A Fault Diagnosis Method for Cellular Networks Based on Deep Learning and Knowledge Graph

Abstract

The present invention relates to the technical field of communication networks and is a cellular network fault diagnosis method based on deep learning and knowledge graphs. The method uses a graph convolutional neural network (GCN) to pre-diagnose network alarm data. This method takes into account the correlation between alarm data and effectively improves the accuracy of network fault diagnosis. The long short-term memory network (LSTM) and conditional random field (CRF) are used to extract unstructured knowledge, and web crawler technology is used to extract semi-structured knowledge. Finally, structured data, semi-structured data, and unstructured data are fused to construct a comprehensive fault knowledge graph for 5G networks. The constructed knowledge graph is used to verify and infer the diagnosis results of the graph convolutional neural network. This method takes into account the complex relationships between alarms and faults, as well as the relationships between faults and causes and methods, greatly improving the accuracy of network fault diagnosis, reducing the threshold of network operation and maintenance, and improving the efficiency of network operation and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of communication networks, and specifically, to a cellular network fault diagnosis method based on deep learning and knowledge graph. Background Art

[0002] With the advent of the big data era and the rapid development of technologies such as deep learning, people can, with the support of powerful computing power, use complex neural network models to mine and extract key information from massive data. Especially in a complex heterogeneous network environment, thousands of network nodes generate a large amount of network operation information every day. In this development trend of network convergence and heterogeneity, fault diagnosis is a key research direction.

[0003] Fault diagnosis is one of the main tasks for managing any network. Traditional network fault diagnosis mainly compares the alarm information of network performance indicators with an expert experience library and manually analyzes and troubleshoots faults. However, in today's large-scale and complex heterogeneous wireless network environment, the diagnosis method based on manual analysis will consume a large amount of human and material resources and increase the maintenance cost. Therefore, there is an urgent need for a dynamic and adaptive network fault diagnosis method that can achieve accurate detection and diagnosis of network faults in a complex network environment, effectively alleviate the harms caused by fault propagation such as service interruption and network paralysis, which is of great significance to the evolution of wireless networks. Exploring and researching more efficient and intelligent fault diagnosis technologies in heterogeneous networks will surely become one of the important topics in future heterogeneous network research.

[0004] Due to the increasing maturity of computer technology, the fault diagnosis method based on deep learning has shown good strength in the field of fault diagnosis. As a branch of deep learning, GCN shows excellent performance in big data processing. GCN has now been preliminarily applied to the field of mechanical fault diagnosis. The network fault diagnosis method based on GCN first needs to obtain the feature attributes of data when obtaining the topological association graph between data, and then calculates the similarity according to the feature attributes to determine the topological association graph of the data set. In fact, this method based on the spectral clustering idea lacks interpretability and is difficult to further improve the classification accuracy of the GCN model. However, for the separate naive Bayes algorithm and GCN algorithm, the achieved diagnostic accuracy does not meet the requirements, the threshold for operation and maintenance personnel is relatively high, and the network operation and maintenance efficiency is also very low. Summary of the Invention

[0005] The present invention proposes a fault diagnosis method for cellular networks based on deep learning and knowledge graphs. The GCN model is used for network fault pre-diagnosis, which improves the accuracy of the model. The LSTM model and CRF model are used for knowledge extraction from unstructured data, and the crawler technology is used for knowledge extraction from semi-structured data. A comprehensive fault knowledge graph is constructed by integrating three-dimensional data of structured data, semi-structured data, and unstructured data to support fault diagnosis and recovery. The pre-diagnosis results of the GCN model are verified and supplemented using the constructed fault knowledge graph, solving the problem that a single GCN model cannot improve the accuracy, and obtaining an interpretable output. This method combines the advantages of deep learning and knowledge graphs. The model is superior to the individual naive Bayes algorithm and GCN algorithm, achieving better diagnostic accuracy, reducing the threshold for operation and maintenance personnel, and improving the efficiency of network operation and maintenance.

[0006] A fault diagnosis method for cellular networks based on deep learning and knowledge graphs disclosed by the present invention can not only improve the accuracy of the fault diagnosis model, but also improve the efficiency of network operation and maintenance. The specific technical solutions adopted are as follows:

[0007] A fault diagnosis method for cellular networks based on deep learning and knowledge graphs includes the following steps:

[0008] Step S1: Collect a labeled network status data set from a dense heterogeneous cellular network environment, and select the optimal subset from the data set through the XGBoost algorithm. The specific selection method is as follows:

[0009] Step S1.1: Obtain the importance scores of each feature through the feature importance ranking function of XGBoost and perform a descending order sorting.

[0010] Step S1.2: XGBoost continuously increases the feature selection threshold according to the importance scores, retains the feature parameters with scores higher than the threshold, and discards the others, thereby obtaining the accuracy of the XGBoost model under different feature combinations.

[0011] Step S1.3: Weigh the model accuracy and the number of features to obtain the optimal subset of network feature parameters.

[0012] Step S2: Map the optimal subset selected in Step S1 to an undirected graph G=(V, E), where V is the node set and E is the edge set, and construct a feature matrix X and an adjacency matrix A according to the data set.

[0013] Step S3: Input the feature matrix and adjacency matrix constructed in Step S2 into the GCN model to obtain the pre-diagnosis results.

[0014] Step S4: Obtain structured data, semi-structured data, and unstructured data from the dense heterogeneous cellular network management system. Use the LSTM and CRF models to extract knowledge from the unstructured data, and use web crawler technology to extract knowledge from the semi-structured data. Finally, construct a fault knowledge graph based on the three-dimensional data;

[0015] Step S5: Input the pre-diagnosis result output in Step S3 into the fault knowledge graph constructed in Step S4 to obtain the final network fault diagnosis result and the interpretability report.

[0016] For a further improvement of the present invention, the specific steps for constructing the feature matrix X and the adjacency matrix A according to the data set in Step S2 are as follows:

[0017] Step S2.1: Construct a feature matrix X ∈ n×k as shown in the following formula:

[0018]

[0019] where k represents that there are k key performance indicators (KPIs) in the data set, and n represents the number of samples in the data set. KPI m,k refers to the value of the kth KPI of the mth sample;

[0020] Step S2.2: Then construct a label matrix Y ∈ R n×c representing the label categories of the samples in the data set, as shown in the following formula:

[0021]

[0022] where c represents the number of fault types in the data set, and n represents the number of samples in the data set. In this method, c is set to 6, namely normal condition, uplink interference, downlink interference, coverage hole, radio interface fault, and base station fault.

[0023] where C 1,2 = 1 ∪ C 1,i (1 ≤ i ≤ c) = 0 indicates that the fault type of the first data sample is uplink interference.

[0024] Step S2.3: Finally, construct an adjacency matrix A ∈ R n×n to represent the connection relationship between nodes. In this method, the Gaussian function is selected to calculate the similarity measure s i,j (0 ≤ s i,j ≤ 1) as shown in the following formula:

[0025]

[0026] where δ = 1 is called the Gaussian function bandwidth parameter. When x iand x j The closer the Euclidean distance between them is, the more similar the \(i\)-th and \(j\)-th nodes in the graph are, and \(s\) i,j is larger; conversely, \(s\) i,j is smaller.

[0027] Then, a threshold \(\alpha\) is artificially initialized. If the node similarity measure \(s\) i,j is greater than or equal to the threshold \(\alpha\), the corresponding element \(A\) in the adjacency matrix i,j = 1; otherwise, \(A\) i,j = 0. Thus, the required adjacency matrix \(A\) is finally obtained as shown in the following formula:

[0028]

[0029] Furthermore, the specific steps for obtaining the preliminary diagnosis result in step S3 are as follows:

[0030] Step S3.1: Input the feature matrix \(X\) and the adjacency matrix \(A\) constructed in step S2 into the GCN model. The forward activation propagation formula defined in GCN is:

[0031]

[0032] where \(\sigma\) is the activation function. Since the adjacency matrix \(A\) only contains the connection information of each node in the graph with its adjacent nodes, after adding the identity matrix \(I\) N the graph convolution operation can aggregate the feature attributes of the node itself and its neighboring nodes. is the degree matrix of the matrix , and the values on its diagonal are the degrees of each node in the graph, and the elements outside the main diagonal are all 0 elements. \(W\) (l) is the trainable weight matrix in the \(l\)-th layer, which is essentially the convolution kernel filter parameter matrix. The parameters in the matrix need to be learned by training the model. During the training process of GCN, the parameters can be updated through error backpropagation and according to the gradient descent method. \(H\) (l) is the input node feature matrix of the \(l\)-th graph convolution layer. For the input layer, \(H\) (0) is equal to the initial node feature matrix \(X\).

[0033] Step S3.2: The GCN model used in this method contains two graph convolution layers. The activation function of the first layer is the ReLU function, and the second layer is the softmax function. The model output is a node feature matrix \(Z\in R\) n×c , where \(c\) is the pre-defined number of network fault categories and \(n\) is the number of dataset samples. For the output result matrix \(Z=(Z\) i,1 , \(Z\) i,2 ,…, \(Z\) i,c ), the sample node \(x\)i The predicted label is In addition, for the output result matrix, if when where λ th Depending on the specific network scenario, λ of this method th is set to 0.0001. At this time, the additional output sample is an anomaly identifier and needs to be input into the knowledge graph for further judgment.

[0034] Step S3.3: During the GCN training process, finally, the cross-entropy loss function needs to be calculated through the labeled samples in the training set. This function is the optimization objective of the model and enables the error to be propagated backward. According to the gradient descent method, the weights of the weight matrices in each graph convolutional layer are optimized.

[0035]

[0036] Among them, l refers to the number of labeled samples, c is the total number of previously defined network state categories, and Y is the previously defined label matrix of nodes.

[0037] Furthermore, the specific steps of the knowledge extraction algorithm for unstructured data based on the LSTM and CRF models in step S4 are as follows:

[0038] Step S4.1: Use the BIO annotation set adopted in the Bakeoff-3 evaluation, that is, B-P and I-P represent the first character and non-first character of the fault name, B-L and I-L represent the first character and non-first character of the relationship name, B-O and I-O represent the first character and non-first character of the entity name of the corresponding fault cause, and O represents that the character does not belong to a part of the named entity.

[0039] Step S4.2: Preprocess the text and data. This method uses one-hot encoding, which is a classic word vectorization method. This method constructs a corpus containing all words in the dataset, and then uses a vector with the same number of elements as the total number of words in the corpus to represent each word. And the value at the position corresponding to this word in the vector is 1, and the rest are 0. Map each character in the sentence to a word vector x obtained from the one-hot vector i Input it into the LSTM model to automatically extract sentence features, then use the combined LSTM and CRF model to score the entity recognition, and finally output the corresponding labels, such as B-P, I-P, O, etc.

[0040] Furthermore, the specific content of obtaining the final network fault diagnosis result and the interpretability report in step S5 is as follows:

[0041] Step S5.1: If the fault output by the GCN diagnostic model is (0, 0, 0, 1, 0, 0), it indicates that the cause of the fault is a coverage hole. Then, input the coverage hole into the KG, and relevant fault definitions, phenomena, possible causes, and solutions can be obtained. The above fault diagnosis and analysis process is detailed and accurate, effectively supporting the operation and maintenance personnel to quickly troubleshoot and recover the fault.

[0042] Step S5.2: If the fault output by the GCN diagnostic model is (0, 1, 0, 1, 0, 0), it indicates that the cause of the fault may be a coverage hole or uplink interference or both. Such faults need to be raised and input into the fault knowledge graph for further diagnosis. The knowledge graph first searches for detailed information about uplink interference and coverage holes, compares and checks them with the alarm information, and then inputs them into the fault rule base for reasoning to determine the root cause of the fault.

[0043] Advantages of the present invention: The present invention solves the problem of low accuracy of using the GCN model alone and the problem of lack of interpretable output in the current artificial intelligence-based diagnostic method; doing so can not only greatly improve the accuracy of the fault diagnosis model but also greatly improve the network operation and maintenance efficiency. Brief Description of the Drawings

[0044] Figure 1 It is a model diagram of the network fault diagnosis system provided by the present invention.

[0045] Figure 2 It is a diagram of the importance ranking of feature attributes in the present invention.

[0046] Figure 3 It is a GCN model diagram in the present invention.

[0047] Figure 4 It is a knowledge extraction model diagram in the present invention.

[0048] Figure 5 It is a schematic diagram of the intelligent auxiliary diagnosis process of network faults in the present invention. Detailed Embodiments

[0049] To deepen the understanding of the present invention, the following will further describe the present invention in detail with reference to the drawings and embodiments. The embodiments are only used to explain the present invention and do not limit the protection scope of the present invention.

[0050] The present invention considers network fault detection and diagnosis in the heterogeneous cellular network scenario. First, historical data is obtained from the heterogeneous wireless network historical database, including the fault category variable set, the fault variable set and its key performance indicators KPI, historical logs and other data. Then, the cause of the fault is initially diagnosed based on the GCN model, and finally, the final diagnosis result is obtained based on the knowledge graph.

[0051] Based onFigure 1 For the system model diagram shown, the present invention proposes a method for diagnosing faults in a cellular network based on deep learning and knowledge graphs. The specific steps are as follows:

[0052] Step S1: Collect a labeled network status data set from a dense heterogeneous cellular network environment, and select an optimal subset from the data set through the XGBoost algorithm. The specific selection method is as follows:

[0053] Step S1.1: Obtain the importance scores of each feature through the feature importance ranking function of XGBoost, and perform a descending order sorting.

[0054] Step S1.2: XGBoost continuously increases the feature selection threshold according to the importance scores, retains the feature parameters with scores higher than this threshold, and discards the others, thereby obtaining the accuracy of the XGBoost model under different feature combinations.

[0055] Step S1.3: Weigh the model accuracy and the number of features to obtain an optimal subset of network feature parameters.

[0056] Step S2: Map the optimal subset selected in Step S1 to an undirected graph G=(V, E), where V is the node set and E is the edge set, and construct a feature matrix X and an adjacency matrix A according to the data set.

[0057] Step S3: Input the feature matrix and the adjacency matrix constructed in Step S2 into the GCN model to obtain a preliminary diagnosis result.

[0058] Step S4: Obtain structured data, semi-structured data, and unstructured data from the dense heterogeneous cellular network management system, use the LSTM and CRF models to extract knowledge from the unstructured data, use web crawler technology to extract knowledge from the semi-structured data, and finally construct a fault knowledge graph based on the three-dimensional data.

[0059] Step S5: Input the preliminary diagnosis result output in Step S3 into the fault knowledge graph constructed in Step S4 to obtain the final network fault diagnosis result and an interpretability report.

[0060] The specific steps for constructing the feature matrix X and the adjacency matrix A according to the data set in Step S2 are as follows:

[0061] Step S2.1: Construct a feature matrix X∈ n×k , as shown in the following formula:

[0062]

[0063] where k represents that the data set has k key performance indicators (KPIs), and n represents the number of samples in the data set. KPIm,k Denote the value of the k-th KPI of the m-th sample;

[0064] Step S2.2: Then construct a label matrix Y ∈ R n×c Denote the label categories of the samples in the dataset, as shown in the following formula:

[0065]

[0066] where c represents the number of fault types in the dataset, and n represents the number of samples in the dataset. In this method, c is set to 6, namely normal condition, uplink interference, downlink interference, coverage hole, air interface fault, and base station fault.

[0067] where C 1,2 = 1 ∪ C 1,i (1 ≤ i ≤ c) = 0 indicates that the fault type of the first data sample is uplink interference.

[0068] Step S2.3: Finally, construct an adjacency matrix A ∈ R n×n to represent the connection relationship between nodes. In this method, the Gaussian function is selected to calculate the similarity measure s i,j (0 ≤ s i,j ≤ 1) between any two sample nodes, as shown in the following formula:

[0069]

[0070] where δ = 1 is called the Gaussian function bandwidth parameter. When the Euclidean distance between x i and x j is closer, it means that the i-th node and the j-th node in the graph are more similar, and s i,j is larger; conversely, s i,j is smaller.

[0071] Then, artificially initialize a threshold α. If the node similarity measure s i,j is greater than or equal to the threshold α, the corresponding element A i,j in the adjacency matrix is 1; otherwise, A i,j is 0. Thus, the required adjacency matrix A is finally obtained as shown in the following formula:

[0072]

[0073] The specific steps of the pre-diagnosis result obtained in step S3 are as follows:

[0074] Step S3.1: Input the feature matrix X and the adjacency matrix A constructed in step S2 into the GCN model. The forward activation propagation formula defined in GCN is:

[0075]

[0076] Among them, σ is the activation function. Since the adjacency matrix A only contains the connection information between each node and its adjacent nodes in the graph, plus the identity matrix I N After that, the graph convolution operation can aggregate the feature attributes of the node itself and its neighboring nodes. is the matrix of the degree matrix, where the values on the diagonal are the degrees of each node in the graph, and the elements outside the main diagonal are all 0 elements. W (l) is the trainable weight matrix in the l-th layer, which is essentially the convolutional kernel filter parameter matrix. The parameters in the matrix need to be learned by training the model. During the training process of the GCN, the parameters can be updated through error backpropagation and according to the gradient descent method. H (l) is the input node feature matrix of the l-th layer graph convolution layer. For the input layer, H (0) is equal to the initial node feature matrix X.

[0077] Step S3.2: The GCN model used in this method contains two graph convolution layers. The activation function of the first layer is the ReLU function, and the second layer is the softmax function. The model output is a node feature matrix Z ∈ R n×c , where c is the pre-defined number of network fault categories, and n is the number of dataset samples. For the output result matrix Z = (Z i,1 , Z i,2 , …, Z i,c ), the predicted label of the sample node x i is In addition, for the output result matrix, if when , where λ th is determined according to the specific network scenario. In this method, λ th is set to 0.0001. At this time, the additional output sample is an anomaly identifier and needs to be input into the knowledge graph for further judgment.

[0078] Step S3.3: During the GCN training process, finally, the cross-entropy loss function needs to be calculated through the labeled samples in the training set. This function is the optimization objective of the model, and it enables the error to be propagated backward, and the weights of the weight matrices in each graph convolution layer are optimized according to the gradient descent method.

[0079]

[0080] Among them, l refers to the number of labeled samples, c is the total number of previously defined network state categories, and Y is the previously defined node label matrix.

[0081] The specific steps of the knowledge extraction algorithm for unstructured data based on the LSTM and CRF models in step S4 are as follows:

[0082] Step S4.1: Use the BIO annotation set adopted in the Bakeoff-3 evaluation, that is, B-P and I-P represent the first character and non-first character of the fault name respectively, B-L and I-L represent the first character and non-first character of the relationship name respectively, B-O and I-O represent the first character and non-first character of the corresponding fault cause entity name respectively, and O represents that the character does not belong to a part of the named entity.

[0083] Step S4.2: Preprocess the text and data. This method uses one-hot encoding, which is a classic word vectorization method. This method constructs a corpus containing all words in the dataset, and then represents each word with a vector of the same length as the total number of words in the corpus, and the value at the position corresponding to this word in the vector is 1, and the rest are 0. Map each character in the sentence to a character vector x obtained from the one-hot vector i Input it into the LSTM model to automatically extract sentence features, then use the combined LSTM and CRF model to score the recognition of entities, and finally output the corresponding labels, such as B-P, I-P, O, etc.

[0084] The specific content of the final network fault diagnosis result and the interpretability report obtained in step S5 is as follows:

[0085] Step S5.1: If the fault output by the GCN diagnosis model is (0, 0, 0, 1, 0, 0), it means that the fault cause is a coverage hole. Then input the coverage hole into the KG, and relevant fault definitions, phenomena, possible causes, and solutions can be obtained. The above fault diagnosis and analysis process is detailed and accurate, effectively supporting the operation and maintenance personnel to quickly troubleshoot and recover the fault.

[0086] Step S5.2: If the fault output by the GCN diagnosis model is (0, 1, 0, 1, 0, 0), it means that the fault cause may be a coverage hole or an uplink interference or both. Such faults need to be proposed and input into the fault knowledge graph for further diagnosis. The knowledge graph first searches for detailed information about uplink interference and coverage holes, compares and checks with the alarm information, and then inputs it into the fault rule base for reasoning to determine the root cause of the fault.

[0087] To illustrate the effectiveness of the method proposed in the present invention, an example is given below:

[0088] Step S1: Collect a labeled network state dataset from a dense heterogeneous cellular network environment, perform data preprocessing, and select the optimal subset from the dataset through the XGBoost algorithm;

[0089] First, obtain the scores of each feature through the feature importance ranking function of XGBoost, and then perform a descending order ranking. The feature importance ranking is asFigure 2 As shown, feature screening is then carried out. If all KPIs are retained, the model accuracy is the highest. However, when the number of features is less than 11, the model accuracy drops significantly. Therefore, considering both the diagnostic accuracy and the complexity during training, the first 11 features are selected as the features of the dataset after dimensionality reduction.

[0090] Step S2: Map the optimal subset selected in Step S1 to an undirected graph G=(V, E), where V is the node set and E is the edge set. Construct the feature matrix X and the adjacency matrix A according to the dataset. Manually initialize a threshold α. If the node similarity measure s i,j is greater than or equal to the threshold α, then the corresponding element A in the adjacency matrix i,j =1; otherwise, A i,j =0. Thus, the required adjacency matrix A is finally obtained as shown in the following formula:

[0091]

[0092] Regarding the specific value of the threshold α, in this method, α is first initialized to 0.90 and then decreased by an interval of 0.05 successively. The rationality of the threshold selection is evaluated according to the diagnostic accuracy of the GCN model. As shown in the following table,

[0093] Table 1 Influence of α on the accuracy of GCN

[0094] Value range of α Accuracy of GCN model 0.90 84.46％ 0.85 87.84％ 0.80 88.06％ 0.75 86.26 0.70 78.72％

[0095] Step S3: Input the feature matrix and the adjacency matrix constructed in Step S2 into the GCN model. The GCN model diagram is as Figure 3 shown, and the pre-diagnosis result is obtained.

[0096] The GCN model used in this method contains two graph convolutional layers. The activation function of the first layer is the ReLU function, and the second layer is the softmax function. For the output result matrix Z=(Z i,1 , Z i,2 , …, Z i,c ), the predicted label of the sample node x i is In addition, for the output result matrix, if when , where λ th is determined according to the specific network scenario. In this method, λ th is set to 0.0001. At this time, the additional output sample is an abnormal identifier and needs to be input into the knowledge graph for further judgment.

[0097] Since the input feature dimension of the sample data is 11 and the final output feature dimension is 6, so in the first graph convolutional layer of the GCN model, W (0) ∈R11×7 , where \(W\) in the second-layer graph convolutional layer (1) \(\in\mathbb{R}\) 7×6 . In addition, a dropout layer with a rate of 0.25 is added to the input data of each convolutional layer in this method to mitigate the overfitting problem. Additionally, to prevent overfitting, early stopping is set in the training of this chapter. The stopping condition is that the loss function of the model in the network training process no longer decreases on the test data set, then the training process of the network is stopped, and the training parameter values of the previous round are output to reduce the training time of the model.

[0098] Step S4: Obtain structured data, semi-structured data, and unstructured data from the dense heterogeneous cellular network management system. Use the LSTM and CRF models to extract knowledge from the unstructured data, and use web crawler technology to extract knowledge from the semi-structured data. Finally, construct a comprehensive fault knowledge graph based on the three-dimensional data, as Figure 4 shown.

[0099] First, use the BIO annotation set adopted in the Bakeoff-3 evaluation, that is, B-P and I-P represent the first character and non-first character of the fault name, B-L and I-L represent the first character and non-first character of the relationship name, B-O and I-O represent the first character and non-first character of the entity name corresponding to the fault cause, and O represents that the character does not belong to a part of the named entity, as shown in the following table:

[0100] Table 2 Annotation data set

[0101] Weak B-P Yes O Cover I-P Parameter B-O Cover I-P Number I-O Of O Set I-O Can O Set I-O Possible O Not I-O Original B-L Combine I-O Reason I-L Reasonable I-O

[0102] Then, preprocess the text and data. This method uses one-hot encoding, which is a classic word vectorization method. This method constructs a corpus containing all words in the data set, and then represents each word with a vector of the same length as the total number of words in the corpus. And the value at the position corresponding to this word in the vector is 1, and the rest are 0. Map the character vector \(x\) obtained by mapping each character in the sentence by the one-hot vector i into the LSTM model to automatically extract sentence features, and then use the LSTM and CRF joint model to score the recognition of entities, and finally output the corresponding labels, such as B-P, I-P, O, etc.

[0103] Step S5: Input the pre-diagnosis result output in Step S3 into the knowledge graph constructed in Step S4 to obtain the final network fault diagnosis result and the interpretability report, as Figure 5 shown, and the specific content is as follows:

[0104] Step S5.1: If the fault output by the GCN diagnostic model is (0, 0, 0, 1, 0, 0), it means that the cause of the fault is a coverage hole. Then, input the coverage hole into the KG, and relevant fault definitions, phenomena, possible causes, and solutions can be obtained. The above fault diagnosis and analysis process is detailed and accurate, effectively supporting the operation and maintenance personnel to quickly troubleshoot and recover the fault.

[0105] Step S5.2: If the fault output by the GCN diagnostic model is (0, 1, 0, 1, 0, 0), it means that the cause of the fault may be a coverage hole or uplink interference or both. Such faults need to be proposed and input into the fault knowledge graph for further diagnosis. The knowledge graph first searches for detailed information about uplink interference and coverage holes, compares and checks it with the alarm information, and then inputs it into the fault rule base for reasoning to determine the root cause of the fault.

[0106] So far, by using the GCN model and the knowledge graph, the final network fault diagnosis task is completed.

[0107] The above are the exemplary embodiments of the present invention, and thus do not limit the protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A cellular network fault diagnosis method based on deep learning and knowledge graph, characterized in that, It includes the following steps: Step S1: Collect a labeled network state dataset from a dense heterogeneous cellular network environment, and select an optimal subset from the dataset through the XGBoost algorithm; Step S2: Map the optimal subset selected in Step S1 to an undirected graph G=(V, E), where V is the node set and E is the edge set, and construct a feature matrix X and an adjacency matrix A according to the dataset; Step S3: Input the feature matrix and the adjacency matrix constructed in Step S2 into the GCN model to obtain a pre-diagnosis result; Step S4: Obtain structured data, semi-structured data, and unstructured data from the dense heterogeneous cellular network management system, use the LSTM and CRF models to extract knowledge from the unstructured data, use web crawler technology to extract knowledge from the semi-structured data, and finally construct a fault knowledge graph based on the three-dimensional data; Step S5: Input the pre-diagnosis result output in Step S3 into the fault knowledge graph constructed in Step S4 to obtain the final network fault diagnosis result and an interpretability report; The specific steps for obtaining the pre-diagnosis result in Step S3 are as follows: Step S3.1: Input the feature matrix X and the adjacency matrix A constructed in Step S2 into the GCN model. The forward activation propagation formula defined in GCN is: Among them, σ is the activation function, A = A + I N , D is the degree matrix of matrix A, the values on its diagonal are the degrees of each node, and the elements outside the main diagonal are all 0 elements, W (l) is the trainable weight matrix in the l-th layer, H (l) is the input node feature matrix of the l-th layer graph convolutional layer. For the input layer, H (0) is equal to the initial node feature matrix X; Step S3.2: The GCN model contains two graph convolutional layers. The activation function of the first layer is the ReLU function, and the second layer is the softmax function. The model output is a node feature matrix Z ∈ R n×c , where c is the predefined number of network fault categories, n is the number of dataset samples. For the output result matrix Z = (Z i，1 Z i，2 ,..., Z i，c ), the predicted label of the sample node x i is In addition, for the output result matrix, if when , where λ th is determined according to the specific network scenario, and λ th is set to 0.0001, an additional output sample is an anomaly flag, which needs to be input into the knowledge graph for further judgment; Step S3.3: During the GCN training process, finally, the cross-entropy loss function needs to be calculated through the labeled samples in the training set. This function is the optimization target of the model, and it enables the error to be propagated backward. According to the gradient descent method, the weights of the weight matrices in each graph convolutional layer are optimized: where l refers to the number of labeled samples, c is the total number of previously defined network state categories, and Y is the previously defined label matrix of the nodes.

2. The method for diagnosing faults in a cellular network based on deep learning and knowledge graph according to claim 1, wherein, The specific selection method for selecting the optimal subset from the dataset through the XGBoost algorithm in Step S1 is as follows: Step S1.1: Obtain the importance scores of each feature through the feature importance ranking function of XGBoost and perform a descending order ranking; Step S1.2: XGBoost continuously increases the feature selection threshold according to the importance scores, retains the feature parameters with scores higher than this threshold, and discards the others, thereby obtaining the accuracy of the XGBoost model under different feature combinations. Step S1.3: Weigh the model accuracy and the number of features to obtain the optimal network feature parameter subset.

3. The method for diagnosing cellular network faults based on deep learning and knowledge graph according to claim 2, characterized in that, The specific steps for constructing the feature matrix X and the adjacency matrix A according to the dataset in Step S2 are as follows: Step S2.

1. Construct a feature matrix \(X\in\) n×k , as shown in the following formula: Among them, k represents that there are k key performance indicators (KPIs) in the data set, n represents the number of samples in the data set, and KPI m，k refers to the value of the k-th KPI of the m-th sample; Step S2.

2. Construct a label matrix \(Y\in\mathbb{R}\) n×c which represents the label categories of the samples in the dataset, as shown in the following formula: where c represents the number of fault types in the dataset, and n represents the number of samples in the dataset; Step S2.3: Construct an adjacency matrix \(A\in\mathbb{R}\) n×n to represent the connection relationships between nodes.

4. The method for diagnosing faults in a cellular network based on deep learning and knowledge graph according to claim 3, wherein In the step S2.2, set c to 6, which are normal situation, uplink interference, downlink interference, coverage hole, radio interface failure, and base station failure, where C 1，2 = 1 ∪ C 1，i (1 ≤ i ≤ c) = 0 indicates that the failure type of the first data sample is uplink interference.

5. The method for diagnosing faults in a cellular network based on deep learning and a knowledge graph according to claim 3, wherein In the step S2.3, the Gaussian function is selected to calculate the similarity measure s between any two sample nodes i，j (0 ≤ s i，j ≤ 1), as shown in the following formula: Among them, δ = 1 is called the Gaussian function bandwidth parameter. When x i and x j are closer to each other in Euclidean distance, it means that the i-th node and the j-th node in the graph are more similar, and s i,j is larger. On the contrary, s i,j is smaller; initially set a threshold α. If the node similarity measure s i,j is greater than or equal to the threshold α, then the corresponding element A in the adjacency matrix i，j = 1; otherwise, A i，j = 0. Finally, the required adjacency matrix A is obtained as shown in the following formula:

6. The method for diagnosing faults in a cellular network based on deep learning and knowledge graph according to claim 5, characterized in that, The specific steps of the knowledge extraction algorithm for unstructured data based on the LSTM and CRF models in Step S4 are as follows: Step S4.1: Use the BIO annotation set adopted in the Bakeoff-3 evaluation, that is, B-P and I-P represent the first character and non-first character of the fault name, B-L and I-L represent the first character and non-first character of the relationship name, B-O and I-O represent the first character and non-first character of the entity name of the corresponding fault cause, and O represents that the character does not belong to a part of the named entity. Step S4.2: Preprocess the text and data. Use one-hot encoding with a corpus that contains all the words in the dataset. Then, represent each word using a vector with the same number of elements as the total number of words in the corpus. The value at the position corresponding to this word in the vector is 1, and the rest are 0. Map each character in the sentence to a character vector x obtained from the one-hot vector. i Input it into the LSTM model to automatically extract sentence features. Then, use the combined LSTM and CRF model to score the entity recognition, and finally output the corresponding labels.

7. The method for diagnosing cellular network faults based on deep learning and knowledge graph according to claim 6, wherein The specific content of the final network fault diagnosis result and the interpretability report obtained in step S5 is as follows: Step S5.1: If the fault output by the GCN diagnosis model is (0, 0, 0, 1, 0, 0), it means that the fault cause is a coverage hole. Then, input the coverage hole into the KG, and relevant fault definitions, phenomena, possible causes, and solutions can be obtained; Step S5.2: If the fault output by the GCN diagnosis model is (0, 1, 0, 1, 0, 0), it means that the fault cause may be a coverage hole or uplink interference or both. Such faults need to be put forward and input into the fault knowledge graph for further diagnosis. The knowledge graph first searches for detailed information about uplink interference and coverage holes, compares and checks them with the alarm information, and then inputs them into the fault rule base for reasoning to determine the root cause of the fault.

Images