A complex industrial process fault diagnosis method and its application

By converting the fault diagnosis of complex industrial processes into graph classification problems, using heterogeneous graphs and multiple edge types to represent the interactive relationship between sensors, the problem of difficult to represent the relationship between sensors in the prior art is solved, and more accurate failures are achieved Identify and diagnose effects.

CN114298095BActive Publication Date: 2024-12-06NO 719 RES INST CHINA SHIPBUILDING IND
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202111554256.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-17
Publication Date
2024-12-06
Estimated Expiration
2041-12-17

AI Technical Summary

Technical Problem

It is difficult for the prior art to effectively use machine learning methods to diagnose complex industrial processes, especially in high-dimensional, multi-dimensional and high-interaction data environments, and the multiple interactions between sensors are difficult to accurately represent and utilize.

Method used

Transform fault diagnosis problems into graph classification problems, use multiple different types of edges to represent the interaction relationship between sensors in heterogeneous graphs, and identify and extract nodes through the interaction perception layer based on attention mechanism and graph convolutional neural network module. The interactive relationship and feature vectors of the two, construct heterogeneous graphs to improve the fault recognition effect.

Benefits of technology

Through heterogeneous pattern structure and the representation of multiple edge types, the complex interactive relationships between sensors can be captured more accurately, the recognition effect of fault diagnosis of complex industrial processes can be improved, and the identification ability of fault categories can be enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114298095B_ABST
    Figure CN114298095B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and application for fault diagnosis of complex industrial processes. The method comprises the steps of: collecting measurement signals of multiple sensors, constructing a fully connected graph, wherein each node in the fully connected graph corresponds to a sensor; inputting the fully connected graph into a diagnostic model, wherein the diagnostic model comprises a relationship discovery module, a feature extraction module and a fault identification module, wherein the relationship discovery module is used to identify the interaction relationship between nodes and construct a heterogeneous graph, wherein the heterogeneous graph uses a plurality of different types of edges to represent different types of interaction relationships between nodes, wherein the feature extraction module is used to extract feature vectors from the heterogeneous graph, and wherein the fault identification module is used to output a fault diagnosis result according to the feature vectors. The present invention converts the fault diagnosis problem into a graph classification problem, and uses a plurality of different types of edges to represent different types of interaction relationships between sensors, which can improve the recognition effect of fault diagnosis of complex industrial processes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and more specifically, relates to a complex industrial process fault diagnosis method and application. Background Art

[0002] With the rapid development of complex industrial processes such as power systems and chemical processes, an effective and reliable health monitoring system has become increasingly important in order to reduce maintenance costs and ensure the safe operation of industrial systems. Generally speaking, a monitoring system includes many sensors and equipment for collecting measurements from multiple sensors. Due to the large-scale and complex control engineering of the process industry, the collected data has the characteristics of high dimension, multi-dimensionality and high interaction. For monitoring systems used on-site, a large number of maintenance records are usually established, which makes it possible to use supervised machine learning methods. However, the prior art still lacks a feasible and effective method for using machine learning methods to effectively diagnose complex industrial process faults. Summary of the invention

[0003] In response to at least one defect or improvement need in the prior art, the present invention provides a complex industrial process fault diagnosis method and application, which converts the fault diagnosis problem into a graph classification problem, and uses a variety of different types of edges to represent different types of interaction relationships between sensors, which can improve the recognition effect of fault diagnosis in complex industrial processes.

[0004] To achieve the above object, according to a first aspect of the present invention, a complex industrial process fault diagnosis method is provided, comprising the steps of:

[0005] Collecting measurement signals of multiple sensors and constructing a fully connected graph, wherein each node in the fully connected graph corresponds to a sensor, and an initial feature of each node is a measurement signal of the corresponding sensor;

[0006] The fully connected graph is input into a diagnosis model, which includes a relationship discovery module, a feature extraction module and a fault identification module. The relationship discovery module is used to identify the interaction relationship between nodes to construct a heterogeneous graph. The heterogeneous graph uses multiple different types of edges to represent different types of interaction relationships between nodes. The feature extraction module is used to extract feature vectors from the heterogeneous graph. The fault identification module is used to output fault diagnosis results based on the feature vectors.

[0007] Furthermore, the construction of the fully connected graph comprises the steps of:

[0008] The number of sensors is recorded as n, and the measurement signals of n sensors are expressed as s i , where 1≤i≤n, a time window of length m is used to measure the signal s i Divide into multivariate time series segments x i, using n multivariate time series segments x of the same time window i Construct the fully connected graph, which includes n nodes, each of which is x i .

[0009] Furthermore, the heterogeneous graph is represented as G = (V, E, R e ), V is a node set, including node v i ∈V, E is the edge set, including edge e i,j ∈E,R e A collection of edge types.

[0010] Furthermore, the heterogeneous graph includes multiple subgraphs, each subgraph representing a type of interaction relationship between nodes.

[0011] Furthermore, the relationship discovery module includes multiple independent attention-based interaction perception layers, each of which is used to identify a type of interaction relationship between nodes.

[0012] Furthermore, the interaction perception layer is used to implement the following steps:

[0013] Let the two nodes be x i and x j , the learned attention parameters for

[0014]

[0015] Where σ(·) is the activation function, CONCAT represents the concatenation operation, are the parameters of the parameter sharing layer;

[0016] Compute Node x i and x j The standard relationship score of i,j

[0017]

[0018] Where τ(·) is the threshold function, and all relationship scores less than the threshold will be set to 0. sparsemax(·) represents the sparsemax function, and max means taking the maximum value.

[0019] Furthermore, the feature extraction module includes multiple independent graph convolutional neural network modules, multiple independent aggregation modules and a splicing module, the graph convolutional neural network module is used to obtain the node features of each subgraph, the aggregation module is used to obtain the global features of each subgraph, and the splicing module is used to splice the global features of all subgraphs.

[0020] According to a second aspect of the present invention, there is provided a complex industrial process fault diagnosis system, comprising:

[0021] A data acquisition and preprocessing module, used to acquire measurement signals of multiple sensors and construct a fully connected graph, wherein each node in the fully connected graph corresponds to a sensor, and the initial feature of each node is the measurement signal of the corresponding sensor;

[0022] A diagnosis module is used to input the fully connected graph into a diagnosis model, wherein the diagnosis model includes a relationship discovery module, a feature extraction module and a fault identification module. The relationship discovery module is used to identify the interaction relationship between nodes to construct a heterogeneous graph. The heterogeneous graph uses multiple different types of edges to represent different types of interaction relationships between nodes. The feature extraction module is used to extract feature vectors from the heterogeneous graph. The fault identification module is used to output fault diagnosis results based on the feature vectors.

[0023] According to a third aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0024] According to a fourth aspect of the present invention, there is provided a storage medium having a computer program stored thereon, wherein the computer program implements any one of the above methods when executed by a processor.

[0025] In general, the present invention transforms the fault diagnosis problem into a graph classification problem, and fully considers the different types of interaction relationships between sensors. It uses multiple different types of edges to represent the different types of interaction relationships between sensors to construct heterogeneous graphs, which can improve the recognition effect of fault diagnosis in complex industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of a complex industrial process fault diagnosis method according to an embodiment of the present invention;

[0027] Figure 2 Schematic diagram of a diagnostic model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention 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 invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] A complex industrial process fault diagnosis method according to an embodiment of the present invention can perform fault diagnosis on a complex industrial process.

[0030] When a subunit of a complex industrial process fails, due to the complex control relationship, the fault will propagate to other subsystems and even affect the operation of the entire system. In principle, due to the interaction between different subsystems, the readings of multiple sensors will deviate from the normal state. At the same time, some sensor measurements will respond to different faults, that is, one fault category is related to multiple variables and the distribution is irregular. Therefore, in the fault diagnosis of large-scale industrial processes, it is of great significance to mine the interaction and fusion information of multi-sensor measurements.

[0031] Due to the complex interactions between sensor measurements, a structured graph is a suitable way to describe the characteristics and relationships of sensor data. Each sensor measurement can be mapped to a node, and these nodes are connected by implicit edges representing interactions. When multi-sensor signals are converted into graph-structured data, the graph can contain multiple fault information such as the root cause of the fault and fault propagation. A key task of fault diagnosis is to identify the fault category, so we convert multi-sensor signals into graph data and define the fault identification problem as a graph classification problem.

[0032] When fault identification is extended to graph classification, the following issues need to be addressed: (1) the graph construction problem of sensor data; and (2) the learning method of graph representation. Graph neural network (GNN) is an effective method to fuse information from multi-sensor readings through a message passing mechanism. In general, exploring the interactions between sensor measurements can be achieved through empirical analysis, for example, by constructing graphs based on the system structure or similar relationships in multi-sensor signals (such as K-nearest neighbor graphs). Although these graph structures are interpretable, due to the fixedness of the system structure and dynamic model, the above graph construction methods will result in different faults having similar graph structures. Similar graph structures will lead to the performance degradation of GNNs because the features of nodes are always aggregated from a few fixed neighbors. At the same time, in actual process industries, there are multiple relationships between sensor signals, including but not limited to similarity, correlation, and functional relationships. For example, in power systems, feedwater pump work and condensate pump work are calculated using the same equation related to steam mass flow, inlet feedwater enthalpy, and outlet feedwater enthalpy. Therefore, there is similarity between the two variables, feedwater pump work and condensate pump work, and they are related to the other three variables through functional equations, so there is also functional correlation between the above variables. Therefore, one relationship cannot well describe multiple interactions between sensor measurements. In the embodiment of the present invention, a heterogeneous graph structure is introduced to represent the complex interactions of sensor measurements, and multiple different types of edges are used to represent different types of interaction relationships between nodes.

[0033] A complex industrial process fault diagnosis method according to an embodiment of the present invention considers multiple interactions between sensor nodes in a graph and uses GNN to fuse the measurement information of each sensor. An attention mechanism is used to learn the relationship scores between nodes in different representation spaces, thereby constructing a heterogeneous sensor network. Since the actual relationship between nodes is difficult to model, a relationship score is used to represent the importance of the interaction between nodes. These learned relationship scores constitute the edge attributes of the graph. The heterogeneous graph can be divided into multiple subgraphs according to the edge type. Then, a parallel GNN module is applied to update the node features of different edge types respectively, and joint training is performed. The subgraphs are divided according to different edge types, and the read out function is used to read the updated node embedding based on the edge type to obtain the subgraph representation. These subgraph features are used to generate the final graph embedding through an aggregation function. Therefore, the multivariate time series measured in the complex industrial process can be fused, and the fused information includes not only the data information but also the interaction relationship between the sensor signals, which can better promote fault identification.

[0034] like Figure 1 As shown, a complex industrial process fault diagnosis method according to an embodiment of the present invention comprises the following steps:

[0035] S101, collecting measurement signals of multiple sensors and constructing a fully connected graph, where each node in the fully connected graph corresponds to a sensor, and the initial feature of each node is the measurement signal of the corresponding sensor.

[0036] The number of sensors is recorded as n, and these signals are continuous multivariate time series. The measurement signals of n sensors can be expressed as Where 1≤i≤n, t is the sampling point, and a time window of length m is used to measure the signal s i Partition into multivariate time series segments Using n multivariate time series segments x of the same time window i Construct the fully connected graph, which includes n nodes x i . n multivariate time series segments x i It is input as a sample into the diagnosis model in step S102.

[0037] S102, input the fully connected graph into a diagnosis model, the diagnosis model includes a relationship discovery module, a feature extraction module and a fault identification module, the relationship discovery module is used to identify the interaction relationship between nodes to construct a heterogeneous graph, the heterogeneous graph uses a variety of different types of edges to represent different types of interaction relationships between nodes, the feature extraction module is used to extract feature vectors from the heterogeneous graph, and the fault identification module is used to output fault diagnosis results according to the feature vectors.

[0038] Furthermore, the diagnosis model is implemented based on interaction-aware graph neural networks (IAGNN).

[0039] Furthermore, the heterogeneous graph is represented as G = (V, E, R e ), V is a node set, including node v i ∈V, E is the edge set, including edge e i,j ∈E,R e is a set of edge types, where nodes and edges have corresponding attributes, represented as as well as

[0040] Furthermore, typical interaction relationships in sensors include but are not limited to similarity, correlation and functional relationships. Several types of edges are used to represent these interaction relationships.

[0041] Furthermore, the heterogeneous graph includes multiple subgraphs, each of which represents a type of interaction relationship between nodes. Assuming that there are two types of interaction relationships between all sensors, the constructed heterogeneous graph includes subgraphs. Figure 1 Kazuko Figure 2 ,son Figure 1 Represents the interaction relationship between all nodes 2, Figure 2 Indicates the interaction relationship 2 between all nodes. If two nodes have an interaction relationship 1, then Figure 1 There is an edge connecting the two nodes, otherwise Figure 1 There is no edge between the two nodes. If there is an interaction relationship 2 between the two nodes, then Figure 2 There is an edge connecting the two nodes, otherwise Figure 2 There is no edge between these two nodes.

[0042] The preferred implementation of the diagnostic model is described in detail below.

[0043] The overall framework of the diagnostic model is as follows Figure 2 As shown in the figure, the framework is an end-to-end supervised graph neural network model, which consists of three basic parts and is jointly trained: (1) learning pairwise relationships between nodes and automatically constructing heterogeneous graphs. (2) using independent graph neural networks to learn node representations for heterogeneous graphs with different edge types. (3) using the application readout function to aggregate the node representations into a graph representation and classify various faults.

[0044] (1) Relationship discovery module (also known as relationship learning layer)

[0045] The relationship discovery module includes multiple independent attention-based interaction perception layers, each of which is used to identify a type of interaction relationship between nodes.

[0046] Multiple pairwise interactions between nodes constitute an implicit sensor network. Based on domain knowledge, the implicit sensor network is converted into explicit graph structures such as K-nearest neighbor graphs and correlation graphs based on different correlations. At the same time, due to the subtle differences in node interactions between different fault types, graphs constructed with prior knowledge may not be suitable for model learning. Therefore, in order to distinguish the graph structures of various faults, an embodiment of the present invention adopts multiple independent interaction perception layers based on the attention mechanism, each of which is used to identify a type of interaction relationship between nodes, and represents the measurement data of the sensor network as a heterogeneous graph forming a node type and multiple edge types.

[0047] The relational learning layer takes the fully connected graph constructed in the first step as input. The goal of this layer is to learn the i and x j The relationship score is obtained through the attention mechanism. The embodiment of the present invention defines a parameter sharing layer, whose parameters are Let the two nodes be x i and x j , the input of this layer is x i and x j , the learned attention parameters Defined as:

[0048]

[0049] Where σ(·) is an activation function, such as a ReLU activation function, and CONCAT represents a concatenation operation. A high relationship score indicates that the two nodes have a strong connection relationship, and these strong connections should be retained. When the node relationship score is relatively low, it means that the relationship between the two nodes is not very strong, so the edge should be deleted. In order to obtain a sparse graph structure and a standardized relationship score that can be compared between nodes of the same type, an embodiment of the present invention uses a sparsemax() function, which returns the Euchildean projection of the input to a probability simplex. It has similar properties to the softmax function, but it can return a sparse probability distribution. Calculate the node x i and x j The standard relationship score of e i,j , the standardized relationship score e i,j ∈R n is defined as follows:

[0050]

[0051] Where τ(·) is the threshold function, and all relationship scores less than the threshold will be set to 0, which is equivalent to deleting the corresponding edge from the fully connected graph. sparsemax(·) represents the sparsemax function, and max represents the maximum value. The standardized relationship score will be used as an edge feature in the graph feature extraction process. In order to reveal the various interactions between two nodes, IAGNN constructs multiple independent relationship learning layers to represent the relationship between nodes in different representation spaces.

[0052] (2) Feature extraction module

[0053] The feature extraction module includes multiple independent graph convolutional neural network modules, multiple independent aggregation modules and a splicing module. The graph convolutional neural network module is used to obtain the node features of each subgraph, the aggregation module is used to obtain the global features of each subgraph, and the splicing module is used to splice the global features of all subgraphs.

[0054] Feature extraction uses graph convolutional neural network (GCN) as the basic module for feature extraction. Graph convolutional neural network is based on the information propagation mechanism on the graph, which can be defined as follows:

[0055]

[0056] in is the node representation of the (l+1)th layer, σ(·) is the activation function, and N(i) is the node x i The set of neighbor nodes of is the node representation of the (l)th layer, W (l) is a learnable parameter, is the standardized term, D i,i For node x i Degree, D j,j For node x j The degree.

[0057] For each group of multi-sequence fragments, the relational learning layer provides the structural information required by the graph convolutional network module, and the relational learning layer obtains the edge weight e i,j Can represent node x j For node x i Therefore, it is necessary to introduce edge weights into graph feature extraction. If there is an edge weight between two nodes, the normalization term can be defined as Since the graph obtained by the relational learning layer is a heterogeneous graph with multiple edge types, the heterogeneous graph can be divided into multiple subgraphs according to the edge type. In order to maintain the independence of each subgraph, each subgraph corresponds to an L-layer GCN module for updating the node representation, and the parameters of the GCN modules corresponding to each subgraph are not shared.

[0058] The GCN module updates the node representation. In order to obtain the representation of the global graph, the readout function is used to aggregate the node features of the last GCN layer. The readout function considers the average characteristics and local maximum characteristics of the sensor network. Therefore, for the kth subgraph, mean pooling and maximum pooling operations are used to obtain the global subgraph representation r k for:

[0059] r k =CONCAT{r mean ,r max}

[0060] The mean pooling r mean And the maximum pooling r max for:

[0061]

[0062] in is the node embedding of the last layer of GCN, and N is the number of nodes.

[0063] In order to maintain the independence of subgraph representations with different edge types, the embedded node only aggregates information from adjacent nodes with the same edge type. Therefore, the subgraph representation cuts off the features in different feature spaces. A simple way to obtain the complete graph representation is to concatenate the features of multiple subgraphs. The final graph representation R is:

[0064] R = CONCAT{r k ∈R 2×n×d}

[0065] The fault identification module is implemented using a fully connected layer, and the last layer is the classification probability output of softmax. The model uses cross entropy loss to train the model, and the loss is defined as follows:

[0066]

[0067] Where W f is the parameter of the model, p i is the sample label, q i is the probability that the sample belongs to fault category i.

[0068] A complex industrial process fault diagnosis system according to an embodiment of the present invention includes:

[0069] The data acquisition and preprocessing module is used to collect the measurement signals of multiple sensors and construct a fully connected graph. Each node in the fully connected graph corresponds to a sensor, and the initial feature of each node is the measurement signal of the corresponding sensor.

[0070] The diagnosis module is used to input the fully connected graph into the diagnosis model. The diagnosis model includes a relationship discovery module, a feature extraction module and a fault identification module. The relationship discovery module is used to identify the interaction relationship between nodes to construct a heterogeneous graph. The heterogeneous graph uses a variety of different types of edges to represent different types of interaction relationships between nodes. The feature extraction module is used to extract feature vectors from the heterogeneous graph. The fault identification module is used to output fault diagnosis results based on the feature vectors.

[0071] The implementation principle and technical effects of the system are similar to the above method and will not be repeated here.

[0072] This embodiment also provides an electronic device, which includes at least one processor and at least one memory, wherein a computer program is stored in the memory. When the computer program is executed by the processor, the processor executes the steps of the complex industrial process fault diagnosis method in the above method embodiment, which will not be repeated here. In this embodiment, the types of processor and memory are not specifically limited. For example, the processor can be a microprocessor, a digital information processor, an on-chip programmable logic system, etc. The memory can be a volatile memory, a non-volatile memory, or a combination thereof, etc.

[0073] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the technical solution of any of the above-mentioned complex industrial process fault diagnosis method embodiments. Its implementation principle and technical effect are similar to those of the above-mentioned method, and will not be repeated here.

[0074] It must be noted that in any of the above embodiments, the method is not necessarily executed in sequence according to the sequence number. As long as it cannot be inferred from the execution logic that it must be executed in a certain order, it means that it can be executed in any other possible order.

[0075] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A complex industrial process fault diagnosis method, characterized in that: Includes steps: Collecting measurement signals of multiple sensors and constructing a fully connected graph, wherein each node in the fully connected graph corresponds to a sensor, and an initial feature of each node is a measurement signal of the corresponding sensor; The fully connected graph is input into a diagnosis model, wherein the diagnosis model includes a relationship discovery module, a feature extraction module and a fault identification module, wherein the relationship discovery module is used to identify the interaction relationship between nodes to construct a heterogeneous graph, wherein the heterogeneous graph uses multiple different types of edges to represent different types of interaction relationships between nodes, the feature extraction module is used to extract a feature vector from the heterogeneous graph, and the fault identification module is used to output a fault diagnosis result according to the feature vector; The relationship discovery module includes multiple independent interaction perception layers based on the attention mechanism, each interaction perception layer is used to identify a type of interaction relationship between nodes; The interaction perception layer is used to implement the following steps: Let the two nodes be x i and x j , the learned attention parameters for Where σ(·) is the activation function, CONCAT represents the concatenation operation, are the parameters of the parameter sharing layer; Compute Node x i and x j The standard relationship score of e i,j Where τ(·) is the threshold function, and all relationship scores less than the threshold will be set to 0. sparsemax(·) represents the sparsemax function, and max means taking the maximum value. The feature extraction module includes multiple independent graph convolutional neural network modules, multiple independent aggregation modules and a splicing module. The graph convolutional neural network module is used to obtain the node features of each subgraph, the aggregation module is used to obtain the global features of each subgraph, and the splicing module is used to splice the global features of all subgraphs.

2. A complex industrial process fault diagnosis method as claimed in claim 1, characterized in that: The construction of the fully connected graph comprises the steps of: The number of sensors is recorded as n, and the measurement signals of n sensors are expressed as s i , where 1≤i≤n, a time window of length m is used to measure the signal s i Divide into multivariate time series segments x i , using n multivariate time series segments x of the same time window i Construct the fully connected graph, which includes n nodes, each of which is x i .

3. A complex industrial process fault diagnosis method as claimed in claim 1, characterized in that: The heterogeneous graph is represented as G = (V, E, R e ), V is a node set, including node v i ∈V, E is the edge set, including edge e i,j ∈E,R e A collection of edge types.

4. A complex industrial process fault diagnosis method as claimed in claim 1, characterized in that: The heterogeneous graph includes multiple subgraphs, and each subgraph represents a type of interaction relationship between nodes.

5. A complex industrial process fault diagnosis system, characterized in that: include: A data acquisition and preprocessing module, used to acquire measurement signals of multiple sensors and construct a fully connected graph, wherein each node in the fully connected graph corresponds to a sensor, and the initial feature of each node is the measurement signal of the corresponding sensor; A diagnosis module, used for inputting the fully connected graph into a diagnosis model, wherein the diagnosis model includes a relationship discovery module, a feature extraction module and a fault identification module, wherein the relationship discovery module is used for identifying the interaction relationship between nodes to construct a heterogeneous graph, wherein the heterogeneous graph uses multiple different types of edges to represent different types of interaction relationships between nodes, wherein the feature extraction module is used for extracting a feature vector from the heterogeneous graph, and wherein the fault identification module is used for outputting a fault diagnosis result according to the feature vector; The relationship discovery module includes multiple independent interaction perception layers based on the attention mechanism, each interaction perception layer is used to identify a type of interaction relationship between nodes; The interaction perception layer is used to implement the following steps: Let the two nodes be x i and x j , the learned attention parameters for Where σ(·) is the activation function, CONCAT represents the concatenation operation, are the parameters of the parameter sharing layer; Compute Node x i and x j The standard relationship score of e i,j Where τ(·) is the threshold function, and all relationship scores less than the threshold will be set to 0. sparsemax(·) represents the sparsemax function, and max means taking the maximum value. The feature extraction module includes multiple independent graph convolutional neural network modules, multiple independent aggregation modules and a splicing module. The graph convolutional neural network module is used to obtain the node features of each subgraph, the aggregation module is used to obtain the global features of each subgraph, and the splicing module is used to splice the global features of all subgraphs.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Citation Information

Patent Citations

  • Terminal operation state prediction method and system based on graph neural network

    CN112364975A

  • Rotating machine fault diagnosis method based on weighted horizontal visible graph

    CN113092044A