Fan blade icing detection method and system based on beta variational graph autoencoder

Abstract

The application discloses a fan blade icing detection method and system based on a Beta variational graph attention autoencoder, belongs to the technical field of fan equipment fault diagnosis, and constructs a weighted adjacent matrix of an attribute graph by using multi-element sensor data of a wind power plant; the weighted adjacent matrix and node embedding features are taken as inputs of a beta-VGATAE model, a GAT layer is used to obtain a mean value and a variance of a Gaussian distribution of a node vector; a reparameterization sampling technology is used to obtain a node vector representation, a weighted adjacent matrix is reconstructed based on the similarity between the node vector representations; the reconstructed weighted adjacent matrix is constrained by using a loss function, a normal distribution of the loss function is obtained by using normal samples, a threshold value is determined, a new distribution of the loss function of unknown samples is counted, and a blade icing state is detected based on the threshold value comparison. The application combines graph machine learning and unsupervised learning technology, realizes fan blade icing detection based on unlabeled sensor data, and keeps the detection precision at a high level.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine equipment fault diagnosis, and more specifically, relates to a wind turbine blade icing detection method and system based on a Beta variational graph attention autoencoder. Background Art

[0002] Vigorously developing wind power generation is a key measure to reduce carbon emissions. Wind turbine capacity is growing rapidly in cold climates; however, wind turbine blades are susceptible to icing. Icing can significantly disrupt aerodynamic properties, leading to turbine downtime, power loss, and damage to components. Therefore, detecting icing on wind turbine blades is of great significance. Existing methods for blade icing detection include physical mechanism methods and data-driven methods. Physical mechanism methods require the installation of additional sensors, making them less cost-effective. Furthermore, they rely heavily on complex expertise. Data-driven methods, which overcome the limitations of physical mechanism methods, have attracted increasing attention. However, existing data-driven methods remain imperfect. These methods fail to consider the contribution of non-Euclidean spatial structure information from sensors to icing detection and rely excessively on large numbers of manually labeled samples. Manually labeled samples are costly and prone to labeling errors. Summary of the Invention

[0003] Traditional data-driven icing detection methods ignore the non-Euclidean spatial structure information between multiple sensors, rely on large amounts of manually labeled sensor data samples, and are unable to effectively process unlabeled sensor data. This paper provides a wind turbine blade icing detection method and system based on a Beta variational graph attention autoencoder. This method utilizes a GAT module to model the spatial correlation of sensor data, thereby extracting the non-Euclidean spatial structure feature information of multiple sensors. The unsupervised learning mechanism of the β-VAE module addresses the technical challenges of wind turbine icing detection using unlabeled sensor data.

[0004] To achieve the above objectives, according to one aspect of the present invention, a method for detecting icing on wind turbine blades based on a Beta variational graph attention autoencoder is provided, comprising:

[0005] Collect multi-sensor data from SCADA system;

[0006] Constructing a weighted adjacency matrix of the attribute graph based on the similarity information between multivariate sensor data;

[0007] The weighted adjacency matrix and node embedding features of the attribute graph are used as input to the Beta Variational Graph Attention Autoencoder β-VGATAE model, and the GAT module is used to obtain the mean and variance of the Gaussian distribution of the node vector. The β-VGATAE model includes the Graph Attention Network GAT module and the Beta Variational Autoencoder β-VAE module.

[0008] The vector representation of the nodes is obtained using the reparameter sampling technique, and the weighted adjacency matrix is ​​reconstructed based on the similarity between the node vector representations.

[0009] The loss function is used to constrain the reconstruction of the weighted adjacency matrix. Normal samples are used to obtain the normal distribution of the loss function and determine the loss probability threshold. In the test phase, the new distribution of the loss function of unknown samples is statistically analyzed, and the blade icing status is detected based on the loss probability threshold comparison.

[0010] In some optional implementations, constructing a weighted adjacency matrix of the attribute graph based on similarity information between multivariate sensor data includes:

[0011] Assume that the multivariate sensor data is {x 1 ,x 2 ,...,x n}, define a weighted undirected graph G = (X, A), and model the sensor data and the correlation between sensor data as the node embedding features X and A of the graph G, respectively, where X = {x 1 ,x 2 ,...,x n}, A∈R n×n , A is the weighted adjacency matrix, any two sensor data x i and x j The mutual information MI correlation between them is expressed as MI(i,j), x i and x j The spatial distance between them is dist(i,j), All elements in the sparse weighted adjacency matrix A are: δ represents the standard deviation of the spatial distance between n sensors; ε represents the threshold, and the average distance between n sensors can be used.

[0012] In some optional implementation schemes, the obtaining the mean and variance of the Gaussian distribution of the node vector using the GAT module includes:

[0013] The encoder of β-VGATAE contains two GAT layers. The mean and variance of the node vector obtained by the GAT layer are: μ = GAT μ (X, A), logσ=GAT σ(X, A), where GAT(·) represents the output function of the GAT layer;

[0014] The posterior probability distribution of the node vector is expressed as: Among them, z i is the embedding vector generated by the β variational autoencoder, and Z is the embedding vector z i The corresponding matrix, μ i and are the mean and variance of the vector of node i, respectively.

[0015] In some optional embodiments, the step of obtaining vector representations of nodes using a reparameter sampling technique and reconstructing a weighted adjacency matrix based on similarities between the vector representations of the nodes comprises:

[0016] Reparameter sampling is performed by z = μ + η·σ, where η ~ N(0,1), so that μ + η·σ ~ N(μ,σ 2 ), z represents the embedding vector after re-parameter sampling;

[0017] Based on the similarity of any two node vector representations, the decoder of β-VGATAE is used to reconstruct the weighted adjacency matrix, where the similarity can reflect the probability of edge connections between nodes, that is:

[0018]

[0019] is the joint probability distribution.

[0020] In some optional implementation schemes, the use of a loss function to constrain the reconstruction of the weighted adjacency matrix includes:

[0021] From L=L r +L KL =E q(Z|X,A) [logp(A|Z)]-βKL[q(Z|X,A)||p(Z)] determines the loss function of the β-VGATAE model, where L r and L KL are the reconstruction loss and divergence loss, respectively; p(A|Z) is the conditional distribution probability; p(Z)~N(0,1); β is the weight parameter of the divergence loss, which is used to adjust the ratio between the reconstruction loss and the divergence loss; I represents the identity matrix, and KL(·) is the Kullback-Leibler divergence operation.

[0022] According to another aspect of the present invention, a wind turbine blade icing detection system based on a Beta variational graph attention autoencoder is provided, comprising:

[0023] Data acquisition module, used to collect multi-sensor data from the SCADA system;

[0024] A weighted adjacency matrix construction module is used to construct a weighted adjacency matrix of the attribute graph based on the similarity information between multivariate sensor data;

[0025] The mean-variance acquisition module is used to take the weighted adjacency matrix and node embedding features of the attribute graph as input to the Beta Variational Graph Attention Autoencoder β-VGATAE model, and use the GAT module to obtain the mean and variance of the Gaussian distribution of the node vector. The β-VGATAE model includes the Graph Attention Network GAT module and the Beta Variational Autoencoder β-VAE module.

[0026] A weighted adjacency matrix reconstruction module is used to obtain the vector representation of the nodes using the reparameter sampling technique and reconstruct the weighted adjacency matrix based on the similarity between the node vector representations;

[0027] The loss constraint module is used to use the loss function to constrain the reconstruction of the weighted adjacency matrix, use normal samples to obtain the normal distribution of the loss function and determine the loss probability threshold, so as to statistically calculate the new distribution of the loss function of unknown samples in the testing phase, and detect the blade icing status based on the loss probability threshold comparison.

[0028] In some optional embodiments, the weighted adjacency matrix building module is used to assume that the multivariate sensor data is {x 1 ,x 2 ,...,x n}, define a weighted undirected graph G = (X, A), and model the sensor data and the correlation between sensor data as the node embedding features X and A of the graph G, respectively, where X = {x 1 ,x 2 ,...,x n}, A∈R n×n , A is the weighted adjacency matrix, any two sensor data x i and x j The mutual information MI correlation between them is expressed as MI(i,j), x i and x j The spatial distance between them is dist(i,j), All elements in the sparse weighted adjacency matrix A are: δ represents the standard deviation of the spatial distance between n sensors; ε represents the threshold, and the average distance between n sensors can be used.

[0029] In some optional implementations, the mean variance acquisition module is used to include two GAT layers in the encoder of β-VGATAE, and the mean and variance of the node vector obtained by the GAT layer are: μ=GAT μ (X, A), logσ=GAT σ (X, A), where GAT(·) represents the output function of the GAT layer; the posterior probability distribution of the node vector is expressed as: Among them, z i is the embedding vector generated by the β variational autoencoder, and Z is the embedding vector z i The corresponding matrix, μ i and are the mean and variance of the vector of node i, respectively.

[0030] In some optional embodiments, the weighted adjacency matrix reconstruction module is used to perform parameter resampling by z=μ+η·σ, where η~N(0,1), so that μ+η·σ~N(μ,σ 2 ), z represents the embedded vector after reparameter sampling; based on the similarity of any two node vector representations, the decoder of β-VGATAE is used to reconstruct the weighted adjacency matrix, where the similarity can reflect the probability of edge connections between nodes, that is: P(A,Z) is the joint probability distribution.

[0031] In some optional embodiments, the loss constraint module is used to calculate L=L r +L KL =E q(Z|X,A) [logp(A|Z)]-βKL[q(Z|X,A)||p(Z)] determines the loss function of the β-VGATAE model, where L r and L KL are the reconstruction loss and divergence loss, respectively; p(A|Z) is the conditional distribution probability; p(Z)~N(0,1); β is the weight parameter of the divergence loss, which is used to adjust the ratio between the reconstruction loss and the divergence loss; I represents the identity matrix, and KL(·) is the Kullback-Leibler divergence operation.

[0032] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0033] The present application combines graph machine learning and unsupervised learning technology, overcomes many defects of traditional fan blade icing detection, for example: ignoring the non-Euclidean space structure information of the sensor, over-reliance on a large number of artificial annotation samples. The beta-VGATAE method proposed in the present application still has a high icing detection accuracy rate using unlabeled sensor data. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a whole architecture diagram of a beta-VGATAE icing detection model provided by an embodiment of the present application;

[0035] Figure 2 is a sparse weighted adjacency matrix result diagram of an attribute graph provided by an embodiment of the present application;

[0036] Figure 3 is a structure diagram of an attribute graph provided by an embodiment of the present application;

[0037] Figure 4 is a weighted adjacency matrix loss distribution diagram of a normal sample provided by an embodiment of the present application;

[0038] Figure 5 is a weighted adjacency matrix loss distribution diagram of an unknown sample provided by an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0040] The present application combines graph machine learning and unsupervised learning technology, overcomes many defects of traditional fan blade icing detection, for example: ignoring the non-Euclidean space structure information of the sensor, over-reliance on a large number of artificial annotation samples. The beta-VGATAE method proposed in the present application still has a high icing detection accuracy rate using unlabeled sensor data.

[0041] (1) Design a Beta-Variational Graph Attention AutoEncoder (Beta-VGATAE) fan blade icing detection model, the architecture diagram of which is as shown in Figure 1 .

[0042] The overall architecture consists of two modules: a Graph Attention Network (GAT) module and a Beta-Variational Auto-Encoder (β-VAE) module. The β-VAE module is used for unsupervised blade ice detection, while the GAT module, as the β-VAE feature extractor, mines the non-Euclidean spatial structure feature information of multiple sensors.

[0043] (2) Collect multi-sensor data from the wind farm's Supervisory Control and Data Acquisition (SCADA) system. To improve data quality, data cleaning, normalization, and feature selection based on manual experience can be used.

[0044] (3) The weighted adjacency matrix of the attribute graph is constructed based on the similarity information between the multi-sensor data. The result is as follows: Figure 2 As shown, the weighted adjacency matrix is ​​sparse, and its determined graph structure is as follows Figure 3 As shown;

[0045] In an embodiment of the present invention, the weighted adjacency matrix of the attribute graph can be constructed in the following manner:

[0046] Construct a weighted adjacency matrix. Assume that the multivariate sensor data is {x 1 ,x 2 ,...,x n}, we can define a weighted undirected graph G = (X, A). Here, the sensor data and the correlation between sensor data are modeled as the node embedding features X and A of the graph G, that is, X = {x 1 ,x 2 ,...,x n}, A∈R n×n , where A is the weighted adjacency matrix. Any two sensor data x i and x j The mutual information (MI) correlation between them is expressed as MI(i,j). The larger the value, the better the x i and x j The greater the degree of correlation in non-Euclidean space. Here we define x i and x j The spatial distance between them is dist(i,j), which can be described as:

[0047]

[0048] All elements in the sparse weighted adjacency matrix A can be described as:

[0049]

[0050] In the formula: delta represents the standard deviation of the spatial distance between n sensors; epsilon represents a threshold value, which can use the average distance between n sensors.

[0051] (4) The weighted adjacency matrix of the attribute graph and the node embedding feature are taken as the input of the beta-VGATAE model, and the GAT layer is used to obtain the mean and variance of the Gaussian distribution of the node vector;

[0052] In the embodiment of the application, the mean and variance of the node vector obtained by GAT can be realized in the following manner:

[0053] GAT obtains the mean and variance of the node vector. GAT is a graph neural network with attention mechanism, and its information expression capability is better than that of the graph convolutional neural network. The encoder of beta-VGATAE contains two GAT layers. The node vector obtained by the GAT layer needs to be sampled in a high-dimensional Gaussian distribution, and the mean and variance thereof can be described as:

[0054] μ=GAT μ (X,A) (3)

[0055] logσ=GAT σ (X,A) (4)

[0056] In the formula: GAT(·) represents the output function of the GAT layer.

[0057] The posterior probability distribution of the node vector can be expressed as:

[0058]

[0059]

[0060] In the formula: z i is the embedding vector generated by the beta variational autoencoder, and Z is the embedding vector z i corresponding matrix; μ i and are the mean and variance of the vector of node i respectively.

[0061] (5) The vector representation of the node is obtained by using the reparameterization sampling technology, and the weighted adjacency matrix is reconstructed based on the similarity between the node vector representations;

[0062] In the embodiment of the application, the vector representation of the node is obtained by using the reparameterization sampling technology, and the weighted adjacency matrix is reconstructed based on the similarity between the node vector representations, which can be realized in the following manner:

[0063] Reparameter sampling technology. Reparameter sampling helps solve the problem of gradient disappearance. Its calculation process is as follows:

[0064] z=μ+η·σ (7)

[0065] In the formula: η~N(0,1), such that μ+η·σ~N(μ,σ 2 ), z represents the embedding vector after reparameter sampling.

[0066] Reconstructing the weighted adjacency matrix. The decoder of β-VGATAE reconstructs the weighted adjacency matrix mainly based on the similarity between any two node vectors. This similarity can reflect the probability of edge connections between nodes, that is:

[0067]

[0068]

[0069] Where: P(A,Z) is the joint probability distribution.

[0070] (6) Use the loss function to constrain the reconstruction of the weighted adjacency matrix. During the training phase, use normal samples to obtain the normal distribution of the loss function. The results are as follows: Figure 4 shown.

[0071] In an embodiment of the present invention, the loss function is used to constrain the reconstruction of the weighted adjacency matrix, which can be achieved in the following way:

[0072] Loss function. The loss function of the β-VGATAE model is:

[0073] L=L r +L KL =E q(Z|X,A) [logp(A|Z)]-βKL[q(Z|X,A)||p(Z)] (10)

[0074]

[0075] Where: L r and L KL are the reconstruction loss and divergence loss respectively; p(A|Z) is the conditional distribution probability; p(Z)~N(0,1); β is the weight parameter of the divergence loss, which is used to adjust the ratio between the reconstruction loss and the divergence loss; KL(·) is the Kullback-Leibler divergence operation; I represents the unit matrix.

[0076] according to Figure 4 It can be seen that the loss distribution of the normal data set is concentrated in the probability interval [0, 0.5], so the loss probability threshold can be set to 0.5. In the test phase, the new distribution of the loss function of the unknown sample is statistically analyzed, and the results are as follows Figure 5 As shown. Figure 5 As can be seen, the loss distribution of the unknown dataset is concentrated in the probability interval [1.45, 2.05], and the vast majority of ice samples exceed the loss probability threshold of normal samples. Therefore, unsupervised wind turbine blade ice detection can be achieved by comparing the loss probability threshold of the normal dataset.

[0077] This paper presents a wind turbine blade icing detection method based on a Beta Variational Graph Attention Autoencoder. By using GAT to extract spatial features from sensor data, the icing detection model can consider the non-Euclidean spatial structure between multiple sensors, thereby improving its performance. Furthermore, the proposed model, based on the β-VAE architecture, can address the problem of unsupervised learning with unlabeled sensor data.

[0078] In another embodiment of the present invention, a wind turbine blade icing detection system based on a Beta variational graph attention autoencoder is provided, comprising:

[0079] Data acquisition module, used to collect multi-sensor data from the SCADA system;

[0080] A weighted adjacency matrix construction module is used to construct a weighted adjacency matrix of the attribute graph based on the similarity information between multivariate sensor data;

[0081] The mean-variance acquisition module is used to take the weighted adjacency matrix and node embedding features of the attribute graph as input to the Beta Variational Graph Attention Autoencoder β-VGATAE model, and use the GAT module to obtain the mean and variance of the Gaussian distribution of the node vector. The β-VGATAE model includes the Graph Attention Network GAT module and the Beta Variational Autoencoder β-VAE module.

[0082] A weighted adjacency matrix reconstruction module is used to obtain the vector representation of the nodes using the reparameter sampling technique and reconstruct the weighted adjacency matrix based on the similarity between the node vector representations;

[0083] The loss constraint module is used to use the loss function to constrain the reconstruction of the weighted adjacency matrix, use normal samples to obtain the normal distribution of the loss function and determine the loss probability threshold, so as to statistically calculate the new distribution of the loss function of unknown samples in the testing phase, and detect the blade icing status based on the loss probability threshold comparison.

[0084] The specific implementation of each module can refer to the description of the above method embodiment, and will not be repeated in this embodiment.

[0085] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0086] It will be easily understood by those skilled in the art that the above description is merely 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 scope of protection of the present invention.

Claims

1. A wind turbine blade icing detection method based on Beta variational graph attention autoencoder, characterized in that: include: Collect multi-sensor data from SCADA system; Constructing a weighted adjacency matrix of the attribute graph based on the similarity information between multivariate sensor data; The weighted adjacency matrix and node embedding features of the attribute graph are used as Beta variational graph attention autoencoders β -VGATAE model input, and use the GAT module to obtain the mean and variance of the node vector Gaussian distribution, where, β -The VGATAE model includes the graph attention network GAT module and Beta variational autoencoder β -VAE module; The vector representation of the nodes is obtained using the reparameter sampling technique, and the weighted adjacency matrix is ​​reconstructed based on the similarity between the node vector representations. The loss function is used to constrain the reconstruction of the weighted adjacency matrix. The normal distribution of the loss function is obtained using normal samples and the loss probability threshold is determined. In the test phase, the new distribution of the loss function of unknown samples is statistically analyzed. The blade icing status is detected based on the loss probability threshold comparison. The method of constructing a weighted adjacency matrix of the attribute graph based on similarity information between multivariate sensor data includes: Assume that the multivariate sensor data is { x 1 , x 2 ,..., x n }, define a weighted undirected graph G = ( X , A ), the sensor data and the correlation between sensor data are modeled as graphs G Node embedding features X and A ,in, X = { x 1 , x 2 ,..., x n }, A ∈ R n×n , A is the weighted adjacency matrix, any two sensor data x i and x j The mutual information MI correlation between MI ( i , j ), x i and x j The spatial distance between dist ( i , j ), , a sparse weighted adjacency matrix A All the elements in are: , δ express n The standard deviation of the spatial distance between sensors; ε Indicates the threshold value, using n The average distance between sensors.

2. The method according to claim 1, characterized in that The method of using the GAT module to obtain the mean and variance of the node vector Gaussian distribution includes: exist β -VGATAE's encoder contains two GAT layers. The mean and variance of the node vectors obtained using the GAT layers are: , ,in, GAT (·) represents the output function of the GAT layer; The posterior probability distribution of the node vector is expressed as: , ,in, z i for β The embedding vector generated by the variational autoencoder, Z is the embedding vector z i The corresponding matrix, μ i and Node i The mean and variance of a vector.

3. The method according to claim 2, characterized in that The method of obtaining vector representations of nodes by using the reparameter sampling technique and reconstructing a weighted adjacency matrix based on similarities between the vector representations of nodes includes: Depend on Perform reparameter sampling, where ~ N (0, 1), so that ~ N ( μ , σ 2 ), z represents the embedding vector after reparameter sampling; Similarity utilization based on vector representation of any two nodes β -VGATAE’s decoder reconstructs a weighted adjacency matrix, where similarity reflects the probability of an edge connection between nodes, i.e.: , , P ( A , Z ) is the joint probability distribution.

4. The method according to claim 3, characterized in that The method of utilizing a loss function to constrain the reconstruction of a weighted adjacency matrix includes: Depend on Sure β -The loss function of the VGATAE model, where , L r and L KL They are reconstruction loss and divergence loss respectively; is the conditional distribution probability; p ( Z ) ~ N (0, 1); β is the weight parameter of the divergence loss, which is used to adjust the ratio between the reconstruction loss and the divergence loss; I represents the identity matrix, KL (·) is the Kullback-Leibler divergence operation.

5. A wind turbine blade icing detection system based on Beta variational graph attention autoencoder, characterized in that: include: Data acquisition module, used to collect multi-sensor data from the SCADA system; A weighted adjacency matrix construction module is used to construct a weighted adjacency matrix of the attribute graph based on the similarity information between multivariate sensor data; The mean-variance acquisition module is used to use the weighted adjacency matrix and node embedding features of the attribute graph as Beta variational graph attention autoencoders β -VGATAE model input, and use the GAT module to obtain the mean and variance of the node vector Gaussian distribution, where, β -The VGATAE model includes the graph attention network GAT module and Beta variational autoencoder β -VAE module; A weighted adjacency matrix reconstruction module is used to obtain the vector representation of the nodes using the reparameter sampling technique and reconstruct the weighted adjacency matrix based on the similarity between the node vector representations; The loss constraint module is used to constrain the reconstruction of the weighted adjacency matrix using the loss function. It uses normal samples to obtain the normal distribution of the loss function and determine the loss probability threshold. This module then calculates the new distribution of the loss function for unknown samples during the test phase and detects the blade icing status based on the loss probability threshold comparison. The weighted adjacency matrix building module is used to assume that the multivariate sensor data is { x 1 , x 2 ,..., x n }, define a weighted undirected graph G = ( X , A ), the sensor data and the correlation between sensor data are modeled as graphs G Node embedding features X and A ,in, X = { x 1 , x 2 ,..., x n }, A ∈ R n×n , A is the weighted adjacency matrix, any two sensor data x i and x j The mutual information MI correlation between MI ( i , j ), x i and x j The spatial distance between dist ( i , j ), , a sparse weighted adjacency matrix A All the elements in are: , δ express n The standard deviation of the spatial distance between sensors; ε Indicates the threshold value, using n The average distance between sensors.

6. The system according to claim 5, characterized in that The mean variance acquisition module is used to β -VGATAE's encoder contains two GAT layers. The mean and variance of the node vectors obtained using the GAT layers are: , ,in, GAT (·) represents the output function of the GAT layer; the posterior probability distribution of the node vector is expressed as: , ,in, z i for β The embedding vector generated by the variational autoencoder, Z is the embedding vector z i The corresponding matrix, μ i and Node i Mean and variance of a vector.

7. The system according to claim 6, characterized in that The weighted adjacency matrix reconstruction module is used to Perform reparameter sampling, where ~ N (0, 1), so that ~ N ( μ , σ 2 ), z Represents the embedded vector after re-parameter sampling; based on the similarity of any two node vector representations β -VGATAE’s decoder reconstructs a weighted adjacency matrix, where similarity reflects the probability of an edge connection between nodes, i.e.: , , P ( A , Z ) is the joint probability distribution.

8. The system according to claim 7, characterized in that The loss constraint module is used to Sure β -The loss function of the VGATAE model, where , L r and L KL They are reconstruction loss and divergence loss respectively; is the conditional distribution probability; p ( Z ) ~ N (0, 1); β is the weight parameter of the divergence loss, which is used to adjust the ratio between the reconstruction loss and the divergence loss; I represents the identity matrix, KL (·) is the Kullback-Leibler divergence operation.

Images