CN113709779B - Cellular network fault diagnosis method - Google Patents

Abstract

The invention discloses a cellular network fault diagnosis method, which comprises the following steps: Step 1, determining the network fault data set; Step 2, obtaining the network fault data set after dimension reduction; The fault data set is expressed in the form of a feature matrix; the label information of the network fault data set after dimensionality reduction in step 2 is expressed in the form of a label matrix; the introduced weight matrix is converted into an adjacency matrix with only 0 and 1 matrix elements; Step 4. Fault diagnosis based on graph convolutional neural network. This new cellular network fault diagnosis method deeply studies the intelligent fault diagnosis of heterogeneous wireless networks, combines the big data processing method to analyze the similarity between samples, converts the existing network fault parameter data set into graph structure data, and utilizes graph convolution The neural network extracts features from the graph structure data to complete the classification task for the sample nodes and predict the fault type of the cell.

Description

A method for diagnosing cellular network faults

Technical Field

The present invention relates to the field of communication technology, and in particular to a cellular network fault diagnosis method.

Background Art

In recent years, in the face of the surge in mobile data traffic and the different demands of various services, heterogeneous cellular networks have gradually become one of the important ways to improve system capacity. However, with the expansion of the scale and complexity of cellular networks, the operation and maintenance tasks of cellular networks have also become complicated and cumbersome. Although the end-to-end user experience in terms of throughput and latency has been significantly improved, cellular networks have also become more prone to failures. When failures occur or are about to occur, how to predict and locate them has become a huge challenge.

In the past, traditional network fault detection and diagnosis were mainly completed through manual operations, which made it difficult to accurately obtain the mapping relationship between network symptoms and fault categories, and also consumed a lot of manpower and material resources.

With the rapid development of artificial intelligence, most of the most commonly used intelligent fault diagnosis methods are based on machine learning. Intelligent fault diagnosis has changed from traditional model-based fault diagnosis methods to data-driven fault diagnosis methods. Network fault diagnosis technology based on machine learning learns the mapping relationship between network events by mining information from a large amount of training data, then establishes a fault diagnosis model based on these mapping relationships, and finally applies the trained model to the newly observed network symptoms to achieve the prediction of network status. However, most of the existing fault diagnosis solutions use supervised learning methods in machine learning. Most of these fault diagnosis methods rely on massive data sets and require that the samples in the data sets have sufficient labeled information. However, in actual situations, it is very difficult to obtain labeled information, labeling samples is often very time-consuming, the cost of manual category labeling is too high, and the historical data sets are too few and the acquisition cost is also expensive. Most of these fault diagnosis methods are based on fine-tuning and classification of supervised learning, and a large amount of unlabeled sample information cannot be fully utilized, and the unlabeled sample data is wasted.

In summary, in the process of 4G/5G heterogeneous wireless network fault diagnosis, how to perform accurate network fault diagnosis when there is very little effective marking information is a technical problem that needs to be urgently solved by technical personnel in this field.

Summary of the invention

The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a cellular network fault diagnosis method. The present invention combines the big data processing method to analyze the similar characteristics between samples, deeply studies the intelligent fault diagnosis algorithm of heterogeneous wireless networks, converts the existing network fault parameter data set into graph structure data, and uses the graph convolutional neural network to extract features from the graph data, completes the classification task of sample nodes, and predicts the fault type of the cell.

The present invention adopts the following technical solutions to solve the above technical problems:

A cellular network fault diagnosis method proposed in the present invention comprises the following steps:

Step 1: Determine the network fault data set; the details are as follows:

Step 101: The data in the network fault data set includes original feature parameters and label categories, and defines non-network fault conditions and five types of network faults as label categories, wherein the five types of network faults are uplink interference, downlink interference, coverage holes, air interface failures, and base station failures;

Step 102: all key performance indicators used in identifying a faulty cell are used as original characteristic parameters, and the original characteristic parameters include 16 characteristic parameters, including reference signal received power, reference signal received quality, uplink packet loss rate, downlink packet loss rate, uplink signal-to-noise ratio, downlink signal-to-noise ratio, radio resource control connection establishment success rate, evolved radio access bearer establishment success rate, call drop rate, handover success rate, uplink average throughput, downlink average throughput, node outgoing average throughput, node incoming average throughput, handover delay, and link bit error rate;

Step 2: Using the XGBoost algorithm, select the first n feature parameters that make the XGBoost algorithm have the highest diagnostic accuracy from the original feature parameters as the feature parameters of the data in the network fault data set, thereby obtaining a network fault data set after dimensionality reduction, 0＜n＜16;

Step 3: Graph data conversion, as follows:

Step 301: The network fault data set after the dimension reduction in step 2 is represented in the form of a feature matrix, where each row vector in the feature matrix corresponds to a feature parameter vector other than the category information in the network fault data set after the dimension reduction;

Step 302: The label information of the network fault data set after the dimension reduction in step 2 is represented in the form of a label matrix; in the label matrix, the label row vector of the labeled data is in the form of one-hot encoding, and the label row vector of the unlabeled data is a zero vector; the labeled data refers to the data with a label, and the unlabeled data is the data without a label;

Step 303: Map the network fault data set after dimensionality reduction into an undirected graph G = (V, E), where the undirected graph consists of two types of elements, namely, a node set V and an edge set E; introduce a weight matrix to represent the similarity between nodes in the undirected graph, where the elements in the weight matrix are the similarities between each node pair, and the similarity between each node pair is obtained by calculating the Euclidean distance between two nodes and normalizing them; then compare all elements in the weight matrix with the parameter threshold through a set parameter threshold, and if the current element is greater than the parameter threshold, set the element to 1, otherwise set it to 0; thereby converting the weight matrix into an adjacency matrix whose matrix elements are only 0 and 1, and the adjacency matrix represents the adjacent relationship between nodes;

Step 4: Fault diagnosis based on graph convolutional neural network, as follows:

Step 401: Adjust the parameters of the graph convolutional neural network and select a hierarchical structure to build a graph convolutional neural network model;

The feature matrix of step 402, step 301 and the label matrix of step 302 are the input of the graph convolutional neural network, which are used to train the parameters in the graph convolutional neural network model, and according to the trained graph convolutional neural network model, the adjacency matrix of step 303, and the propagation formula between layers, the high-order aggregate feature attributes of each node are obtained; the high-order aggregate feature attributes of each node are input into the Softmax layer in the graph convolutional neural network model to obtain the final fault classification diagnosis result.

As a further optimization scheme of the cellular network fault diagnosis method according to the present invention, step 2 is specifically as follows:

Step 201: Use the XGBoost algorithm to obtain the importance score of each feature parameter in the original feature parameters;

Step 202: sort the feature parameters in descending order according to the importance scores of the feature parameters;

Step 203: According to the original feature parameters sorted in step 202, the first n feature parameters that make the diagnosis accuracy of the XGBoost algorithm the highest are selected as the feature parameters of the data in the network fault data set, thereby obtaining a network fault data set after dimensionality reduction, 0＜n＜16;

As a further optimization scheme of a cellular network fault diagnosis method described in the present invention, the parameters in the graph convolutional neural network model include the number of graph convolutional layers, the probability size of the dropout layer, and the size of the filter matrix.

Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

(1) The present invention is based on a semi-supervised fault diagnosis algorithm, which reduces the amount of labeled data used for training;

(2) The present invention uses the XGBoost algorithm to solve the optimal feature parameter combination, which effectively avoids the problems of reducing the model training speed and generating dimensionality disaster;

(3) The complex nonlinear relationship between network key performance indicator parameters and fault types is fully utilized, and the similarity between samples is well utilized to further improve the performance;

(4) Rationally construct the graph convolutional neural network hierarchy and reasonably set the convolutional neural network parameters to improve the effectiveness and reliability of model fault diagnosis.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a 4G/5G heterogeneous wireless network scenario diagram of the present invention.

FIG. 2 is a structural diagram of data in the present invention.

FIG3 is a hierarchical structure diagram of the graph convolutional neural network in the present invention.

DETAILED DESCRIPTION

The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

This paper proposes a new intelligent fault diagnosis algorithm model based on graph convolutional neural network, which makes full use of the label information of labeled data and the characteristic parameter information of unlabeled data, and can achieve good diagnostic accuracy when only a small amount of labeled labeled data is used for training. The model can quickly detect network faults and further identify possible network fault types, thereby accelerating the recovery speed of faulty cells.

A 4G/5G heterogeneous wireless network scenario diagram is shown in Figure 1, and the content mainly includes the selection of the optimal feature combination. The present invention considers a heterogeneous wireless network scenario with overlapping coverage of macro cells, micro cells and femto cells. In this scenario, due to the diversity of the network, the system is more complex and network management is more difficult. The present invention considers the network fault diagnosis and prediction in this scenario. First, it is necessary to analyze the key performance indicators that measure network performance and common network faults. This part is the preliminary work for building a network fault diagnosis model.

1) Feature selection, including the following steps:

Step 101, based on the real data set obtained from the route test and combined with the actual situation, non-network fault conditions and five other common network faults are selected as label categories of the network fault data set, where the five network faults are uplink interference, downlink interference, coverage holes, air interface failures, and base station failures;

Step 102, the key performance indicator parameters used to identify the faulty cell are used as the original characteristic parameters of the sample data, including reference signal reception power, reference signal reception quality, uplink packet loss rate, downlink packet loss rate, uplink signal-to-noise ratio, downlink signal-to-noise ratio, radio resource control connection establishment success rate, evolved radio access bearer establishment success rate, call drop rate, switching success rate, uplink average throughput, downlink average throughput, node outgoing average throughput, node incoming average throughput, switching delay, link bit error rate, a total of 16 key performance indicator parameters.

2) Optimal feature combination selection. If the data samples in the data set contain too many feature parameters, it will not only be detrimental to the construction of the adjacency matrix of the subsequent graph, resulting in an unreasonable adjacency matrix, but also reduce the speed of model training in the fault diagnosis stage, causing dimensionality disaster and other problems. Therefore, it is necessary to select a small number of important network parameters to identify faulty cells based on the importance of network parameters. The following steps are included:

Step 201, use the XGBoost algorithm to select the optimal feature combination. XGBoost will create multiple boosted trees. In each iteration of training, XGBoost will build a new decision tree model in the gradient descent direction of the previous model loss function, so that the residual between the predicted value and the true value of the previous model is gradually reduced. The XGBoost algorithm can obtain an importance score for each feature parameter. The more times an attribute is used to build a decision tree in the model, the more important it is;

Step 202, using the feature importance sorting function of the XGBoost framework, sorting the 16 feature parameters in descending order according to the importance scores of the feature parameters;

Step 203, selecting the first n feature parameters that make the diagnosis accuracy of the XGBoost algorithm the highest from the original feature parameters sorted in step 202 as the feature parameters of the network fault data set after dimensionality reduction, 0＜n＜16;

3) Graph data conversion, the data in the data set needs to be converted into graph data, a non-Euclidean data, so as to conform to the input format of the graph convolutional neural network. As shown in Figure 2, the method proposed in the present invention first needs to map the original network fault data set into an undirected graph G = (V, E), and the graph consists of two types of elements, namely the node set V and the edge set E. In the present invention, each sample data in the original data set corresponds to a node in the graph, each node has its own attribute characteristics, and the similarity measure between nodes can be measured by the edge weight. It includes the following steps:

Step 301, the characteristic parameters of the sample data in the network fault data set after dimensionality reduction are expressed in the form of a characteristic matrix, and each row vector in the characteristic matrix corresponds to a characteristic parameter vector formed by removing the category information from a piece of sample data in the network fault data set after dimensionality reduction;

Step 302: The label information of the samples in the original data set is represented in the form of a label matrix. The label information of the labeled samples in the matrix is represented by a row vector similar to one-hot encoding, which corresponds to the different types of faults currently suffered by the cell. The row vector corresponding to the unlabeled data is a zero vector, indicating that the label information is unknown.

其中，w_ij表示节点x_i和x_j之间的相似性度量，x_i和x_j分别表示第i和第j两个不同的样本，δ被称为高斯带宽参数，可根据实际情况自主定义。然后，设置合理的参数阈值α，将权重矩阵中的所有元素与该阈值α进行比较，值大于该阈值的元素w_ij，重新置为1，否则全部置为0。于是，把权重矩阵转换为元素只有0和1的邻接矩阵A，其中0表示两个样本之间没有关系，而1表示两个样本之间相似性程度很高。Step 303, first construct a weight matrix to record the similarity between nodes in the graph. The elements w _ij in the weight matrix, i.e., the similarity between each node pair, are obtained by calculating the Euclidean distance between each pair and normalizing them, i.e.

Among them, w _ij represents the similarity measure between nodes _xi and x _j , _xi and x _j represent two different samples, δ, which is called the Gaussian bandwidth parameter and can be defined according to the actual situation. Then, a reasonable parameter threshold α is set, and all elements in the weight matrix are compared with the threshold α. The elements w _ij with values greater than the threshold are reset to 1, otherwise all are reset to 0. Thus, the weight matrix is converted into an adjacency matrix A with only 0 and 1 elements, where 0 indicates that there is no relationship between the two samples, and 1 indicates that the similarity between the two samples is very high.

4) Fault diagnosis based on graph convolutional neural network. In order to improve the accuracy of network fault diagnosis, it is necessary to determine the optimal neural network structure and parameters before training GCN. In order to illustrate the structure and workflow of GCN, the present invention shows a GCN model defined by the present invention as shown in FIG3, which includes the following steps:

Step 401, select reasonable GCN parameters and hierarchical structure to construct a graph convolutional neural network. The structure and parameter selection of GCN mainly include the number of graph convolution layers, the size of the filter matrix of each graph convolution layer, and the probability size of the dropout layer. According to actual conditions, it is best when the number of graph convolution layers is 2. At this time, it can not only reasonably aggregate the feature parameter information of the neighboring nodes of each central node, but also avoid the overfitting problem caused by too many training parameters due to too many convolution layers. The probability size of the dropout layer is generally selected to be 0.25 or 0.5. The size of the filter matrix needs to be determined according to the size of the previous convolution layer, and factors such as attribute dimensionality reduction in the operation must also be considered;

通过划分训练集，训练图卷积神经网络模型，利用训练完成的GCN模型去聚合各节点以及邻居节点的特征参数，得到各节点的高阶聚合特征属性。其中，H^(l+1)表示当前图卷积层输出的各节点的高阶聚合特征属性所构成的矩阵，H^(l)是之前第l层的输出，

表示步骤303中的邻接矩阵加了自环，

是关于

的度矩阵，σ表示激活函数，W^(l)表示第l层图卷积层的可训练权重矩阵；Step 402, after GCN is constructed, the feature matrix in step 301 and the label matrix in step 302 are used as the input of GCN. According to the propagation formula between layers defined by GCN,

By dividing the training set and training the graph convolutional neural network model, the trained GCN model is used to aggregate the feature parameters of each node and its neighboring nodes to obtain the high-order aggregate feature attributes of each node. Among them, H ^(l+1) represents the matrix composed of the high-order aggregate feature attributes of each node output by the current graph convolution layer, and H ^(l) is the output of the previous lth layer.

It means that the adjacency matrix in step 303 has self-loops added.

About

, σ represents the activation function, and W ^(l) represents the trainable weight matrix of the l-th graph convolutional layer;

Step 403, input the obtained high-order aggregated feature attributes of each node into the Softmax layer to obtain the final fault classification diagnosis result. The Softmax activation function needs to be applied to each row of the feature matrix output by the last graph convolution layer. The diagnosis result will be represented by the node-level feature matrix Z, and each row in the matrix Z is similar to the network fault corresponding to the row vector of the one-hot encoding, which represents the different types of network faults suffered by the current cell.

The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims (3)

1. A cellular network fault diagnosis method, is characterized in that, comprises the following steps: Step 1. Determine the network fault data set; the details are as follows: Step 101, the data in the network fault data set includes original characteristic parameters and label categories, and defines non-network fault conditions and 5 types of network faults as label categories, wherein the 5 types of network faults are uplink interference, downlink interference, coverage holes, and air interfaces Faults and base station failures; Step 102, using all key performance indicators used when identifying the faulty cell as original characteristic parameters, the original characteristic parameters include 16 characteristic parameters, and these 16 characteristic parameters include reference signal received power, reference signal received quality, uplink packet loss rate, downlink packet loss rate, uplink signal-to-noise ratio, downlink signal-to-noise ratio, radio resource control connection establishment success rate, evolved radio access bearer establishment success rate, call drop rate, handover success rate, uplink average Throughput, downlink average throughput, node outgoing average throughput, node incoming average throughput, switching delay and link bit error rate; Step 2. Using the XGBoost algorithm, select the first n characteristic parameters that make the diagnosis accuracy of the XGBoost algorithm the highest from the original characteristic parameters as the characteristic parameters of the data in the network fault data set, so as to obtain the network fault data set after dimensionality reduction, 0< n<16; Step 3, graph data conversion, as follows: Step 301, express the network fault data set after dimension reduction in step 2 in the form of a feature matrix, and each row vector in the feature matrix corresponds to a feature parameter in the network fault data set after dimension reduction except category information vector; Step 302, express the label information of the network fault data set after dimension reduction in step 2 in the form of a label matrix; in the label matrix, the label row vectors of labeled data are in the form of one-hot encoding, while the label row vectors of unlabeled data Then it is a zero vector; labeled data refers to data with labels, and unlabeled data refers to data without labels; Step 303, map the network fault data set after dimensionality reduction into an undirected graph G=(V, E), the undirected graph is composed of two types of elements, namely node set V and edge set E; introduce a weight matrix To represent the similarity between nodes in the undirected graph, the elements in the weight matrix are the similarity between each node pair, and the similarity between each node pair is calculated by the Euclidean distance between two nodes and normalized Obtained; then compare all the elements in the weight matrix with the parameter threshold through the set parameter threshold, if the current element is greater than the parameter threshold, set the element to 1, otherwise set to 0; thus transform the weight matrix It is an adjacency matrix whose matrix elements only have 0 and 1, and the adjacency matrix represents the adjacency relationship between nodes; Step 4. Fault diagnosis based on the graph convolutional neural network, as follows: Step 401, adjusting the parameters of the graph convolutional neural network and selecting a hierarchical structure to construct a graph convolutional neural network model; Step 402, the feature matrix of step 301 and the label matrix of step 302 are the input of the graph convolutional neural network, which are used to train the parameters in the graph convolutional neural network model, and according to the trained graph convolutional neural network model, step 303 The adjacency matrix and the propagation formula between layers are used to obtain the high-order aggregate feature attributes of each node; the high-order aggregate feature attributes of each node are input into the Softmax layer in the graph convolutional neural network model to obtain the final fault classification and diagnosis result. 2. A kind of cellular network fault diagnosis method according to claim 1, is characterized in that, step 2 is specifically as follows: Step 201, using the XGBoost algorithm to obtain the importance score of each feature parameter in the original feature parameters; Step 202, sort the feature parameters in descending order according to the importance scores of the feature parameters; Step 203, according to the original characteristic parameters sorted in step 202, select the first n characteristic parameters that make the diagnosis accuracy of the XGBoost algorithm the highest as the characteristic parameters of the data in the network fault data set, so as to obtain the network fault data after dimension reduction Set, 0<n<16. 3. A kind of cellular network fault diagnosis method according to claim 1, is characterized in that, the parameter in the graph convolution neural network model comprises graph convolution layer number of layers, dropout layer probability size and filter matrix size.

Images