Network fault prediction method and system based on fully heterogeneous meta-contrastive learning model

Abstract

This invention discloses a network fault prediction method and system based on a fully heterogeneous meta-contrastive learning model, comprising: S100: converting network fault data into time series data, wherein the network fault data includes at least the fault type, fault occurrence time, and fault occurrence location; S200: constructing a network fault prediction model, which includes an embedding module, a heterogeneous meta-path network, a self-supervised contrastive learning module, and a prediction module. The embedding module is used to receive time series data and convert it into an embedding vector; the heterogeneous meta-path network is used to integrate fault node information; the self-supervised contrastive learning module is used to compare and learn the fault network views of the fault nodes; and the prediction module is used to perform fault prediction using a loss function. This invention can improve the accuracy of network fault prediction.

Description

Technical Field

[0001] The present application belongs to the technical field of network fault prediction, and specifically relates to a network fault prediction method and system based on a fully heterogeneous meta-contrastive learning model. Background Art

[0002] The proliferation of electronic devices and the development of various software features have made our lives more convenient, but new types of network failures have also emerged. Efficiently resolving network failures is a widespread concern for major operators. These failures include service interruptions, low network speeds, and network noise. However, the complexity and randomness of network failures make it difficult to effectively predict them using traditional methods. Traditional statistical methods are overly mechanistic in predicting network failures. However, the use of machine learning offers the potential for accurate prediction of network failures. As features are added to the fault data, the prediction results of machine learning methods become increasingly accurate. Summary of the Invention

[0003] The purpose of this application is to provide a network fault prediction method and system based on a fully heterogeneous meta-contrast learning model, which can improve the accuracy of network fault prediction.

[0004] To achieve the above objectives, the present application provides, on the one hand, a network fault prediction method based on a fully heterogeneous meta-contrast learning model, including:

[0005] S100: Converting network fault data into time series data, wherein the network fault data at least includes fault type, fault occurrence time, and fault occurrence location;

[0006] S200: Constructing a network fault prediction model, the network fault prediction model comprising an embedding module, a heterogeneous meta-path network, a self-supervised contrastive learning module, and a prediction module; wherein the embedding module is configured to receive time series data and convert the time series data into an embedding vector; the heterogeneous meta-path network comprises a first heterogeneous meta-path network and a second heterogeneous meta-path network, each constructed based on a different meta-path; the self-supervised contrastive learning module is configured to perform contrastive learning on fault network views of a faulty node; and the prediction module is configured to predict fault probabilities using a loss function;

[0007] The first heterogeneous meta-path network is constructed by obtaining fault type and fault occurrence time data from time series data, using fault type and fault occurrence time as node objects, and using relationships between faults and fault occurrence times, relationships between faults, and relationships between fault occurrence times and fault occurrence times as edges; adding random walks to calculate and output the correlation between nodes in each instance path;

[0008] The second heterogeneous meta-path network is constructed as follows: fault type, fault occurrence time and fault location data are obtained from time series data, and the fault type, fault occurrence time and fault location are used as node objects, and the relationship between fault type and fault occurrence time, the relationship between fault type and fault type, the relationship between fault occurrence time and fault occurrence time, and the relationship between fault type and fault location are used as edges to construct the second heterogeneous meta-path network; the projection features of each node object in the second heterogeneous meta-path network are calculated respectively, two meta-paths using different node objects are designed, and the embeddings of the two meta-paths are fused and output using the attention mechanism.

[0009] Furthermore, the correlation between nodes in each instance path Among them, Pro(a f ,a f+1 ) represents the node a in the random walk of instance path p f to a f+1 The probability of a f+1 For node a f A randomly selected node around f+1 For node a f A random neighbor node of .

[0010] Furthermore, the embedding of the output of the second heterogeneous meta-path network in:

[0011] βp n is the important weight to measure the meta-path,

[0012] ωp n Used to weigh p n The importance of

[0013] T is transpose, tanh represents the activation function, is the feature encoded by the convolutional neural network, W∈R d×d and r∈R d×1 are all learnable parameters. represents the semantic attention vector, and V represents the node set.

[0014] Furthermore, the self-supervised contrastive learning module is used to construct a view fault network of the fault node for self-supervised learning, further including:

[0015] The constructed view fault networks include positive pairs and negative pairs. The view fault networks of the same fault node are taken as positive pairs, and the two view fault networks of the same fault node are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively. The view fault networks of any two different fault nodes are taken as negative pairs, and the view fault networks of two different fault nodes are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively.

[0016] Another aspect of the present application provides a network fault prediction system based on a fully heterogeneous meta-contrastive learning model, including:

[0017] A first module is used to convert network fault data into time series data, wherein the network fault data at least includes fault type, fault occurrence time and fault occurrence location;

[0018] The second module further includes an embedding module, a heterogeneous meta-path network, a self-supervised contrastive learning module, and a prediction module; wherein the embedding module is used to receive time series data and convert the time series data into an embedding vector; the heterogeneous meta-path network includes a first heterogeneous meta-path network and a second heterogeneous meta-path network, each constructed based on a different meta-path; the self-supervised contrastive learning module is used to compare and learn the fault network view of the fault node; and the prediction module is used to predict the fault probability using a loss function;

[0019] The first heterogeneous meta-path network is constructed by obtaining fault type and fault occurrence time data from time series data, using fault type and fault occurrence time as node objects, and using relationships between faults and fault occurrence times, relationships between faults, and relationships between fault occurrence times and fault occurrence times as edges; adding random walks to calculate and output the correlation between nodes in each instance path;

[0020] The second heterogeneous meta-path network is constructed as follows: fault type, fault occurrence time and fault location data are obtained from time series data, and the fault type, fault occurrence time and fault location are used as node objects, and the relationship between fault type and fault occurrence time, the relationship between fault type and fault type, the relationship between fault occurrence time and fault occurrence time, and the relationship between fault type and fault location are used as edges to construct the second heterogeneous meta-path network; the projection features of each node object in the second heterogeneous meta-path network are calculated respectively, two meta-paths using different node objects are designed, and the embeddings of the two meta-paths are fused and output using the attention mechanism.

[0021] Compared with the prior art, this application has the following advantages and beneficial effects:

[0022] By constructing meta-paths using two different meta-learning methods, we achieve richer representation and feature extraction of fault node information, thereby better understanding the impact relationships between different faults. Furthermore, through a self-supervised comparative learning framework, we maximize the interaction of all fault information between the two views, enabling global learning and further improving the model's prediction accuracy. Therefore, this application can improve the accuracy of network fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;

[0024] Figure 2 2 is a framework diagram of the heterogeneous meta-contrast learning model in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0026] Typically, the prediction operation for a fault only conforms to some of its characteristics, that is, the network itself has certain potential feature representations for certain faults. To address the above problems, inspired by meta-learning representation learning, this application proposes a heterogeneous meta-contrastive learning model and a prediction module. Through self-supervised contrastive learning, it captures high-level node information of the fault to generate different prediction methods, making the model more generalizable and thus achieving accurate prediction of network faults.

[0027] To make the technical solution and beneficial effects of this application more clear, the following will further illustrate the embodiments of this application in conjunction with the embodiments. It should be understood that the embodiments described herein are only used to explain this application and are not used to limit this application.

[0028] The network fault prediction method of the embodiment of the present application is based on the fully heterogeneous meta-contrast learning model, and its specific steps are shown in Figure 1 ,include:

[0029] S100: Convert network fault data into time series data.

[0030] In an embodiment of the present application, the network fault data includes at least the fault type, fault occurrence time, and fault occurrence location. Converting the network fault data into time series data includes: converting the network fault data into time series data based on the time node of the fault occurrence and the time series benchmark.

[0031] S200: Construct a network fault prediction model, which includes an embedding module, a heterogeneous meta-path network, a self-supervised comparative learning module and a prediction module, wherein the embedding module is used to receive time series data and convert the time series data into an embedding vector; the heterogeneous meta-path network includes a first heterogeneous meta-path network and a second heterogeneous meta-path network respectively constructed based on different meta-paths; the self-supervised comparative learning module is used to compare and learn the fault network views of the fault node; and the prediction module is used to use a loss function to perform fault prediction.

[0032] The following is a detailed description of the method for constructing a network fault prediction model.

[0033] S210: Construct an embedding module, which is used to receive time series data and convert the time series data into an embedding vector x = (x1, x2, x3, ..., x t ), t represents time.

[0034] S220: Construct a first heterogeneous meta-path network and a second heterogeneous meta-path network.

[0035] The heterogeneous meta-path network can be expressed as G = (V, E, A, R), where V and E represent sets of node objects and edges, respectively; A represents a node object. In the embodiment of the present application, the node objects include a fault node F, a fault occurrence time node T, and a fault occurrence location node L; R represents an inter-node relationship type. The inter-node relationship type refers to the relationship between node objects. In the heterogeneous meta-path network of the present application, the inter-node relationship is an edge, and |A+R|>2. The inter-node relationship type includes four categories: the relationship between the fault and the time of fault occurrence, the relationship between faults, the relationship between the time of fault occurrence and the time of fault occurrence, and the relationship between the fault and the location of the fault. For example, if a fault occurs at a certain time and place, this is the relationship between the corresponding fault node, the fault occurrence time node, and the fault occurrence location node; for another example, if two faults are associated or similar, this is the relationship between the corresponding fault nodes. Define the meta-path in the heterogeneous meta-path network Among them, A1, A2, ...A n Represents different node objects, R1, R2, ...R n Represents the relationship between node objects.

[0036] In the embodiment of the present application, two meta-paths with different connection modes are constructed, and the two meta-paths are encoded to obtain a first heterogeneous meta-path network and a second heterogeneous meta-path network. The construction of the first heterogeneous meta-path network and the second heterogeneous meta-path network will be described below.

[0037] (1) Construction of the first heterogeneous meta-path network

[0038] In the first type of meta-path construction, the node objects only consider the fault node F and the fault occurrence time node T, and only consider meta-paths with a path length of 2 or 3. Based on this, the meta-path is constructed. For the multiple meta-paths constructed, the eigenvalue Eig(P) of the meta-path P is calculated using the following formula:

[0039]

[0040] In formula (1), the eigenvalue Eig(P) is the sum of the correlations of the multiple meta-paths constructed, p represents the instance path of P, and cor(p) represents the correlation between nodes under the instance path.

[0041] For the instance path p=(a1,a2,…,a k ), where a1∈F, a k ∈F, a1 and a k All of them represent fault nodes, and the others are intermediate nodes of the instance path. The intermediate node object type is not limited. Assume that cor(p) is a particle that travels from a1 to a along the instance path p. k , then the estimation formula of cor(p) is as follows:

[0042]

[0043] In formula (2), Pro(a f ,a f+1 ) represents the node a in the random walk of instance path p f to a f+1 The probability of a f+1 For node a f A randomly selected node around f+1 For node a f A random neighbor node of .

[0044] Pro(a f ,a f+1 ) is calculated as follows:

[0045]

[0046] In formula (2), ω(a f ,a f+1 ) represents node a f and a f+1 The weight between instances; instance path p=(v1,v2,…,v n+1 ), v i ∈V,ω(a f ,v i ) represents node a f and v i The weight between f) represents node a in instance path p f Among the neighbor nodes of a f+1 A collection of nodes of the same type.

[0047] Based on the above, the method for constructing the first heterogeneous meta-path network in the embodiment of the present application is:

[0048] Fault type and fault occurrence time data are obtained from time series data. The fault type and fault occurrence time are used as node objects, and the relationship between faults and fault occurrence time, the relationship between faults, and the relationship between fault occurrence time and fault occurrence time are used as edges to construct the first heterogeneous meta-path network.

[0049] Calculate the eigenvalues ​​Eig(P) of all meta-paths in the first heterogeneous meta-path network. The eigenvalue is the sum of the correlations of all meta-paths. Add a random walk to calculate the correlations cor(p) between nodes in each instance path and output them. For the calculation of the correlation cor(p), see formula (2).

[0050] (2) Construction of the Second Heterogeneous Meta-Path Network

[0051] In the second meta-path construction, the node object also adds the node L where the fault occurs, but only considers the meta-path with path 2. Based on this, M meta-paths {P1, P2, P3, ..., P M Considering that there are different types of node objects and their features exist in different spaces, all types of node objects are projected into a common potential vector space, and the node objects of the same type are placed into a type-specific mapping matrix, so that the feature x of the node object is b Transformed into the common space, feature x b Extract from embedding vector; learn node embedding from the perspective of high-order meta-path structure. Specifically, from M meta-paths {P1, P2, P3, ..., P M} randomly select two different meta-paths and record them as P n and P m , starting from the randomly selected node b1, we get the node b1 based on the meta-path P n and P m The set of neighbor nodes

[0052] Metapath P n The specific encoding features are shown below:

[0053]

[0054]

[0055] In formulas (4) to (5): represents the projection feature of node b1; σ(·) represents the activation function; Represents the characteristics of node b1; represents the vector deviation; φ b1 represents a node of the same type as node b1, Represents a mapping matrix of a specific type, which refers to the type of node b1; d b and d b' represents the degree of nodes b and b', h b and h b' They represent the projection features of nodes b and b' respectively, which can be obtained by formula (4).

[0056] For node b1, there are M meta-paths {P1,P2,P3,...,P M} in the embed

[0057] Using the attention mechanism Fused into the final embedding:

[0058]

[0059]

[0060]

[0061] In the above formula, z b represents the node information fused under the current constructed path structure, that is, the output value of the second heterogeneous meta-path network; βp n is the measure of the meta-path P n The important weight, W∈R d×d and r∈R d×1 are all learnable parameters. represents the semantic attention vector, V represents the node set, ωp n is used to weigh p n The importance of, T is transposed, tanh represents the activation function, Features encoded for convolutional neural networks.

[0062] Based on the above, the second heterogeneous meta-path network in the embodiment of the present application is constructed as follows:

[0063] Fault type, fault occurrence time and fault location data are obtained from time series data. The fault type, fault occurrence time and fault location are used as node objects, and the relationship between fault type and fault occurrence time, the relationship between fault type and fault type, the relationship between fault occurrence time and fault occurrence time, and the relationship between fault type and fault location are used as edges to construct a second heterogeneous meta-path network.

[0064] Calculate the projection features of each node object in the second heterogeneous meta-path network respectively, design two meta-paths using different node objects and use the attention mechanism to fuse the embeddings of the two meta-paths, and output

[0065] S230: Construct a self-supervised contrastive learning module, which is used to construct a fault network view of the faulty node for self-supervised learning.

[0066] The fault network view of the same fault node is used as the positive The two fault network views of the same fault node are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively; the fault network views of any two different fault nodes are used as negative pairs. Similarly, the fault network views of two different fault nodes are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively.

[0067] Following SimCLR and using contrast loss InfoNCE for modeling, the final network fault prediction loss function value L is obtained. loss Here are the results:

[0068]

[0069]

[0070]

[0071] In formulas (9) to (11), s(·) is used to measure the similarity between two vectors. In the embodiment of the present application, the cosine similarity function is used. τ represents the temperature parameter, which is generally set to 0.1 or 0.2. V represents the node set. and are the loss function values ​​generated from the perspectives of the two meta-paths, and η is the balance and The coefficient between is generally set to 0.5.

[0072] The smaller the prediction loss value, the higher the prediction accuracy. The goal of self-supervised contrastive learning is to maximize the mutual information between two views, thereby achieving global learning.

[0073] S240: The prediction module is used to receive a given candidate node sequence and predict the probability of the candidate node becoming the next fault.

[0074] Note that the above are only preferred embodiments of the present application and the technical principles employed. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the scope of protection of the present application, all of which fall within the scope of protection of the present application.

Claims

1. A network fault prediction method based on a fully heterogeneous meta-contrastive learning model, characterized by: include: S100: Converting network fault data into time series data, wherein the network fault data at least includes fault type, fault occurrence time, and fault occurrence location; S200: Constructing a network fault prediction model, the network fault prediction model comprising an embedding module, a heterogeneous meta-path network, a self-supervised contrastive learning module, and a prediction module; wherein the embedding module is configured to receive time series data and convert the time series data into an embedding vector; the heterogeneous meta-path network comprises a first heterogeneous meta-path network and a second heterogeneous meta-path network, each constructed based on a different meta-path; the self-supervised contrastive learning module is configured to perform contrastive learning on fault network views of a faulty node; and the prediction module is configured to predict fault probabilities using a loss function; The first heterogeneous meta-path network is constructed by obtaining fault type and fault occurrence time data from time series data, using fault type and fault occurrence time as node objects, and using relationships between faults and fault occurrence times, relationships between faults, and relationships between fault occurrence times and fault occurrence times as edges; adding random walks to calculate and output the correlation between nodes in each instance path; The second heterogeneous meta-path network is constructed by: acquiring fault type, fault occurrence time, and fault occurrence location data from time series data, using fault type, fault occurrence time, and fault occurrence location as node objects, and using the relationship between fault type and fault occurrence time, the relationship between fault types, the relationship between fault occurrence time and fault occurrence time, and the relationship between fault type and fault occurrence location as edges to construct the second heterogeneous meta-path network; The projection features of each node object in the second heterogeneous meta-path network are calculated separately, two meta-paths using different node objects are designed, and the embeddings of the two meta-paths are fused and output using the attention mechanism.

2. The network fault prediction method based on the fully heterogeneous meta-contrastive learning model according to claim 1 is characterized by: The correlation between nodes in each instance path Among them, Pro(a f ,a f+1 ) represents the node a in the random walk of instance path p f to a f+1 The probability of a f+1 For node a f A randomly selected node around f+1 For node a f A random neighbor node of .

3. The network fault prediction method based on the fully heterogeneous meta-contrastive learning model according to claim 1 is characterized by: The embedding of the output of the second heterogeneous meta-path network in: βp n is the important weight to measure the meta-path, ωp n Used to weigh p n The importance of T represents transpose, tanh represents activation function, is the feature encoded by the convolutional neural network, W∈R d×d and r∈R d×1 are all learnable parameters. represents the semantic attention vector, and V represents the node set.

4. The network fault prediction method based on the fully heterogeneous meta-contrastive learning model according to claim 1 is characterized by: The self-supervised contrastive learning module is used to construct a view fault network of the fault node for self-supervised learning, further comprising: The constructed fault network views include positive pairs and negative pairs. The fault network views of the same fault node are taken as the positive pair, and the two fault network views of the same fault node are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively. The fault network views of any two different fault nodes are taken as the negative pair, and the fault network views of two different fault nodes are the outputs of the first heterogeneous meta-path network and the second heterogeneous meta-path network respectively.

5. A network fault prediction system based on a fully heterogeneous meta-contrastive learning model, characterized by: include: A first module is used to convert network fault data into time series data, wherein the network fault data at least includes fault type, fault occurrence time and fault occurrence location; The second module further includes an embedding module, a heterogeneous meta-path network, a self-supervised contrastive learning module, and a prediction module; wherein the embedding module is used to receive time series data and convert the time series data into an embedding vector; the heterogeneous meta-path network includes a first heterogeneous meta-path network and a second heterogeneous meta-path network, each constructed based on a different meta-path; the self-supervised contrastive learning module is used to compare and learn the fault network view of the fault node; and the prediction module is used to predict the fault probability using a loss function; The first heterogeneous meta-path network is constructed by obtaining fault type and fault occurrence time data from time series data, using fault type and fault occurrence time as node objects, and using relationships between faults and fault occurrence times, relationships between faults, and relationships between fault occurrence times and fault occurrence times as edges; adding random walks to calculate and output the correlation between nodes in each instance path; The second heterogeneous meta-path network is constructed by: acquiring fault type, fault occurrence time, and fault occurrence location data from time series data, using fault type, fault occurrence time, and fault occurrence location as node objects, and using the relationship between fault type and fault occurrence time, the relationship between fault types, the relationship between fault occurrence time and fault occurrence time, and the relationship between fault type and fault occurrence location as edges to construct the second heterogeneous meta-path network; The projection features of each node object in the second heterogeneous meta-path network are calculated separately, two meta-paths using different node objects are designed, and the embeddings of the two meta-paths are fused and output using the attention mechanism.

Images