Method and device for learning optical network failure prediction by comparing unwrapping hock chart

Abstract

The application provides a contrast disentangled Hodge plot learning optical network fault prediction method, first, optical network fault data is preprocessed, and the optical network fault data is converted into time series data, the optical network fault data includes port fault, board fault, network break, power failure, fiber break fault, network node device information and time node information. Then, a contrast disentangled Hodge plot learning model is constructed, the model includes an embedding module, a fault representation learning module, a self-supervised contrast learning module and a prediction module. The time series data is input into the contrast disentangled Hodge plot learning model to predict the fault in the optical network fault data. The optical network fault prediction method of the application is constructed by constructing the contrast disentangled Hodge plot learning model, thereby providing a technical solution for accurately predicting the optical network fault, and improving the accuracy of network fault prediction. The application also provides a corresponding contrast disentangled Hodge plot learning optical network fault prediction device.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical network fault prediction, and more particularly relates to a method and device for learning optical network fault prediction based on a contrast learning idea. BACKGROUND

[0002] The communication industry in China is developing rapidly, and optical fiber communication based on optical cables has the advantages of light weight, small size and wide frequency band, and has been increasingly important in modern communication life. In the process of optical fiber communication based on optical cables, the optical cable detection system is realized by real-time detection and alarm of optical power, and a large amount of fault alarm data is generated in the communication process. These information is stored in the database. The massive historical data store a large amount of valuable information, and effective mining of these data can find an efficient model of optical cable fault. How to understand the optical power information in the database to predict the future behavior of optical fiber, and how to detect, handle and locate the fault are important problems that need to be further studied in optical fiber communication.

[0003] The occurrence of optical network fault can be regarded as a time-based fault sequence according to time. The traditional statistical-based method arranges workers on different routes in the same paragraph for manual monitoring, which cannot meet the requirements of current fault judgment. In recent years, with the development of deep learning technology, using neural network technology for optical network fault prediction has become a trend. However, the current optical network fault prediction based on deep learning technology has the following difficulties: most methods cannot fully consider the complex interaction relationship behind the influence of optical network fault on nodes. In order to accurately predict the next fault influence node, it is more important to mine which factors and features in fault warning and fault influence reason. SUMMARY

[0004] The purpose of the present application is to provide an optical network fault prediction method to improve the warning ability of optical fiber fault mode. In order to solve the above problems, a method for learning optical network fault prediction based on contrast learning idea is provided.

[0005] To achieve the above purpose, according to one aspect of the present application, a method for learning optical network fault prediction based on contrast unwinding Hock chart is provided, comprising the following steps:

[0006] S1. Preprocessing the optical network fault data to convert it into time series data, each time series data representing all fault types occurring over time of the current device;

[0007] S2. Construct a comparative disentangled Hodge chart learning model, input the time series data into the comparative disentangled Hodge chart learning model, and the comparative disentangled Hodge chart learning model realizes prediction of faults in optical network fault data; the comparative disentangled Hodge chart learning model comprises an embedding module, a fault representation learning module, a self-supervised contrast learning module and a prediction module, wherein the embedding module is used for realizing conversion of input time series into dense vector representation with the fault time series of the optical network device as input; the fault representation learning module is used for mining complex relationships between faults based on the disentangled graph convolution network by using the input vector representation; the self-supervised contrast learning module is used for self-supervised learning by constructing multiple views of nodes, reducing interference of noise on the learning process and accelerating the training process; and the prediction module is used for realizing accurate prediction of optical network faults.

[0008] In one embodiment of the present application, the optical network fault data is preprocessed and converted into time series data, specifically: the collected network fault data is preprocessed, the time node of fault occurrence and the time series data reference are converted into time series samples, and all fault type data sequences on the time series are obtained according to the time stamp.

[0009] In one embodiment of the present application, the optical network fault data comprises fault types, network node device information and time node information; the fault types comprise port fault, board fault, network break, power failure and fiber break fault.

[0010] In one embodiment of the present application, the embedding module in the step S2 takes the fault time series of the optical network device as input, converts the input fault time series into an embedding vector, and specifically realizes:

[0011] When the fault time series is input, each fault time is represented as a one-hot vector, in the embedding module, an embedding matrix O is used to convert the one-hot vector into a low-dimensional dense vector representation, wherein d is the dimension of embedding, and ψ is the length of the time series; a graph G is constructed d×ψ is used to convert a one-hot vector into a low-dimensional dense vector representation, wherein d is the dimension of embedding, and ψ is the length of the time series; a graph G is constructed s (V s ,E s ) is a graph with a set of fault influence nodes V s and a set of edges E s , in the graph G s , each fault influence node represents a different fault type in the time series data, and each directed edge from v i to v i+1 is a fault change interaction v t →v i in E i+1 ; in the graph that has been constructed, each directed edge from node v ito v i+1 a directed edge E t represents a fault change interaction transition, where v i and v i+1 both represent fault influence nodes belonging to a node set V s E t both represent an edge belonging to an edge set E s are ordered by their time in a time series so as to distinguish their fault characteristics at different times.

[0012] In an embodiment of the present application, the fault representation learning module in step S2 uses the input vector representation to mine complex relationships between faults based on the unwound graph convolutional network, specifically including:

[0013] First, represents an input fault sequence (i.e. a vector representation of an input fault time series), where sn is the number of different faults in Sv; for each different fault Each fault is represented by an embedding R d represents an embedding space, and the fault representation is The embedding of each fault is converted into k blocks, each of which represents a potential cause behind a complex relationship:

[0014]

[0015] where c i,k represents the embedding representation of the kth cause of the fault influence node i, W k and b k are parameters of the kth cause, ω is a nonlinear activation function, and || ||2 represents the use of L2 regularization to avoid model overfitting; when not yet input, at the 0th layer of graph convolution, the initial embedding of the fault sequence is: where

[0016] The basic idea of iterative graph convolution is then performed by propagating and aggregating on the graph to learn the representation of the block:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] denotes the final t-th layer output factor embedding representation, denotes the i-th row factor embedding representation of the t-th layer, and the average embedding representation of the k-th factor finally output by t-1 graph convolution layers is and denote the update gate and the reset gate, respectively, and denote the in-degree similarity matrix between the fault influence node and the adjacent fault influence node and the out-degree similarity matrix of the i-th row, concat() denotes the aggregation operation in the graph convolution, H in , H out , U z , U r and U o are all weight matrices in the learning propagation process, b in , b out are constraint terms to prevent model overfitting, σ is a sigmoid activation function, and ⊙ denotes an element-wise multiplication operation.

[0023] In an embodiment of the present application, in order to solve the over-smoothing problem, the embedding of the fault influence node is updated by aggregating the embedding learned from the adjacent fault influence node and the original embedding with a specified attention weight based on an l-layer attention network, and the embedding of the fault influence node is obtained by re-connecting all the learned embeddings of the T factors of the l-layer graph convolution layer, that is, denotes, as follows:

[0024]

[0025]

[0026] where w f is a weight vector, W q is a weight matrix, β l is the attention weight of the l-th layer in the attention network, and the time information in the sequence is captured by using a GRU to capture the time dependence in the above fault influence representation, and the time representation at the τ-th day is

[0027] h τ = GRU (e τ , h τ-1 ), t-T≤τ≤t-1

[0028] where eτdenotes the fault influence node embedding representation at the τ-th day

[0029] A HoG attention mechanism φ() is used to learn the attention weights ρ τ Hidden representations from different days are combined As follows

[0030]

[0031] Final aggregate per-day latent representations γ τ Wρ is a learned linear transformation function, and the final temporal representation is:

[0032]

[0033] η is an excitation parameter, μ is an attenuation parameter, and Δt is the time difference between the current and the τth day in the past;

[0034] The above temporal features are fused into the embedded fault impact node representation, and the final fault impact node update representation is as follows:

[0035]

[0036] Aggregate embedded representations:

[0037]

[0038] δ i，k = θ T σ(W1c n，k +W2c i，k +b)

[0039] χ g is the fault time series embedding representation, δ i，k is the attention weight of the fault impact node i on the kth factor, W1, W2, W t is a weight matrix, and b is a constraint term used to prevent overfitting.

[0040] In one embodiment of the present application, the self-supervised contrast learning module in step S2 performs self-supervised learning by constructing multiple views of nodes, specifically including:

[0041] Two strategy operators are designed based on the above graph structure, namely node dropping s1 and edge dropping s2:

[0042] s1(G) = (M' ⊙ V s , E s ), s2(G) = (M'' ⊙ V s , E s ),

[0043] s1(G) = (V s , M1 ⊙ E s ), s2(G) = (V s, M2⊙E s ),

[0044] where, are two masking vectors that are applied to the set of nodes V s , construct multiple self-views of the fault-affected nodes, so different node l-layer self-views are obtained based on the above 2 operators and

[0045]

[0046] Z′ 1，u , Z″ 2，u are different self-views constructed by node u based on s1 and s2 operators, Z″ 2，v is the self-view of node v constructed based on the s2 operator, the self-views of the same node are regarded as positive pairs (i.e. {(Z′ 1，u , Z″ 2，u )|u∈U} and the self-views of any different nodes are regarded as negative pairs (i.e. {(Z′ 1，u , Z″ 2，v )|u, v∈U, u≠v}), and Info-NCE is used as the learning goal, which has a standard binary cross-entropy loss between samples from the ground truth (positive) and contaminated samples (negative), and the self-supervised contrastive learning objective function is as follows:

[0047]

[0048] where s() represents the similarity between the views constructed based on different nodes and different operators, and τ is a hyperparameter.

[0049] In an embodiment of the present application, the prediction module in step S2 realizes accurate prediction of optical network faults, and specifically includes:

[0050] In the case of a given sequence, the score Γ i of all candidate nodes i∈I is calculated by performing an inner product between the node embedding e g learned from the above and the sequence embedding χ i :

[0051] Γ i = χ g T e i

[0052] The softmax function is used to predict the probability of a node becoming the next fault:

[0053]

[0054] In one embodiment of the present application, a cross-entropy loss function is used to learn and optimize the failure impact node prediction objective function:

[0055]

[0056] where y is the ground truth hot encoding vector, the L2 regularization term is omitted and L is minimized using Adam r ;

[0057] The final learning objective function is defined as:

[0058]

[0059] where is a variable factor that controls the self-supervised contrastive learning task.

[0060] According to another aspect of the present application, a contrastive unwrapped Hough graph learning optical network failure prediction device is also provided, comprising at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions executable by the at least one processor, and the instructions are used to complete the contrastive unwrapped Hough graph learning optical network failure prediction method after being executed by the processor.

[0061] Overall, compared with the prior art, the above technical solutions conceived by the present application have the following beneficial effects:

[0062] (1) By integrating decoupled representation learning into the gated graph neural network, the failure representation caused by multiple independent latent causes is represented. By representing the failure with separate independent latent causes, the main purpose of mining in the target failure representation is better captured, ensuring the effectiveness and accuracy of the prediction;

[0063] (2) By using the method of multi-strategy self-contrastive learning based on failure impact nodes, the influence of noisy current values on optical power values is alleviated, and the robustness of the prediction model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is a flowchart of a contrastive unwrapped Hough graph learning optical network failure prediction method according to the present application. DETAILED DESCRIPTION

[0065] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be given below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and not to 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.

[0066] As shown in Figure 1 The present application provides a comparative disentangled Hodge plot learning optical network fault prediction method, comprising the following steps:

[0067] S1. Preprocessing the optical network fault data to convert it into time series data, each time series data representing all fault types occurring over time of the current device;

[0068] Specifically, the collected network fault data is preprocessed, the time node of fault occurrence and the time series data reference are converted into time series samples, and all fault type data sequences on the time series are obtained according to the time stamp.

[0069] Specifically, the optical network fault data includes fault type, network node device information and time node information; the fault type includes but is not limited to port fault, board fault, network break, power failure, fiber break fault;

[0070] S2. Constructing a comparative disentangled Hodge plot learning model, inputting the time series data into the comparative disentangled Hodge plot learning model, and realizing the prediction of the fault in the optical network fault data; the comparative disentangled Hodge plot learning model includes an embedding module, a fault representation learning module, a self-supervised contrast learning module and a prediction module, wherein the embedding module is used to input the fault time series of the optical network device to realize the conversion of the input time series into dense vector representation; the fault representation learning module is used to mine the complex relationship between faults based on the disentangled graph convolution network using the input vector representation; the self-supervised contrast learning module is used to perform self-supervised learning by constructing multiple views of nodes to reduce the interference of noise on the learning process and speed up the training process; the prediction module is used to realize the accurate prediction of the optical network fault. Specifically comprising:

[0071] S21. Constructing an embedding module to input the fault time series of the optical network device to convert the input fault time series into embedding vectors. Specifically,

[0072] When the fault time series is input, each fault time is represented as a one-hot vector, and in the embedding module, the embedding matrix O ∈ R d×ψTransform a one-hot vector into a low-dimensional dense vector representation, where d is the embedding dimension and ψ is the length of the time series; construct G. s (V s E s V is the set of nodes affected by the fault. s The set of edges E s The image in G s In the diagram, each fault-affected node represents a different fault type in the time series data, and each data point from v... i to v i+1 The directed edge is E t Fault Change Interaction Transformation v i →v i+1 In the already constructed graph, each node v i to v i+1 The directed edge E t This represents the interactive transition of fault changes, where v i and v i+1 All of these indicate that they belong to the node set V. s The fault affects the node, E t All statements belong to the edge set E s An edge in the time series; sort them by their time in the time series in order to distinguish the fault characteristics they have at different times.

[0073] S22. Constructing a fault representation learning module: This module utilizes an unwrapped graph convolutional network to uncover complex relationships between faults. The unwrapping method is employed to obtain fine-grained information about various underlying causes of faults, and a spatiotemporal Hawke process model is used to acquire the correlations between fault occurrences among optical network device nodes. The specific process is as follows:

[0074] first, Let s represent the input fault sequence (i.e., the vector representation of the input fault time series), where s n It is S v The number of different faults; for each different fault Each fault By embedding It means that R d Represents the embedded space, and the fault is represented as Embedding of each fault It is transformed into k blocks, each block representing a potential reason behind the complex relationship:

[0075]

[0076] Among them, c i,k W represents the embedding representation of the cause of the failure affecting node i. k and b kis the parameter of the kth cause, ω is a nonlinear activation function, || ||2 represents the L2 regularization method to avoid model overfitting, when has not been input, at the 0th layer of graph convolution, the initial embedding of the fault sequence is: where

[0077] Then the basic idea of iterative graph convolution is performed to learn the representation of the block by propagation and aggregation on the graph:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] represents the final tth layer output factor embedding representation, represents the ith row of the tth layer factor embedding representation, and the average embedding representation of the kth factor finally output by t-1 graph convolution layers is and represent the update gate and the reset gate, respectively, and represent the in-degree similarity matrix and the out-degree similarity matrix between the fault influence node and the adjacent fault influence node, respectively, concat() represents the aggregation operation in the graph convolution, H in , H out , U z , U r , and U o are all learned weight matrices in the propagation process, b in , b out is a constraint term to prevent model overfitting, σ is a sigmoid activation function, and ⊙ represents an element-wise multiplication operation.

[0084] In order to solve the over-smoothing problem, the embedding of the fault influence node is updated by aggregating the embedding learned from the adjacent fault influence nodes based on the l-layer attention network and the original embedding with specified attention weights, and the embedding of the fault influence node is obtained by reconnecting all the learned embeddings of the T factors of the l-layer graph convolution, i.e. represents, as follows:

[0085]

[0086]

[0087] where w f is the weight vector, W q is the weight matrix, β l is the attention weight of the l-th layer in the attention network, the temporal information in the sequence is captured by GRU to capture the temporal dependency in the above failure impact representation, and the time representation at day τ is

[0088] h τ = GRU(e τ , h τ-1 ), t - T ≤ τ ≤ t - 1

[0089] where e τ represents the failure impact node embedding representation at day τ

[0090] An attention mechanism φ() is used to learn the attention weight ρ τ The time hidden representation from different days is as follows

[0091]

[0092] The final aggregated per-day latent representation γ τ is learned by a linear transformation function Wρ, and the final time representation is:

[0093]

[0094] η is the excitation parameter, μ is the decay parameter, and Δt is the time difference between the current and the past day τ.

[0095] The above temporal features are fused into the embedding failure impact node representation, and the final failure impact node update representation is as follows:

[0096]

[0097] The aggregated embedding representation is:

[0098]

[0099] δ i，k = θ T σ(W1c n，k +W2c i，k +b)

[0100] χ g is the failure time series embedding representation, δ i，k is the attention weight of the failure impact node i on the k-th factor, W1, W2, Wt is a weight matrix, and b is a constraint term to prevent overfitting.

[0101] S23. Constructing a self-supervised contrastive learning module based on the fault-affected node representation.

[0102] Two strategy operators are designed based on the above graph structure, namely node dropping s1 and edge dropping s2:

[0103] s1(G)=(M′⊙V s ,E s ),s2(G)=(M″⊙V s ,E s ),

[0104] s1(G)=(V s ,M1⊙E s ),s2(G)=(V s ,M2⊙E s ),

[0105] wherein, are two mask vectors, which are applied to the node set V s , to construct multiple self-views of the fault-affected nodes, and different node l-layer self-views and

[0106]

[0107] Z′ 1，u , Z″ 2，u are different self-views of node u constructed based on the s1 and s2 operators, and Z″ 2，v is a self-view of node v constructed based on the s2 operator, the self-views of the same node are regarded as positive pairs (i.e. {(Z′ 1，u , Z″ 2，u )|u∈U} and the self-views of any different nodes are regarded as negative pairs (i.e. {(Z′ 1,u ,Z″ 2,v )|u,v∈U,u≠v}, and Info-NCE is adopted as the learning goal, which has a standard binary cross-entropy loss between samples from the ground truth (positive) and contaminated samples (negative), and the self-supervised contrastive learning objective function is as follows:

[0108]

[0109] wherein s() represents the similarity between views constructed based on different nodes and different operators, and τ is a hyperparameter.

[0110] S24. Construct a prediction module that computes the score Γ i for all candidate nodes i ∈ I by taking the inner product between the sequence embedding χ g and the node embedding e i learned from above:

[0111] Γ i = χ g T e i

[0112] The softmax function is adopted to predict the probability of a node becoming the next failure:

[0113]

[0114] The cross-entropy loss function is adopted to learn and optimize the failure influence node prediction objective function:

[0115]

[0116] where y is the positive (ground truth) hot encoding vector, the L2 regularization term is omitted and L is minimized using Adam r ,

[0117] The final learning objective function is defined as:

[0118]

[0119] where is a variable factor that controls the self-supervised contrastive learning task.

[0120] Further, the present application also provides a contrastive unwrapping Hock chart learning optical network failure prediction device, comprising at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions executable by the at least one processor, and the instructions are used to complete the contrastive unwrapping Hock chart learning optical network failure prediction method after being executed by the processor.

[0121] Those skilled in the art will easily understand that the above description is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for fault prediction in optical networks based on comparative untangling Hawke's diagram learning, characterized in that, Comprising the following steps: S1. Preprocess the optical network fault data to convert it into time series data, each of which represents all fault types occurring over time for the current device; the optical network fault data includes fault type, network node device information, and time node information; the fault type includes port failure, board failure, network break, power failure, and fiber break failure; S2. Construct a comparative unwinding Hodge chart learning model, input the time series data into the comparative unwinding Hodge chart learning model, and the comparative unwinding Hodge chart learning model realizes the prediction of the faults in the optical network fault data; the comparative unwinding Hodge chart learning model includes an embedding module, a fault representation learning module, a self-supervised contrast learning module, and a prediction module, wherein the embedding module is configured to input the fault time series of the optical network device to realize the conversion of the input time series into a dense vector representation; the fault representation learning module is configured to mine the complex relationship between faults based on the unwinding Hodge chart convolution network using the input vector representation; the self-supervised contrast learning module is configured to perform self-supervised learning by constructing multiple views of nodes to reduce the interference of noise on the learning process and speed up the training process; and the prediction module is configured to realize accurate prediction of optical network faults; The fault representation learning module in step S2 mines the complex relationship between faults based on the unwinding Hodge chart convolution network using the input vector representation, specifically comprising: First, represents the input failure sequence, i.e. the vector representation of the input failure time sequence, where is the number of different failures in for each different failure is represented by an embedding , represents the embedding space, where the failure is represented as the embedding of each failure is converted to blocks, each block representing one potential cause behind the complex relationship: in, Indicates the nodes affected by the fault No. Embedded representation of each reason, and It is the first The parameters for each reason, It is a non-linear activation function. This indicates that L2 regularization is used to avoid model overfitting. When there is no input yet and the model is at layer 0 of the graph convolution, the initial embedding of the fault sequence is: ,in , ; Then the basic idea of graph convolution is iteratively executed to learn the representation of the block by propagation and aggregation on the graph: represents the final factor embedding representation output by the layer, represents the row factor embedding representation of the layer, the final output of the , and represent the update gate and the reset gate, respectively, and represent the first row of the in-degree similarity matrix and the out-degree similarity matrix between the failure-affected node and its adjacent failure-affected nodes, respectively, represents the aggregation operation in the graph convolution, , , , and are the weight matrices in the propagation process of learning, , are the constraint terms to prevent the model from overfitting, is the sigmoid activation function, represents the element-wise multiplication operation.

2. The method of claim 1, wherein the learning optical network failure prediction method is based on a comparative disentanglement Hock chart. The preprocessing of the optical network fault data to convert it into time series data is specifically: preprocessing the collected network fault data, converting the time node of fault occurrence and the time series data reference into time series samples, and obtaining all fault type data sequences on the time series according to the timestamp. ​ 3. The method of claim 1 or 2, wherein the learning optical network failure prediction method is based on a comparative disentanglement Hock chart. The embedding module in step S2 inputs the fault time series of the optical network device to convert the input fault time series into an embedding vector, which is specifically implemented as: ​ When the fault time series is input, each fault time is represented as a one-hot vector. In the embedding module, this is processed by the embedding matrix. Transform a one-hot vector into a low-dimensional dense vector representation, where It is the dimension of the embedding. It is the length of the time series; construction For the set of nodes affected by the fault Sum of edges The image, in In the diagram, each fault-affected node represents a different fault type in the time series data, and each data point from... arrive The directed edge is Interactive conversion of fault changes in ; In the already constructed graph, each directed edge from a node to a node represents a failure change interaction transition, where and both represent failure impact nodes belonging to the set of nodes , and represents an edge belonging to the set of edges ; ordered by their time in the time series so as to distinguish their failure characteristics at different times.

4. The method of claim 1, wherein the learning optical network failure prediction method is based on a comparative disentanglement Hock chart. To address the oversmoothing problem, the embedding of a failure-affected node is updated by aggregating the embeddings learned from neighboring failure-affected nodes and their original embeddings with specified attention weights, and the embedding of a failure-affected node is updated by layer reconnection all factors of the learned embeddings of the graph convolutional layers to obtain the embedding of the failure-affected node, i.e. is represented as follows: ​ wherein, is a weight vector, is a weight matrix, is an attention weight of the layer in the attention network, the temporal information in the sequence captures the temporal dependency in the above failure impact representation using a GRU at the time representation wherein, represents the day of the week fault impact node embedding represents ; Employing a hock attention mechanism Learning attention weights Concatenating temporal hidden representations from different days As follows Final aggregated per day potential representation , is a learned linear transformation function, final temporal representation: For excitation parameters, For attenuation parameters, It is the present and the past. Time difference; The above time features are fused into the embedding fault influence node representation, and the final fault influence node update representation is as follows: Aggregate the embedding representation: embedding representation for the failure time series, is a failure-affected node attention weight on the th factor, , , is a weight matrix, is a constraint term for preventing overfitting.

5. The method of claim 1 or 2, wherein the learning optical network failure prediction method is based on a comparative disentanglement Hock chart. The self-supervised contrast learning module in step S2 performs self-supervised learning by constructing multiple views of nodes, specifically comprising: ​ Two strategy operators are designed based on the above graph structure, namely node dropping and edge dropping : wherein, are two masking vectors which are applied to the set of nodes , constructing multiple self-views of the fault-affected nodes, thus obtaining different node layer self-views based on the above 2 operators respectively and : is a node based on and different self-views constructed by operators, is a node based on self-views constructed by operators, considering the self-view of the same node as positive pair, i.e. considering the self-view of any different node as negative pair, i.e. adopting Info-NCE as the learning objective, which has standard binary cross-entropy loss between samples from positive and contaminated negative, the self-supervised contrastive learning objective function is as follows: wherein denotes the similarity between views constructed based on different nodes and different operators, is a hyperparameter.

6. The method of claim 1 or 2, wherein the learning optical network failure prediction method is based on a comparative disentanglement Hock chart. The prediction module in step S2 realizes accurate prediction of optical network faults, specifically comprising: In the case of a given sequence, the scores of all candidate nodes are computed by taking the inner product between the node embeddings learned from above and the sequence embedding :​​ The softmax function is used to predict the probability of a node becoming the next fault: 。 7. The method of claim 6, wherein the learning optical network failure prediction method is based on a comparative disentangled Hough diagram. 7 The cross-entropy loss function is used to learn and optimize the fault influence node prediction objective function: wherein is the ground truth hot encoding vector, omitting the L2 regularization term and employing Adam minimization ; The final learning objective function is defined as: wherein is a variable factor controlling the self-supervised contrastive learning task.

8. A comparative unwinding Hodge chart learning optical network fault prediction device, characterized in that: It comprises at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions executable by the at least one processor, and the instructions are used to complete the comparative unwinding Hodge chart learning optical network fault prediction method in any one of claims 1-7 after being executed by the processor.

Images