Network fault prediction method and device based on self-supervised hypergraph

Abstract

This invention discloses a network fault prediction method and apparatus based on a self-supervised hypergraph, comprising the following steps: S1. First, network fault data is collected and converted into time series data, wherein all network fault types correspond to their respective time series data; S2. A neural network model based on a self-supervised hypergraph is constructed, the neural network model including an embedding layer, a query layer, a self-gating layer, and a self-supervised fusion layer, the aforementioned time series data is input into the neural network model based on the self-supervised hypergraph, and the neural network model based on the self-supervised hypergraph predicts faults in the network fault data. This invention mainly addresses the one-sidedness of directly mining the complex relationships of traditional network fault occurrence, and the randomness of network fault occurrence by filtering different network faults and then determining the fault repair order.

Description

Technical Field

[0001] This invention belongs to the field of network fault prediction, and relates to a network fault prediction method and apparatus based on self-supervised hypergraphs. Background Technology

[0002] The internet has become a large-scale market, and although the quality of network services has been continuously improving, the recurring occurrence of network failures remains a major concern for all operators. The most common manifestations of network failures include: complete service interruption, low downlink bandwidth, inability to access websites, noise during phone calls, and inability to establish phone calls. Reducing the occurrence of failures, i.e., increasing the mean time between failures, is a key focus for operators in network improvement. Furthermore, the complexity of networks makes accurate diagnosis difficult, potentially leading to a higher recurrence rate of problem failures.

[0003] Therefore, taking effective measures to provide early warnings of potential network failures and to accurately interpret the relationships between the causes of network failures are areas that need improvement. In recent years, 3GPP introduced network data analysis capabilities in Release 15, which is expected to become an AI engine for network functions. ETSI established the ZSM working group to achieve automated and intelligent network operation and maintenance. Meanwhile, leading global operators and equipment manufacturers are also strengthening cooperation in the field of AI-assisted network development. The industry is applying artificial intelligence technology to achieve intelligent 5G networks.

[0004] Analysis of current technological development shows that research on this technology is intense, but the ability to determine the causes of network failures is relatively low. Therefore, intelligent technologies for predicting network failures need to be improved. Summary of the Invention

[0005] The purpose of this invention is to provide a network fault prediction method that addresses the limitations of traditional prediction methods due to the complexity and inherent randomness of network faults. This method analyzes the impact of different network faults on potentially predicted faults, thereby providing a technical solution capable of accurately predicting network faults. To solve the aforementioned problems, a network fault prediction method using a self-supervised hypergraph neural network is provided.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A self-supervised hypergraph neural network prediction method includes the following steps:

[0008] The stored network fault data is preprocessed to obtain serialized network fault data;

[0009] A neural network model based on a self-supervised hypergraph is constructed. The serialized network fault data is divided into a test set and a training set. The neural network model based on the self-supervised hypergraph is then trained and tested to obtain a trained neural network model based on a self-supervised hypergraph.

[0010] After real-time acquisition and preprocessing of network fault data to obtain serialized network fault data, the data is input into a trained neural network model based on a self-supervised hypergraph. The model then makes predictions based on the network fault data.

[0011] Preferably, the network fault data includes the fault occurrence time, the node where the network fault occurred in a certain device, and then it is converted into time series data. Each time series represents the timestamp of the fault type collected in the current time period, and all fault types that occurred. Specifically, for the node data of the network fault, it is converted into sample serialized data based on the time node of its occurrence and the time series benchmark.

[0012] Preferably, the neural network model includes an embedding layer, a query layer, a self-gated layer, and a self-supervised fusion layer. The time series and label data are input into the self-supervised hypergraph neural network model, which then predicts data in the network fault data.

[0013] As a preferred approach, constructing a neural network model based on a self-supervised hypergraph includes:

[0014] S21. Construct an embedding layer that takes the fault time series and the text sequence of problem feedback as input, and transforms the input sequences into model embeddings e respectively. u and e t ;

[0015] S22. Construct a query layer that uses aggregated consistent feedback to learn node embeddings, which generates a query embedding by mapping the embeddings of fault time series and problem feedback text sequences.

[0016] S23. Construct a self-gated layer to extract more accurate node fault embedding vectors from the higher-order relationships between node faults. Specifically, use an automatic gating module to mine these deep connections. First, aggregate the feature information of the nodes to the hyperedge, and then aggregate the information from the hyperedge to the nodes. These two processes are called node-level attention mechanism and hyperedge-level attention mechanism, respectively.

[0017] S24. Construct a self-supervised fusion layer to incorporate self-supervised learning into the model's training:

[0018]

[0019] To obtain a representation of the sub-hypergraph:

[0020]

[0021] Set the read function R k×d →R d ,in This is to control the extent of P's involvement. It is A c The row vector, corresponding to the center node u, This represents the number of connections in the sub-hypergraph, where the weight of each faulty node in the sub-hypergraph is considered to be a z-factor in forming the sub-hypergraph. u Set up another read function R m×d →R d The resulting sub-hypergraph embeddings are summarized into a hierarchical representation:

[0022]

[0023] Two vectors are taken as input, and their consistency is scored. The hierarchical design in this paper preserves more structural information of the hypergraph in the representation of the faulty node.

[0024] Construct a loss function based on the obtained matrix:

[0025]

[0026] The fault time representation obtained in section S21 Through H obtained in section S22 t The average of the sums is then compared with the text feedback obtained in section S21. The scoring error is obtained by calculating the inner product, where E rating This is the set of all rating edges, and we use Adam as the optimizer. We also jointly learn the tasks of maximizing the objective and hierarchical mutual information to obtain the overall model loss function.

[0027] As a preferred option, in the query layer, the obtained multi-source data features e u ,e t ∈R d The prediction accuracy of the model can be increased by query embedding.

[0028] The formula is as follows:

[0029]

[0030] Where q u,t For query embedding, Indicates a connection, W q ∈R 2d×d σ is the mapping matrix, and σ is the ReLU activation function.

[0031] The framework for embedding aggregation nodes can be formalized as follows:

[0032]

[0033] Where σ is the ReLU activation function. For the implicit embedding of node v at layer l, N v Let W be the sampled neighbors of node v, AGG be the aggregation function, and W be the sampling neighbor. (l) ∈R 2d×d For encoding functions, This is the initial embedding for node v.

[0034] As a preferred option, the self-gating layer is based on the formula:

[0035]

[0036] and Let c ∈ {t, r} represent the training weights and biases, respectively. Let c ∈ {t, r} represent two different channels, which refer to the basic node fault embedding vectors of the hypergraph channels. u It is the initial input node fault embedding. This refers to the explicit basic node fault embedding vector. Let represent the dot product, and σ be the sigmoid function. SGU adjusts the basic node fault embeddings at the feature granularity through dimensionality reweighting, ultimately obtaining the node fault embedding E of the hypergraph channel. u .

[0037] As a preferred option, the specific implementation process of the self-gating layer includes:

[0038] S1. Using the sigmoid function as the activation function in the gated fusion layer, we obtain the attention coefficients and relevance of the nodes' contributions to the hyperedges:

[0039]

[0040]

[0041] a jk u represents the attention coefficient that node k contributes to the hyperedge when its information is aggregated to the hyperedge. k This indicates that node K is on hyperedge e. j Relevance on;

[0042] in, For the representation of hyperedges, the feature information X of nodes in the hypergraph (0) Using a multilayer perceptron, the basic embedding vector of faulty nodes in the hypergraph is used. and text basic embedding vectors Combining them, we get

[0043] σ is a sigmoid function, and W1 is a trainable weight matrix. This refers to the feature information of node k in the l-th layer of the neural network. This represents a series operation between two vectors. It is the basic embedding vector of a node in the hypergraph, and the method for obtaining it is the same as...

[0044] S2. Obtain hyperedge information of the hypergraph through a node-level attention mechanism:

[0045]

[0046] For node v i Updated feature information, W2 represents the weight matrix, b ij Represents node v i In the super-edge e j Attention coefficient.

[0047] A network fault prediction device based on self-supervised hypergraphs includes:

[0048] Preprocessing module: Used to preprocess the stored network fault data to obtain serialized network fault data;

[0049] The building module is used to build a neural network model based on a self-supervised hypergraph. It divides the serialized network fault data into a test set and a training set, and then trains and tests the neural network model based on the self-supervised hypergraph to obtain a trained neural network model based on the self-supervised hypergraph.

[0050] Prediction module: This module collects network fault data in real time, preprocesses it to obtain serialized network fault data, and then inputs it into a trained neural network model based on a self-supervised hypergraph. The model then makes predictions about the data in the network fault data.

[0051] A network fault prediction device based on a self-supervised hypergraph, the device comprising: a memory, a processor, and a fault prediction program stored in the memory and executable on the processor, wherein the fault prediction program, when executed by the processor, implements the steps of the network fault prediction method described above.

[0052] Compared with existing technologies, the present invention, employing the above technical solutions, has the following beneficial effects: 1. It proposes a new framework to mine higher-order information between data, enriching data feature representation by modeling hypergraphs at different levels, thus ensuring the effectiveness of prediction. 2. There are certain relationships between the intrinsic causes of some node failures. We can extract more accurate node failure embedding vectors from the higher-order relationships between node failures, thereby improving the accuracy of network failure prediction. 3. By integrating self-supervised learning into the training of the hypergraph attention network and demonstrating self-supervised auxiliary tasks for learning, it understands the higher-order relationships between different failures, effectively improving the efficiency of failure prediction implementation. Attached Figure Description

[0053] Figure 1 This is a framework diagram of a self-supervised hypergraph neural network model according to an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the network fault prediction method based on a self-supervised hypergraph neural network model according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the network fault prediction device based on a self-supervised hypergraph neural network model according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] like Figure 1 As shown, a self-supervised hypergraph neural network prediction method includes the following steps:

[0058] S1. Preprocess the network fault data and then convert it into time series data;

[0059] S2. Construct a neural network model based on a self-supervised hypergraph, which includes an embedding layer, a query layer, a self-gating layer, and a self-supervised fusion layer, and train and test the model.

[0060] S1 specifically includes:

[0061] The network fault data is preprocessed, including the fault occurrence time and the node where the network fault occurred in a certain device, and then converted into time series data. Each time series represents the timestamp of the fault type collected in the current time period, and all fault types that occurred.

[0062] like Figure 2 As shown, S2 specifically includes:

[0063] S21. Construct an embedding layer that takes the fault time series and the text sequence of problem feedback as input, and transforms the input sequences into model embeddings e respectively. u and e t ;

[0064] S22. Construct a query layer that leverages aggregated consistent feedback to learn node embeddings. This layer generates a query embedding by mapping the embeddings of the fault time series and the problem feedback text sequence. The formula is as follows:

[0065]

[0066] The framework for embedding aggregation nodes can be formalized as follows:

[0067]

[0068] S23. Construct a self-gated layer. In some cases, there is a relationship between the intrinsic causes of certain node failures. We can extract a more accurate node failure embedding vector from the higher-order relationships between node failures, as shown in the following formula:

[0069]

[0070] S24. Construct a self-supervised fusion layer. Since aggregation operations on higher-order relations may lead to the loss of higher-order information, we can incorporate self-supervised learning into the model training to address this issue:

[0071] The objective function for a self-supervised task is defined as follows:

[0072]

[0073] To obtain a representation of the sub-hypergraph:

[0074]

[0075] This invention designs a readout function R k×d →R d ,in This is to control the participation of P, avoid overfitting, and alleviate gradient conflicts between the primary and secondary tasks. It is A c The row vector, corresponding to the center data u, This represents the number of connections in the sub-hypergraph. Thus, the weight of each faulty node in the sub-hypergraph is considered to be a factor contributing to the formation of the sub-hypergraph. u Similarly, define another read function. R m×d →R dThe resulting sub-hypergraph embeddings are summarized into a hierarchical representation:

[0076]

[0077] Two vectors are taken as input, and their consistency is scored. The hierarchical design in this paper preserves more structural information of the hypergraph in the node representations.

[0078] Construct a loss function based on the obtained matrix:

[0079]

[0080] The fault time representation obtained in section S21 Through H obtained in section S22 t The average of the sums is then compared with the text feedback obtained in section S21. The scoring error is obtained by calculating the inner product, where E rating This is the set of all rating edges, and we use Adam as the optimizer. We also jointly learn the tasks of maximizing the objective and hierarchical mutual information to obtain the overall model loss function.

[0081] S23 specifically includes:

[0082] S231. Using the sigmoid function as the activation function in the gated fusion layer, we obtain the attention coefficients and relevance of the nodes' contributions to the hyperedges:

[0083]

[0084]

[0085] in, For the representation of hyperedges, the feature information X of nodes in the hypergraph (0) Using a multilayer perceptron, the basic embedding vector of faulty nodes in the hypergraph is used. Basic embedding vectors for question feedback Combining them, we get

[0086] S232. Obtain hyperedge information of the hypergraph through a node-level attention mechanism:

[0087]

[0088] For node v i Updated feature information, W2 represents the weight matrix, b ij Represents node v i In the super-edge e j Attention coefficient.

[0089] like Figure 3 As shown, according to another aspect of the present invention, a network fault prediction device is provided, the device comprising: a memory 301, a processor 302, and a fault prediction program stored in the memory 301 and executable on the processor 302, wherein the fault prediction program, when executed by the processor, implements the steps of the network fault prediction method described above.

[0090] The memory 301 is used to store long-term operating data of the device and extract multiple indicator visualization data from the operating data.

[0091] When the fault prediction program is executed by the processor 302, the processing module 302 is further configured to implement the following steps of the network fault prediction method:

[0092] For the network failure time series text, the node embedding of the time series text obtained after hypergraph attention processing is combined with the node embedding of the time series text obtained after social consistency processing.

[0093] Meanwhile, the text embedding of network fault feedback is multiplied by the text embedding after social consistency processing and the time series text node embedding combined with the above to obtain the final prediction result value.

[0094] By comparing the predicted values ​​with the actual values ​​again, the loss function of the subject is finally obtained. Combined with a certain loss ratio obtained from the self-supervised part mentioned in the previous part, the model is trained to obtain the optimal parameters of the model.

[0095] Different types of data can be used for prediction, improving prediction accuracy; the time spent resolving faults has almost zero impact on user service quality, reducing the company's operational costs; for different faults or degradations, different time points can be selected for prediction before they occur, reducing the system's computational load and saving system resources.

[0096] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0097] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0098] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A network failure prediction method based on self-supervised hypergraph, characterized in that, It comprises the following steps: (1) preprocessing the stored network fault data to obtain serialized network fault data; the network fault data comprises fault occurrence time, node of network fault occurring in a device, and is further converted into time series data, each time series representing a timestamp of a fault type collected in a current time period, all fault types occurring, and specifically, for node data of network fault, the node data is converted into sample serialized data according to a time node of occurrence and a time series reference; (2) constructing a neural network model based on a self-supervised hypergraph, training and testing the neural network model based on the self-supervised hypergraph after the serialized network fault data is divided into a test set and a training set, and obtaining the trained neural network model based on the self-supervised hypergraph; The neural network model comprises an embedding layer, a query layer, a self-gated layer, and a self-supervised fusion layer, the time series and label data are input into the neural network model based on the self-supervised hypergraph, and the neural network model based on the self-supervised hypergraph predicts data in the network fault data; (3) after real-time collection of network fault data is preprocessed to obtain serialized network fault data, the serialized network fault data is input into the trained neural network model based on the self-supervised hypergraph, and the model predicts data in the network fault data; In the step (2), the neural network model based on the self-supervised hypergraph comprises the following steps: S21. Construct an embedding layer to input the failure time series and the text sequence of problem feedback, and convert the input sequence into model embedding respectively and ; S22. A query layer is constructed to learn node embedding by using the information of aggregated consistent feedback, which generates a query embedding by mapping the embedding of fault time series and problem feedback text sequence; S23. A self-gated layer is constructed to extract more accurate node fault embedding vectors from high-order relationships between node faults, specifically, an automatic gating module is used to mine the deep relationship, and the information is aggregated from the hyperedge to the node, and the two processes are called node-level attention mechanism and hyperedge-level attention mechanism respectively; S24. A self-supervised fusion layer is constructed to add self-supervised learning to the training of the model: In order to obtain the representation of the sub-hypergraph: Setting up a readout function where is to control the participation amplitude of P, is a row vector of u, corresponding to the central data, denotes the number of connections in the sub-hypergraph, the weight of each faulty node in the sub-hypergraph is considered to be the , setting another readout function , the resulting sub-hypergraph is embedded into a level representation: Two vectors are taken as input, and then the consistency between them is scored, and the hierarchical design can retain more structural information of the hypergraph in the representation of the fault node; According to the obtained matrix, a loss function is constructed: The data obtained in S21 section is represented , and the text feedback obtained in S21 section is represented after being added and averaged with the data obtained in S22 section The inner product is calculated to obtain the score error, wherein is the set of all rating edges, using Adam as the optimizer, and the task of maximizing the target and hierarchical mutual information is jointly learned to obtain the overall model loss function. 2.The self-supervised hypergraph-based network failure prediction method of claim 1, wherein, In the query layer, the obtained multi-source data features are used to increase the prediction accuracy of the model through query embedding. The formula is as follows: wherein is a query embedding, denotes a concatenation, is a mapping matrix, is a ReLU activation function; The framework of aggregating node embedding can be formalized as: wherein is a ReLU activation function, is the hidden embedding of node v in layer 1, is a sampled neighbor of node v, AGG is an aggregation function, is an encoding function, is the initial embedding of node v. 3.The self-supervised hypergraph-based network failure prediction method of claim 1, wherein, The self-gated layer is based on the formula: and , representing the weight parameters and bias parameters during training, respectively. Representing two different channels, these refer to the basic node fault embedding vectors of the hypergraph channels. It is the initial input node fault embedding. This refers to the explicit basic node fault embedding vector. Represents the dot product. It is the Sigmoid function. SGU adjusts the basic node fault embeddings at the feature granularity through dimensionality reweighting, ultimately obtaining the node fault embeddings of the hypergraph channels. . 4.The self-supervised hypergraph based network failure prediction method of claim 1, wherein, The specific implementation process of the self-gated layer comprises: S1. The sigmoid function is used as the activation function in the gating fusion layer to obtain the attention coefficient and the correlation degree of the node contribution to the hyperedge: the attention coefficient that the information aggregation of node k contributes to the hyperedge, then represents the relevance of node K on the hyperedge ; wherein, is a representation of the hyperedge, the feature information of the nodes in the hypergraph using a multilayer perceptron to embed the faulty node in the hypergraph into a base embedding vector and a textual feedback base embedding vector combined, resulting in ; is a Sigmoid function, is a trainable weight matrix, refers to the feature information of node k in the l-th layer neural network, represents the operation between two vectors, is the basic embedding vector of the node in the hypergraph, and the obtaining method is the same as ; S2. The hyperedge information of the hypergraph is obtained through the node-level attention mechanism node updated feature information, representing a weight matrix, representing a node attention coefficient on a hyperedge .

5. A network failure prediction apparatus based on the method of claims 1-4, characterized by, It comprises: A preprocessing module for preprocessing the stored network fault data to obtain serialized network fault data; A construction module for constructing a neural network model based on a self-supervised hypergraph, training and testing the neural network model based on the self-supervised hypergraph after the serialized network fault data is divided into a test set and a training set, and obtaining the trained neural network model based on the self-supervised hypergraph; The prediction module is used for inputting the sequenced network fault data obtained by preprocessing the network fault data collected in real time into the trained self-supervised hypergraph-based neural network model, and the model predicts the data in the network fault data.

6. A network failure prediction device characterized by comprising: The device comprises a memory, a processor, and a fault prediction program stored on the memory and executable on the processor, and the fault prediction program, when executed by the processor, implements the steps of the network fault prediction method according to claims 1-4.

Images