label of the th test sample for the th category. In the present invention, the high-order graph convolutional layer is used to extract the global features of the data. Compared with other GCNs, it has a stronger hierarchical representation learning ability; and the residual connection convolutional module introduced by it is used to further learn the local features in the samples from the feature map that has already learned the topological features, solving the problem of the traditional GCN's neglect of local features; the two convolutional layers used in the adjacency attention module calculate and utilize the similarity vector between nodes and use the module to construct a graph to better represent the dependence relationship between nodes. As a graph optimization method, it plays a role in suppressing the redundant edges of the graph structure caused by noise and ensuring the robustness of graph representation learning, and its performance in suppressing noise is superior to that of the traditional GCN.

[0062] The above are only the preferred embodiments of the present application, and are not used to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A GCN hydrodynamic mechanical seal diagnosis method based on an adjacency attention mechanism, characterized in that Including 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 a fast Fourier transform on each sub-sample to display the frequency domain characteristics of each sub-sample; Step 2: Consider each sub-sample as a node of the initial graph, use the K-NN algorithm to find the first k nearest neighbor nodes for each node, and then simplify the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights to obtain a sparse initial graph; Step 3: Adopt an adjacency attention mechanism to use the sample matrix in the sparse initial graph, extract the signal samples representing the nodes in the sample matrix and re-evaluate their proximity relationship to generate a new adjacency matrix and adjacency weights, and obtain an optimized iterative graph; Step 4: Input the matrix of the iterative graph into the GCN module for feature extraction to obtain a feature graph for classification; Step 5: Connect the feature graph for classification output by the GCN module to a fully connected layer, and use Softmax for class discrimination to obtain the fault type.

2. The GCN hydrodynamic mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1, wherein The representation of the neighborhood nodes of the node in step 2 is as follows: ; where: represents the return node in the set the first k nearest neighborhood nodes; represents the parameter of the number of neighborhood nodes corresponding to the node in the K-NN algorithm; represents a subset composed of m nodes; represents the node all adjacent nodes of.

3. The GCN hydrodynamic mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1, wherein, The step of simplifying the adjacency relationship in the initial graph based on the adjacency matrix and adjacency weights is as follows: Among them, the adjacency matrix defines the nodes with a relationship to each other, and does not mark the relationship between two nodes that are substantially unrelated; The adjacency weight represents the degree of closeness of the adjacency relationship; The weight of the relationship between nodes is evaluated using a Gaussian kernel function: In the formula: represents a node and the adjacency weight between them; represents the width of the Gaussian kernel, which is a parameter for controlling the length ratio of the input space, and its value is related to the size of the input graph; represents the base of the natural logarithm; After obtaining the adjacency weights of all sub-samples, retain the first k samples with the highest similarity in each sub-sample, and set the other sub-samples to zero to obtain a sparse initial graph.

4. The GCN hydrodynamic mechanical seal diagnosis method based on the adjacent attention mechanism according to claim 1, characterized in that Step 3 includes the following steps: Calculate the similarity vector for any two node samples \(x\) and \(y\) in the sparse initial graph through two convolutional layers: ; where: and respectively represent any two node samples in the sparse initial graph; and are two different convolutional layers, where represents that the convolution sum size is 1; then activate with the LeackReLU function to obtain the similarity vector between the two nodes ; then calculate the adjacency weight between nodes \(x\) and \(y\) : ; where: represents the set composed of the neighborhood nodes of node \(x\); respectively represent any three node samples in the sparse initial graph; after obtaining the adjacency weights of all nodes with other nodes, use the K-NN algorithm for graph construction, and select the nodes with the highest adjacency weights as the nodes with near neighbor relationships to obtain the adjacency attention score matrix , and finally superimpose it with the adjacency relationship matrix in the sparse initial graph to obtain the iterative adjacency matrix, realize adjacency attention, and obtain the optimized iterative graph: where: represents the adjacency relationship matrix of the sparse initial graph.

5. The GCN hydrodynamic mechanical seal diagnosis method based on the adjacency attention mechanism according to claim 1, wherein 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: ; Wherein: represents the features of the learned hidden nodes; represents the dimension of the subgraph in the l-th iteration; and respectively represent the subgraph constructed by nodes and its neighboring subgraphs; represents the subgraph the set of all subgraphs within the local neighborhood of; represents the features of the previous learned hidden node, represents the weight matrix trained in the 1st and 2nd high-order graph convolutional layers; Adopt the graph pooling method EdgePool for edge relationships in the high-order graph convolutional layer; through the graph pooling method EdgePool, nodes with edges to each other are adaptively merged during the iteration process to establish new nodes, and the node features of the new nodes are calculated by the following function: ; where: and represent the node features of node a and node b; and are Edgepool two learning parameters in the layer; After establishing new nodes, aggregate the features of the new nodes of each sample through global average pooling to obtain a feature graph for classification.

6. The GCN hydrodynamic mechanical seal diagnosis method based on the adjacency attention mechanism according to claim 1, wherein The expression for class discrimination using Softmax is as follows: ; where: is the input 's brand-new feature representation, represents the number of target classes for the classification task, represents the number of hidden layers, H is the feature representation, is a diagonal and satisfies 's degree matrix, represents the weight matrix; For the said perform iterative optimization, where the loss function for iterative optimization is as follows: ； Wherein: represents the label set of the test set, represents the true label of the test sample, represents the th value of the predicted label of the

Images