CN115734274A - Cellular network fault diagnosis method based on deep learning and knowledge graph - Google Patents
Cellular network fault diagnosis method based on deep learning and knowledge graph Download PDFInfo
- Publication number
- CN115734274A CN115734274A CN202211519142.XA CN202211519142A CN115734274A CN 115734274 A CN115734274 A CN 115734274A CN 202211519142 A CN202211519142 A CN 202211519142A CN 115734274 A CN115734274 A CN 115734274A
- Authority
- CN
- China
- Prior art keywords
- fault
- matrix
- knowledge
- network
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention relates to the technical field of communication networks, in particular to a cellular network fault diagnosis method based on deep learning and knowledge maps, which uses a method of a graph convolution neural network (GCN) to pre-diagnose network alarm data, considers the correlation among the alarm data and effectively improves the accuracy of network fault diagnosis; extracting unstructured knowledge by using a long-short term memory network LSTM and a conditional random field CRF, extracting semi-structured knowledge by using a crawler technology, and finally fusing structured data, semi-structured data and unstructured data to construct a comprehensive fault knowledge map facing a 5G network; the constructed knowledge graph is used for verifying and reasoning the diagnosis result of the graph convolution neural network, the method considers the complex relation between the alarm and the fault and the relation between the fault and the reason and the method, the network fault diagnosis accuracy is greatly improved, the threshold of network operation and maintenance is reduced, and the network operation and maintenance efficiency is improved.
Description
Technical Field
The invention relates to the technical field of communication networks, in particular to a cellular network fault diagnosis method based on deep learning and knowledge maps.
Background
With the coming of big data era and the rapid development of technologies such as deep learning, people can utilize a complex neural network model to mine and extract key information in mass data under the support of strong calculation power, and especially in a complex heterogeneous network environment, thousands of network nodes can generate a large amount of network operation information every day. With the development trend of network convergence and isomerization, fault diagnosis is a key research direction.
Troubleshooting is one of the main tasks of managing any network. The traditional network fault diagnosis is mainly to compare the alarm information of the network performance index with an expert experience base and manually analyze and investigate the fault, but in the heterogeneous wireless network environment with huge scale and complex structure, a diagnosis mode based on manpower analysis can occupy a large amount of manpower and material resources and increase the maintenance cost, so that a dynamic and self-adaptive network fault diagnosis method is urgently needed, accurate detection and diagnosis of the network fault in the complex network environment can be realized, the hazards of service interruption, network paralysis and the like caused by fault propagation are effectively relieved, the method has great significance to the evolution of the wireless network, and the exploration and research of a more efficient and more intelligent fault diagnosis technology in the heterogeneous network is one of important subjects of future heterogeneous network research.
Due to the increasing maturity of computer technology, fault diagnosis methods based on deep learning have shown unusual strength in the field of fault diagnosis. As a branch of deep learning, GCN shows excellent performance in large data processing, and GCN has now been primarily applied to the field of mechanical failure diagnosis. When the network fault diagnosis method based on the GCN is used for obtaining the topological correlation diagram among the data, firstly, the characteristic attribute of the data needs to be obtained, and then, the similarity is calculated according to the characteristic attribute so as to determine the topological correlation diagram of the data set. In fact, it is difficult to further improve the accuracy of GCN model classification while lacking interpretability based on such spectral clustering concepts. However, the obtained diagnosis accuracy rate does not meet the requirements, the threshold of operation and maintenance personnel is higher, and the network operation and maintenance efficiency is low by using the single naive Bayesian algorithm and the single GCN algorithm.
Disclosure of Invention
The invention provides a cellular network fault diagnosis method based on deep learning and knowledge maps, which utilizes a GCN model to pre-diagnose network faults, improves the accuracy of the model, utilizes an LSTM model and a CRF model to extract knowledge of unstructured data, adopts a crawler technology to extract knowledge of semi-structured data, integrates three-dimensional data of structured data, semi-structured data and unstructured data to construct a comprehensive fault knowledge map to support fault diagnosis and recovery, utilizes the constructed fault knowledge map to verify and supplement the pre-diagnosis result of the GCN model, solves the problem that the accuracy of a single GCN model cannot be improved, and obtains interpretable output. The method combines the advantages of deep learning and knowledge maps, and the model is superior to a single naive Bayesian algorithm and a single GCN algorithm, thereby obtaining better diagnosis accuracy, reducing the threshold of operation and maintenance personnel, and improving the network operation and maintenance efficiency.
The invention discloses a cellular network fault diagnosis method based on deep learning and knowledge graph, which can improve the accuracy of a fault diagnosis model and the network operation and maintenance efficiency, and adopts the following specific technical scheme:
a cellular network fault diagnosis method based on deep learning and knowledge graph comprises the following steps:
s1, collecting a network state data set with a label from an intensive heterogeneous cellular network environment, and selecting an optimal subset from the data set through an XGboost algorithm; the specific selection method is as follows:
s1.1, obtaining importance scores of all the characteristics through a characteristic importance sorting function of XGboost, and performing descending sorting;
and S1.2, continuously improving the feature selection threshold value by the XGboost according to the importance score, reserving the feature parameters with the score higher than the threshold value, and otherwise, discarding the feature parameters, thereby obtaining the accuracy of the XGboost model under different feature combinations.
And S1.3, balancing the model accuracy and the feature quantity to obtain an optimal network feature parameter subset.
S2, mapping the optimal subset selected in the step S1 into an undirected graph G = (V, E), wherein V is a node set, E is an edge set, and constructing a characteristic matrix X and an adjacent matrix A according to a data set;
s3, inputting the characteristic matrix and the adjacency matrix constructed in the S2 into a GCN model to obtain a pre-diagnosis result;
s4, acquiring structured data, semi-structured data and unstructured data from the intensive heterogeneous cellular network management system, performing knowledge extraction on the unstructured data by using an LSTM (local surface technology) and CRF (fuzzy rule target) model, performing knowledge extraction on the semi-structured data by using a crawler technology, and finally constructing a fault knowledge map based on three-dimensional data;
and S5, inputting the pre-diagnosis result output in the step S3 into the fault knowledge map constructed in the step S4 to obtain a final network fault diagnosis result and an interpretable report.
In a further improvement of the present invention, in the step S2, a feature matrix X is constructed according to the data set, and the specific steps of the adjacency matrix a are as follows:
s2.1, establishing a characteristic matrix X epsilon according to the data set n×k As shown in the following formula:
wherein k represents that the data set has k Key Performance Indicators (KPIs), and n represents the number of samples of the data set. KPI m,k Refers to the value of the kth KPI of the mth sample;
step S2.2, then constructing a label matrix Y epsilon R n×c The label class representing a sample in the dataset is shown as:
where c represents the number of failure types of the data set and n represents the number of samples of the data set. The method sets c to 6, and respectively indicates normal conditions, uplink interference, downlink interference, coverage holes, air interface faults and base station faults.
Wherein C is 1,2 ＝1∪C 1,i (1 ≦ i ≦ c) =0 indicates that the failure type of the first data sample is uplink interference.
Step S2.3, finally constructing an adjacency matrix A belonging to R n×n To represent the connection relationship between nodes. The method selects and uses a Gaussian function to calculate similarity measurement s between any two sample nodes i,j (0≤s i,j 1). Ltoreq.1) as shown by the following formula:
where δ =1 is referred to as a gaussian function bandwidth parameter. When x is i And x j The closer the Euclidean distances between the nodes are, the more similar the ith node and the jth node in the representation are, s i,j The larger; otherwise, s i,j The smaller.
Then, a threshold value alpha is artificially set initially, if the node similarity measure s i,j If the value is larger than or equal to the threshold value alpha, the corresponding element A in the adjacency matrix i,j =1; otherwise, then A i,j And =0. Thus, the desired adjacency matrix a is finally obtained as shown in the following formula:
further, the specific steps of the pre-diagnosis result obtained in step S3 are:
step S3.1: inputting the feature matrix X and the adjacency matrix A constructed in the step S2 into a GCN model, wherein a forward excitation propagation formula defined in the GCN is as follows:
where σ is the activation function.
Since the adjacency matrix A only contains nodes and phases in the figureConnection information of neighboring nodes, plus identity matrix I N The graph convolution operation can then aggregate the characteristic attributes of the node itself and the neighboring nodes.
Is a matrix
The diagonal values of the degree matrix of (1) are degrees of each node in the graph, and the elements outside the main diagonal are all 0 elements. W (l) The weight matrix is a trainable weight matrix in the l-th layer, which is essentially a parameter matrix of a convolution kernel filter, parameters in the matrix need to be obtained through learning of a training model, and the parameters can be updated according to a gradient descent method through error back propagation in the training process of the GCN. H (l) Is the input node feature matrix of the first layer graph convolution layer, for the input layer, H (0) It is equal to the initial node characteristic matrix X.
Step S3.2: the GCN model used by the method comprises two graph convolution layers, wherein the activation function of the first layer is a ReLU function, and the second layer is a softmax function. The output of the model is a node characteristic matrix Z epsilon R n×c Where c is the number of predefined network failure classes and n is the number of data set samples. For output result matrix Z = (Z) i,1 ,Z i,2 ,…,Z i,c ) Sample node x i The prediction label is
In addition, for the output result matrix, if
In which λ is th Lambda of the method, depending on the specific network scenario th Set to 0.0001. At this time, the extra output sample is an abnormal identifier and needs to be input into the knowledge graph for further judgment.
Step S3.3: in the GCN training process, a cross entropy loss function is calculated through the labeled samples in the training set, the function is an optimization target of the model, errors are made to propagate reversely, and the weight of the weight matrix in each graph convolution layer is optimized according to a gradient descent method.
Where l is the number of labeled samples, c is the total number of previously defined network state classes, and Y is the label matrix of previously defined nodes.
Further, the knowledge extraction algorithm for unstructured data based on LSTM and CRF models in step S4 specifically includes the following steps:
step S4.1: a BIO labeling set adopted in Bakeoff-3 evaluation is used, namely B-P, I-P represents a failure first character and a failure non-first character, B-L, I-L represents a relation first character and a relation non-first character, B-O, I-O represents a failure cause entity first character and a failure cause non-first character, and O represents that the character does not belong to one part of a named entity.
Step S4.2: the method adopts single hot coding and is a classical word vectorization method. The method constructs a corpus containing all words in a data set, then uses a vector with the same total word number as the corpus to represent each word, and the value of the position corresponding to the word in the vector is 1, and the rest positions are 0. A word vector x obtained by mapping each word in the sentence by a one-hot vector i Inputting the sentence features into an LSTM model, automatically extracting sentence features, then utilizing an LSTM and CRF combined model to score the identification of the entity, and finally outputting corresponding labels, such as B-P, I-P, O and the like.
Further, the final network fault diagnosis result and the interpretable report obtained in step S5 have the following specific contents:
step S5.1: if the GCN diagnostic model output fault is (0,0,0,1,0,0), the fault is indicated to be a coverage hole. And then inputting the coverage holes into KG, and then obtaining related fault definitions, phenomena, possible reasons and solutions. The fault diagnosis and analysis process is detailed and accurate, and operation and maintenance personnel are effectively supported to rapidly troubleshoot and recover faults.
Step S5.2: if the GCN diagnosis model output fault is (0,1,0,1,0,0), the fault reason is possible to be a coverage hole or uplink interference or two reasons. The faults need to be extracted and input into a fault knowledge graph for further diagnosis. The knowledge graph firstly searches detailed information about uplink interference and coverage holes, compares the detailed information with alarm information, and inputs the information into a fault rule base for reasoning so as to determine a fault root.
The invention has the beneficial effects that: the invention solves the problem of low accuracy of the GCN model used alone and solves the problem of the lack of interpretable output of the current artificial intelligence-based diagnosis method; by the method, the accuracy of the fault diagnosis model can be greatly improved, and the network operation and maintenance efficiency is greatly improved.
Drawings
Fig. 1 is a diagram of a network fault diagnosis system provided by the present invention.
FIG. 2 is a graph of feature attribute importance ranking in the present invention.
FIG. 3 is a diagram of a GCN model in the present invention.
FIG. 4 is a diagram of a knowledge extraction model in the present invention.
Fig. 5 is a schematic diagram of the intelligent auxiliary diagnosis process of the network fault in the invention.
Detailed Description
For the purpose of enhancing the understanding of the present invention, the present invention will be described in further detail with reference to the accompanying drawings and examples, which are provided for the purpose of illustration only and are not intended to limit the scope of the present invention.
The method comprises the steps of firstly obtaining historical data from a heterogeneous wireless network historical database, wherein the historical data comprises data of a fault category variable set, a fault variable set, key performance indicators KPI (key performance indicator), historical logs and the like, then preliminarily diagnosing fault reasons based on a GCN (general knowledge network) model, and finally obtaining a final diagnosis result based on a knowledge graph.
Based on a system model diagram shown in fig. 1, the invention provides a cellular network fault diagnosis method based on deep learning and a knowledge graph, which comprises the following specific steps:
s1, collecting a network state data set with a label from an intensive heterogeneous cellular network environment, and selecting an optimal subset from the data set through an XGboost algorithm; the specific selection method is as follows:
s1.1, obtaining importance scores of all the characteristics through a characteristic importance sorting function of XGboost, and performing descending sorting;
and S1.2, continuously improving the feature selection threshold value by the XGboost according to the importance score, reserving the feature parameters with the score higher than the threshold value, and otherwise, discarding the feature parameters, thereby obtaining the accuracy of the XGboost model under different feature combinations.
And S1.3, balancing the model accuracy and the feature quantity to obtain an optimal network feature parameter subset.
S2, mapping the optimal subset selected in the step S1 into an undirected graph G = (V, E), wherein V is a node set, E is an edge set, and constructing a characteristic matrix X and an adjacent matrix A according to a data set;
s3, inputting the characteristic matrix and the adjacency matrix constructed in the step S2 into a GCN model to obtain a pre-diagnosis result;
s4, acquiring structured data, semi-structured data and unstructured data from the intensive heterogeneous cellular network management system, performing knowledge extraction on the unstructured data by using an LSTM (local surface technology) and CRF (fuzzy rule target) model, performing knowledge extraction on the semi-structured data by using a crawler technology, and finally constructing a fault knowledge map based on three-dimensional data;
and S5, inputting the pre-diagnosis result output in the step S3 into the fault knowledge map constructed in the step S4 to obtain a final network fault diagnosis result and an interpretable report.
In the step S2, a feature matrix X is constructed according to the data set, and the adjacency matrix a specifically includes the steps of:
s2.1, establishing a characteristic matrix X epsilon according to the data set n×k As shown in the following formula:
wherein k represents that the data set has k Key Performance Indicators (KPIs), and n represents the number of samples of the data set. KPI m,k Refers to the value of the kth KPI of the mth sample;
step S2.2, then constructing a label matrix Y epsilon R n×c The label class representing a sample in the dataset is shown as:
where c represents the number of failure types of the data set and n represents the number of samples of the data set. The method sets c to 6, and respectively indicates normal conditions, uplink interference, downlink interference, coverage holes, air interface faults and base station faults.
Wherein C is 1,2 ＝1∪C 1,i (1 ≦ i ≦ c) =0 indicates that the failure type of the first data sample is uplink interference.
Step S2.3, finally constructing an adjacency matrix A belonging to R n×n To represent the connection relationship between nodes. The method selects and uses a Gaussian function to calculate similarity measurement s between any two sample nodes i,j (0≤s i,j 1) as shown below:
where δ =1 is referred to as a gaussian function bandwidth parameter. When x is i And x j The closer the Euclidean distances between the nodes are, the more similar the ith node and the jth node in the representation are, s i,j The larger; otherwise, s i,j The smaller.
Then, a threshold value alpha is artificially set initially, if the node similarity measure s i,j If the value is larger than or equal to the threshold value alpha, the corresponding element A in the adjacency matrix i,j =1; otherwise, then A i,j And =0. Thus, the desired adjacency matrix a is finally obtained as followsIs represented by the formula:
the specific steps of the pre-diagnosis result obtained in the step S3 are as follows:
step S3.1: inputting the feature matrix X and the adjacency matrix A constructed in the step S2 into a GCN model, wherein a forward excitation propagation formula defined in the GCN is as follows:
where σ is the activation function.
Because the adjacent matrix A only contains the connection information of each node and adjacent nodes in the graph, and the unit matrix I is added N The graph convolution operation can then aggregate the characteristic attributes of the node itself and the neighboring nodes.
Is a matrix
The diagonal values of the degree matrix of (1) are degrees of each node in the graph, and the elements outside the main diagonal are all 0 elements. W (l) The weight matrix is a trainable weight matrix in the first layer, which is essentially a parameter matrix of a convolution kernel filter, parameters in the matrix need to be obtained through learning of a training model, and the parameters can be updated according to a gradient descent method through error back propagation in the training process of the GCN. H (l) Is the input node feature matrix of the first layer graph convolution layer, for the input layer, H (0) It is equal to the initial node characteristic matrix X.
Step S3.2: the GCN model used by the method comprises two graph convolution layers, wherein the activation function of the first layer is a ReLU function, and the second layer is a softmax function. The output of the model is a node characteristic matrix Z epsilon R n×c Where c is the number of predefined network failure classes and n is the number of data set samples. For output result matrix Z = (Z) i,1 ,Z i,2 ,…,Z i,c ) Sample node x i The prediction label is
In addition, for the output result matrix, if
In which λ is th Lambda of the method, depending on the specific network scenario th Set to 0.0001. At this time, the extra output sample is an abnormal identifier and needs to be input into the knowledge graph for further judgment.
Step S3.3: in the GCN training process, a cross entropy loss function is calculated through the labeled samples in the training set, the function is an optimization target of the model, errors are made to propagate reversely, and the weight of the weight matrix in each graph convolution layer is optimized according to a gradient descent method.
Where l is the number of labeled samples, c is the total number of previously defined network state classes, and Y is the label matrix of previously defined nodes.
The knowledge extraction algorithm for the unstructured data based on the LSTM and CRF models in the step S4 comprises the following specific steps:
step S4.1: a BIO labeling set adopted in Bakeoff-3 evaluation is used, namely B-P, I-P represents a failure first character and a failure non-first character, B-L, I-L represents a relation first character and a relation non-first character, B-O, I-O represents a failure cause entity first character and a failure cause non-first character, and O represents that the character does not belong to one part of a named entity.
Step S4.2: the method adopts single hot coding and is a classical word vectorization method. This method constructs a corpus containing all words in the dataset, and thenEach word is represented by a vector with the same number of words as the total number of words in the corpus, and the position corresponding to the word in the vector has a value of 1, and the rest positions are 0. A word vector x obtained by mapping each word in the sentence by a one-hot vector i Inputting the sentence features into an LSTM model, automatically extracting sentence features, then utilizing an LSTM and CRF combined model to score the recognition of the entity, and finally outputting a corresponding label, such as B-P, I-P, O and the like.
The specific contents of the final network fault diagnosis result and the interpretable report obtained in the step S5 are as follows:
step S5.1: if the GCN diagnosis model output fault is (0,0,0,1,0,0), the fault cause is a coverage hole. And then inputting the coverage hole into the KG, and obtaining related fault definition, phenomenon, possible reason and solution. The fault diagnosis and analysis process is detailed and accurate, and operation and maintenance personnel are effectively supported to rapidly troubleshoot and recover faults.
Step S5.2: if the GCN diagnosis model outputs that the fault is (0,1,0,1,0,0), the fault is indicated to be caused by a coverage hole or uplink interference or caused by two reasons. The faults need to be extracted and input into a fault knowledge graph for further diagnosis. The knowledge graph firstly searches detailed information about uplink interference and coverage holes, compares the detailed information with alarm information, and inputs the information into a fault rule base for reasoning so as to determine a fault root.
To illustrate the effectiveness of the proposed method of the invention, an example is given below:
s1, collecting a network state data set with a label from an intensive heterogeneous cellular network environment, preprocessing the data, and selecting an optimal subset from the data set through an XGboost algorithm;
the method comprises the steps of firstly obtaining scores of all features through a feature importance sorting function of XGboost, then conducting descending sorting, wherein the feature importance sorting is shown in figure 2, then conducting feature screening, if all KPIs are reserved, the model accuracy is highest, but when the number of the features is lower than 11, the model accuracy is greatly reduced. Therefore, the diagnosis accuracy and the training complexity are comprehensively considered, so that the first 11 features are selected as the features of the reduced-dimension data set.
S2, mapping the optimal subset selected in the step S1 into an undirected graph G = (V, E), wherein V is a node set, E is an edge set, and constructing a characteristic matrix X and an adjacent matrix A according to a data set; artificially initializing a threshold value alpha if the node similarity measure s i,j If the value is larger than or equal to the threshold value alpha, the corresponding element A in the adjacency matrix i,j =1; otherwise, then A i,j And =0. Thus, the desired adjacency matrix a is finally obtained as shown in the following equation:
regarding the specific value problem of the threshold value alpha, the method initializes the alpha to 0.90, then reduces the alpha in sequence at intervals of 0.05, and evaluates the rationality of threshold value selection according to the diagnosis accuracy of the GCN model. As shown in the following table, the following,
TABLE 1 Effect of alpha on GCN accuracy
| Value of alpha | GCN model accuracy |
| 0.90 | 84.46％ |
| 0.85 | 87.84％ |
| 0.80 | 88.06％ |
| 0.75 | 86.26 |
| 0.70 | 78.72％ |
And S3, inputting the characteristic matrix and the adjacency matrix constructed in the step S2 into a GCN model, wherein a GCN model graph is shown in figure 3, and a pre-diagnosis result is obtained.
The GCN model used by the method comprises two graph convolution layers, wherein the activation function of the first layer is a ReLU function, and the second layer is a softmax function. For output result matrix Z = (Z) i,1 ,Z i,2 ,…,Z i,c ) Sample node x i The prediction label is
In addition, for the output result matrix, if so
In which λ is th The method sets lambda according to specific network scene th Is 0.0001. At this time, the extra output sample is an abnormal identifier and needs to be input into the knowledge graph for further judgment.
Since the input characteristic dimension of the sample data is 11 and the final output characteristic dimension is 6, W in the first layer graph convolution layer of the GCN model (0) ∈R 11×7 W in the second layer of the graph convolution layer (1) ∈R 7×6 . In addition, 0.25 dropout layers are added to the input data of each convolution layer in the method, so that the overfitting problem is relieved. In addition, in order to prevent overfitting, the training of the chapter is set to be stopped in advance, if the stopping condition is that the loss function of the model in the network training process on the test data set is not reduced any more, the network training process is stopped, and the training parameter value of the previous round is output to reduce the time for training the model.
And S4, acquiring structured data, semi-structured data and unstructured data from the intensive heterogeneous cellular network management system, performing knowledge extraction on the unstructured data by using an LSTM (local surface technology) and CRF (fuzzy rule target) model, performing knowledge extraction on the semi-structured data by using a crawler technology, and finally constructing a comprehensive fault knowledge map based on three-dimensional data, as shown in FIG. 4.
Firstly, a BIO labeling set adopted in Bakeoff-3 evaluation is used, namely B-P, I-P represents a first character of a fault name and a non-first character of the fault name, B-L, I-L represents a first character of a relation name and a non-first character of the relation name, B-O, I-O represents a first character of a fault cause entity and a non-first character of the fault cause entity, and O represents that the character does not belong to a part of a named entity, as shown in the following table:
TABLE 2 annotated data set
| Weak (weak) | B-P | Is that | O |
| Coating(s) | I-P | Ginseng radix (Panax ginseng C.A. Meyer) | B-O |
| Cover | I-P | Number of | I-O |
| Is/are as follows | O | Is provided with | I-O |
| Can be used for | O | Device for placing | I-O |
| Can be used for | O | Is not limited to | I-O |
| Original source | B-L | Combination of Chinese herbs | I-O |
| Due to the fact that | I-L | Theory of things | I-O |
And then preprocessing the text and the data, wherein the method adopts single hot coding and is a classical word vectorization method. The method constructs a corpus containing all words in a data set, then uses a vector with the same total word number as the corpus to represent each word, and the value of the position corresponding to the word in the vector is 1, and the rest positions are 0. A word vector x obtained by mapping each word in the sentence by a one-hot vector i Inputting the sentence features into an LSTM model, automatically extracting sentence features, then utilizing an LSTM and CRF combined model to score the identification of the entity, and finally outputting corresponding labels, such as B-P, I-P, O and the like.
Step S5, inputting the pre-diagnosis result output in step S3 into the knowledge graph constructed in step S4 to obtain a final network fault diagnosis result and an interpretable report, as shown in fig. 5, the specific contents are as follows:
step S5.1: if the GCN diagnostic model output fault is (0,0,0,1,0,0), the fault is indicated to be a coverage hole. And then inputting the coverage holes into KG, and then obtaining related fault definitions, phenomena, possible reasons and solutions. The fault diagnosis and analysis process is detailed and accurate, and operation and maintenance personnel are effectively supported to rapidly troubleshoot and recover faults.
Step S5.2: if the GCN diagnosis model output fault is (0,1,0,1,0,0), the fault reason is possible to be a coverage hole or uplink interference or two reasons. The faults need to be extracted and input into a fault knowledge graph for further diagnosis. The knowledge graph firstly searches detailed information about uplink interference and coverage holes, compares the detailed information with alarm information, and inputs the information into a fault rule base for reasoning so as to determine a fault root.
At this point, the final network fault diagnosis task is completed by using the GCN model and the knowledge graph.
The above description is an exemplary embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.
Claims (8)
1. A cellular network fault diagnosis method based on deep learning and knowledge graph is characterized by comprising the following steps:
s1, collecting a network state data set with a label from an intensive heterogeneous cellular network environment, and selecting an optimal subset from the data set through an XGboost algorithm;
s2, mapping the optimal subset selected in the step S1 into an undirected graph G = (V, E), wherein V is a node set, E is an edge set, and constructing a characteristic matrix X and an adjacent matrix A according to a data set;
s3, inputting the characteristic matrix and the adjacency matrix constructed in the step S2 into a GCN model to obtain a pre-diagnosis result;
s4, acquiring structured data, semi-structured data and unstructured data from the intensive heterogeneous cellular network management system, performing knowledge extraction on the unstructured data by using an LSTM (local surface technology) and CRF (fuzzy rule target) model, performing knowledge extraction on the semi-structured data by using a crawler technology, and finally constructing a fault knowledge map based on three-dimensional data;
and S5, inputting the pre-diagnosis result output in the step S3 into the fault knowledge map constructed in the step S4 to obtain a final network fault diagnosis result and an interpretable report.
2. The deep learning and knowledge graph-based cellular network fault diagnosis method according to claim 1, wherein in step S1, a specific selection method for selecting an optimal subset from a data set through an XGBoost algorithm is as follows:
s1.1, obtaining importance scores of all the characteristics through a characteristic importance sorting function of XGboost, and performing descending sorting;
s1.2, continuously improving the feature selection threshold value by the XGboost according to the importance score, keeping the feature parameters with the score higher than the threshold value, otherwise, discarding the feature parameters, further obtaining the accuracy of the XGboost model under different feature combinations,
and S1.3, balancing the model accuracy and the feature quantity to obtain an optimal network feature parameter subset.
3. The method for diagnosing the cellular network fault based on the deep learning and the knowledge graph according to claim 2, wherein the step S2 is to construct a feature matrix X according to a data set, and the specific steps of the adjacency matrix a are as follows:
s2.1, establishing a characteristic matrix X epsilon according to the data set n×k As shown in the following formula:
wherein k represents that the data set has k key performance indicators KPI, n represents the number of samples of the data set, KPI m,k Refers to the value of the kth KPI of the mth sample;
step S2.2, constructing a label momentThe matrix Y is formed by R n×c The label class representing a sample in the dataset is shown as:
wherein c represents the number of failure types of the data set, and n represents the number of samples of the data set;
step S2.3, an adjacency matrix A epsilon R is constructed n×n To represent the connection relationship between nodes.
4. The method according to claim 3, wherein in step S2.2, C is set to 6, which is normal condition, uplink interference, downlink interference, coverage hole, air interface fault, and base station fault, where C is 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 deep learning and knowledge-graph based cellular network fault diagnosis method according to claim 3, characterized in that in step S2.3, the similarity measure S between any two sample nodes calculated by using Gaussian function is selected i,j (0≤s i,j 1). Ltoreq.1) as shown by the following formula:
where δ =1 is called the gaussian function bandwidth parameter, when x i And x j The closer the Euclidean distances between the nodes are, the more similar the ith node and the jth node in the representation are, s i,j The larger and vice versa s i,j The smaller; initially setting a threshold value alpha if the node similarity measure s i,j If the value is larger than or equal to the threshold value alpha, the corresponding element A in the adjacency matrix i,j =1; otherwise, then A i,j =0, finally gives the desiredThe adjacency matrix a is shown as follows:
6. the deep learning and knowledge-graph-based cellular network failure diagnosis method according to claim 4 or 5, wherein the steps of obtaining the pre-diagnosis result in the step S3 are as follows:
step S3.1: inputting the feature matrix X and the adjacency matrix A constructed in the step S2 into a GCN model, wherein a forward excitation propagation formula defined in the GCN is as follows:
where, σ is the activation function,
is a matrix
The diagonal values of the degree matrix of (1) are degrees of each node, the elements outside the main diagonal are all 0 elements, W (l) Is a trainable weight matrix in the l-th layer, H (l) Is the input node feature matrix of the first layer graph convolution layer, for the input layer, H (0) Is equal to the initial node feature matrix X;
step S3.2: the GCN model comprises two graph convolution layers, the activation function of the first layer is a ReLU function, the second layer is a softmax function, and the output of the model is a node characteristic matrix Z belonging to R n×c Where c is the predefined number of network fault classes, n is the number of data set samples, and for the output result matrix Z = (Z) i,1 ,Z i,2 ,…,Z i,c ) Sample node x i The prediction label is
In addition, for the output result matrix, if so
In which λ is th Depending on the particular network scenario, λ th Setting the value to be 0.0001, wherein the additional output sample is an abnormal identifier and needs to be input into the knowledge graph for further judgment;
step S3.3: in the GCN training process, a cross entropy loss function needs to be calculated through the labeled samples in the training set, the function is an optimization target of the model, errors are made to propagate reversely, and the weight of a weight matrix in each graph convolution layer is optimized according to a gradient descent method:
where l is the number of labeled samples, c is the total number of previously defined network state classes, and Y is the label matrix of previously defined nodes.
7. The deep learning and knowledge-graph based cellular network fault diagnosis method according to claim 6, wherein the knowledge extraction algorithm for unstructured data based on LSTM and CRF models in step S4 comprises the following specific steps:
step S4.1: using a BIO labeling set adopted in Bakeoff-3 evaluation, namely B-P, I-P represents a failure first character and a failure non-first character, B-L, I-L represents a relationship first character and a relationship non-first character, B-O, I-O represents a failure cause entity first character and a failure cause non-first character, and O represents that the character does not belong to one part of a named entity;
step S4.2: preprocessing text and data, using one-hot encoding, a corpus containing all words in the data set, and then using the same direction as the total number of words in the corpusThe quantity represents each word, the value of the position corresponding to the word in the vector is 1, the rest positions are 0, and the character vector x of each character in the sentence is mapped by a one-hot vector i Inputting the sentence features into an LSTM model, automatically extracting sentence features, then utilizing a LSTM and CRF combined model to score the identification of the entity, and finally outputting a corresponding label.
8. The deep learning and knowledge-graph-based cellular network fault diagnosis method according to claim 7, wherein the final network fault diagnosis result and the interpretable report obtained in step S5 comprise:
step S5.1: if the GCN diagnosis model output fault is (0,0,0,1,0,0), the fault reason is indicated to be a coverage hole, and then the coverage hole is input into KG, so that the related fault definition, phenomenon, possible reason and solution can be obtained;
step S5.2: if the GCN diagnosis model outputs (0,1,0,1,0,0) a fault, the fault reason is possible to be a coverage hole or uplink interference or two reasons, the fault needs to be extracted and input into a fault knowledge graph for further diagnosis, the knowledge graph firstly searches detailed information about the uplink interference and the coverage hole, compares the detailed information with alarm information for troubleshooting, and then inputs the detailed information into a fault rule base for reasoning, so that a fault root factor is determined.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211519142.XA CN115734274A (en) | 2022-11-30 | 2022-11-30 | Cellular network fault diagnosis method based on deep learning and knowledge graph |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211519142.XA CN115734274A (en) | 2022-11-30 | 2022-11-30 | Cellular network fault diagnosis method based on deep learning and knowledge graph |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115734274A true CN115734274A (en) | 2023-03-03 |
Family
ID=85299399
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211519142.XA Pending CN115734274A (en) | 2022-11-30 | 2022-11-30 | Cellular network fault diagnosis method based on deep learning and knowledge graph |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115734274A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116358871A (en) * | 2023-03-29 | 2023-06-30 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
-
2022
- 2022-11-30 CN CN202211519142.XA patent/CN115734274A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116358871A (en) * | 2023-03-29 | 2023-06-30 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
| CN116358871B (en) * | 2023-03-29 | 2024-01-23 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110533631B (en) | SAR image change detection method based on pyramid pooling twin network | |
| CN110542819B (en) | Transformer fault type diagnosis method based on semi-supervised DBNC | |
| CN104572449A (en) | Automatic test method based on case library | |
| CN104155574A (en) | Power distribution network fault classification method based on adaptive neuro-fuzzy inference system | |
| CN110232434A (en) | A kind of neural network framework appraisal procedure based on attributed graph optimization | |
| CN112685504B (en) | Production process-oriented distributed migration chart learning method | |
| CN112039687A (en) | Small sample feature-oriented fault diagnosis method based on improved generation countermeasure network | |
| CN105260738A (en) | Method and system for detecting change of high-resolution remote sensing image based on active learning | |
| CN113283027A (en) | Mechanical fault diagnosis method based on knowledge graph and graph neural network | |
| CN105471647B (en) | A kind of power communication network fault positioning method | |
| CN108133225A (en) | A kind of icing flashover fault early warning method based on support vector machines | |
| CN112756759B (en) | Spot welding robot workstation fault judgment method | |
| CN101546290B (en) | Method for improving accuracy of quality forecast of class hierarchy in object-oriented software | |
| CN112147432A (en) | BiLSTM module based on attention mechanism, transformer state diagnosis method and system | |
| CN106991049A (en) | A kind of Software Defects Predict Methods and forecasting system | |
| CN111680875A (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN108304974A (en) | A kind of civil aviaton NOSHOW predictions based on optimization C5.0 and Apriori and strong factor-analysis approach | |
| CN110990718A (en) | Social network model building module of company image improving system | |
| CN112765415A (en) | Link prediction method based on relational content joint embedding convolution neural network | |
| CN112217674A (en) | Alarm root cause identification method based on causal network mining and graph attention network | |
| CN115734274A (en) | Cellular network fault diagnosis method based on deep learning and knowledge graph | |
| CN111586728A (en) | Small sample characteristic-oriented heterogeneous wireless network fault detection and diagnosis method | |
| CN113343581B (en) | Transformer fault diagnosis method based on graph Markov neural network | |
| CN116842459B (en) | Electric energy metering fault diagnosis method and diagnosis terminal based on small sample learning | |
| CN113989550A (en) | Electric vehicle charging pile operation state prediction method based on CNN and LSTM hybrid network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |