A Fault Diagnosis Method for Cellular Networks Based on the Fusion of Knowledge and Data

Abstract

The present invention discloses a cellular network fault diagnosis method based on knowledge and data fusion. The method uses a method based on a generative adversarial network (GAN) to perform data augmentation on a real dataset, effectively balancing the dataset and reducing the cost of labeling data. The naive Bayes method is used in combination with expert knowledge to perform a pre-diagnosis classification task on the network fault dataset, and an association graph between data is generated according to the pre-diagnosis results, effectively solving the problem of inaccurate topological association graphs between data generated by the GCN. The generated topological association graph and the training dataset are both input into the GCN model for model training, and the GCN model is improved so that it can separately adjust the influence of pre-diagnosis prior knowledge and the scale of the training dataset on the accuracy of the GCN model during training.

Description

Technical Field

[0001] The present invention relates to the technical field of communication networks, and mainly relates to a cellular network fault diagnosis method based on knowledge and data fusion. 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. Fault diagnosis is one of the main tasks for managing any network.

[0003] CN113709779A, a cellular network fault diagnosis method, discloses a cellular network fault diagnosis method, which converts the introduced weight matrix into an adjacency matrix with matrix elements only 0 and 1; studies the intelligent fault diagnosis of heterogeneous wireless networks, analyzes the similarity characteristics between samples by combining big data processing methods, converts the existing network fault parameter data set into graph structure data, and uses a graph convolutional neural network to extract features from the graph structure data, so as to complete the classification task of sample nodes and predict the fault type of the cell.

[0004] 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 manpower 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 hazards such as service interruption and network paralysis caused by fault propagation, and 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.

[0005] 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 initially applied to the field of mechanical fault diagnosis. The network fault diagnosis method based on GCN first needs to obtain the characteristic attributes of the data when obtaining the topological association graph between data, and then calculates the similarity according to the characteristic attributes to determine the topological association graph of the data set. In fact, the topological association graph formed by the similarity of the characteristic attributes between data based on the spectral clustering idea is relatively rough and has no practical significance. While lacking interpretability, it is difficult to further improve the classification accuracy of the GCN model. Summary of the Invention

[0006] Object of the Invention: The present invention provides a method for diagnosing faults in a cellular network based on the fusion of knowledge and data, which can not only reduce the cost of labeling data, but also improve the accuracy of the fault diagnosis model.

[0007] Technical Solution: To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for diagnosing faults in a cellular network based on the fusion of knowledge and data, comprising the following steps:

[0009] Step S1: Collect a labeled network state 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:

[0010] Step S1.1: Obtain the importance scores of each feature through the feature importance ranking function of XGBoost, and perform a descending order sorting.

[0011] 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 rate of the XGBoost model under different feature combinations.

[0012] Step S1.3: Weigh the model accuracy rate and the number of features to obtain an optimal subset of network feature parameters.

[0013] Step S2: Input the preprocessed data set in Step S1 into a generative adversarial network to generate a simulated data set representing different network states with labels, and summarize it with the preprocessed data set in Step S1.

[0014] Step S3: Use the KPI discretization rule to perform a discretization operation on the optimal feature subset obtained by the XGBoost algorithm in Step S1. Use expert knowledge to perform a discretization operation on the KPI. According to the discretized KPI attributes, train a naive Bayes classifier by reasonably dividing the training data set, and use the trained naive Bayes model to classify the remaining data to obtain a preliminary diagnosis result. Use the preliminary diagnosis result set to construct a topological association graph, that is, an adjacency matrix.

[0015] Step S4: According to the obtained topological association graph, improve the GCN on the basis of the original GCN model, and use the improved GCN to obtain the final network fault diagnosis result.

[0016] Further, the specific steps for generating a simulated data set based on the generative adversarial network in Step S2 are as follows:

[0017] Step S2.1. The selected generative adversarial network model is WGAN-GP, and the specific optimization objective is as follows:

[0018]

[0019] D(s) is a scalar representing the probability that s comes from the real data distribution rather than p g . Among them represents the distribution that the data generated by the generator follows, is the distribution that the real data follows, which refers to the data distribution under different network states here. is sampled from the entire data set composed of real data and generated data. Using ∈ that follows a uniform distribution between [0, 1], and then randomly interpolating and sampling on the connection line between and to obtain as the penalty term. In the penalty term, it is hoped that is closer to 1, and the penalty is less. λ is the penalty parameter.

[0020] Step S2.2. The specific steps for generating a simulated data set based on WGAN-GP are as follows:

[0021] (1) Use two fully connected neural networks to form a generator and a discriminator respectively.

[0022] (2) Train the generator so that the generator imitates the small amount of labeled data sets with different network states collected in the heterogeneous wireless network environment respectively to generate simulated data.

[0023] (3) Input the small amount of labeled network state data sets collected from the heterogeneous wireless network environment and the simulated data generated by the generator into the discriminator respectively for training the discriminator.

[0024] (4) Alternately and iteratively train the generator and the discriminator until balance is finally achieved and the model converges.

[0025] (5) Generate a simulated data set with labels representing different network states.

[0026] Furthermore, the specific steps for generating a pre-diagnosis result based on Naive Bayes and obtaining an adjacency matrix in step S3 are as follows:

[0027] Step S3.1. According to the discretized KPI attributes, train a Naive Bayes classifier by reasonably planning the training data set;

[0028] Step S3.2. Use the trained Naive Bayes model combined with expert knowledge to classify the remaining data to obtain a pre-diagnosis result label set Where N represents the total number of samples in the dataset after expansion by WGAN-GP. The specific classification steps are as follows:

[0029] Select the network fault category with the maximum posterior probability as the network fault h suffered by the current network, i.e.: * (X), that is:

[0030]

[0031] To avoid underflow errors, convert the above formula to logarithmic form:

[0032]

[0033] In this chapter, the Laplace smoothing method is used to estimate the prior probability P(y i ) and the conditional probability P(x j |y i ), that is:

[0034]

[0035]

[0036] Where D t represents the total number of samples contained in the training dataset; is the total number of samples in the training set under the network fault y i condition; represents the total number of samples in the training set under the network fault y i condition and the value of the jth KPI parameter is x j ; L is the total number of network fault categories defined previously; S j is the number of all possible values of the jth KPI.

[0037] Step S3.2. Based on the pre-diagnosis result set to construct a topological association graph, that is, the adjacency matrix A. The specific content is as follows:

[0038] In the pre-diagnosis result set , the data diagnosed as having the same network fault type are connected to each other in the graph, while the data with different network fault types are not connected to each other, that is:

[0039]

[0040] Furthermore, the specific content of step S4 based on the adjacency matrix A and using the improved GCN to obtain the final network fault diagnosis result is:

[0041] First, convert the feature parameter vector of the data samples in the dataset into n×d 0The feature matrix X of dimension, and then construct the adjacency matrix A of dimension n×n according to the similarity between nodes, and use X and A as the input of GCN.

[0042] input=(X, A)

[0043] The forward excitation propagation formula defined in GCN is:

[0044]

[0045] In the original GCN model theory, the matrix is directly obtained by adding the adjacency matrix A of the dataset and the identity matrix of the same size, that is:

[0046]

[0047] Improve the original GCN by adding a weight coefficient λ. λ is used to control the influence of the prior knowledge of pre-diagnosis obtained by Naive Bayes and the scale of the training dataset on the model accuracy respectively

[0048]

[0049] Therefore, the propagation formula of the improved GCN model is:

[0050]

[0051] Among them, σ is the activation function, is the degree matrix of the matrix Each element on its main diagonal is obtained by summing all the elements in the corresponding row of the matrix So The formation of completely depends on 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 GCN, they can be updated according to the error backpropagation and the gradient descent method. H (l) is the input node feature matrix of the l-th graph convolutional layer. For the input layer, H (0) is equal to the initial node feature matrix.

[0052] The output of the graph convolutional neural network is a node feature matrix where c is the number of predefined network fault categories. For the output result matrix Z = [Z 1 , Z 2 ,..., Z n , its form is similar to the label matrix Y. Each row vector Z i(1 ≤ i ≤ n) corresponds to the sample node x in the original dataset i to predict the final network fault category. Specifically, for the row vector Z i = [Z i,1 , Z i,2 ,..., Z i,c , the predicted label of the sample node x i is

[0053] During the GCN training process, finally, the cross-entropy loss function needs to be calculated through the labeled samples in the training set, and the error is backpropagated. According to the gradient descent method, the weights of the weight matrices in each graph convolutional layer are optimized.

[0054]

[0055] 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.

[0056] Beneficial effects:

[0057] The present invention proposes a knowledge and data fusion-based cellular network fault diagnosis method. The method of using a generative adversarial network is used to augment real data, reducing the cost spent on labeling data. The naive Bayes method is used to combine expert knowledge to perform a pre-diagnosis classification task on the network fault dataset, and a topological association graph is constructed. The generated topological association graph and the training dataset are simultaneously input into the improved GCN model for model training. This method combines the advantages of pre-diagnosis prior knowledge and deep learning. The model is superior to the individual naive Bayes algorithm and GCN algorithm, achieving better diagnostic accuracy.

[0058] The present invention proposes a knowledge and data fusion-based cellular network fault diagnosis method, which solves the problem that the historical data obtained from the real network is not rich enough to result in an unsatisfactory effect in constructing a diagnostic system. It solves the problem of the difficulty in selecting the number of neural network layers for network fault diagnosis based on GCN and the problem of insufficient model accuracy. This can not only greatly save the time of manually labeling training data but also greatly improve the accuracy of the fault diagnosis model. Brief description of the drawings

[0059] Figure 1 is a schematic diagram of a dense heterogeneous cellular network scenario;

[0060] Figure 2 is the flowchart of the knowledge and data fusion-based cellular network fault diagnosis provided by the present invention;

[0061] Figure 3 Feature attribute importance ranking diagram;

[0062] Figure 4 It is a data preprocessing flow chart;

[0063] Figure 5 It is a graph of a generative adversarial network model;

[0064] Figure 6 Graph convolutional neural network model. DETAILED DESCRIPTION

[0065] The present invention will be further described below in conjunction with the accompanying drawings.

[0066] The present invention Figure 1 The dense heterogeneous wireless network scenario shown is composed of high-power macro base stations and low-power micro base stations with a multi-level network structure. In this scenario, due to the diversity of the network, the system becomes more complex and network management becomes more difficult. The present invention considers network fault detection and diagnosis in this scenario. First, the causes that may cause the fault to occur are analyzed for the specific network scenario, and useful network parameters are screened out. This part is the work that must be done in the early stage of building a network fault diagnosis model. Then, historical data is obtained from the heterogeneous wireless network history database, including a fault category variable set and a fault variable set and its key performance indicator KPI.

[0067] based on Figure 1 The dense heterogeneous cellular network shown in the figure, the present invention proposes a cellular network fault diagnosis method based on knowledge and data fusion, the specific steps are as follows:

[0068] Step S1: Collect a network status data set with labels 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:

[0069] Step S1.1, obtain the importance score of each feature through the feature importance sorting function of XGBoost, and sort them in descending order;

[0070] Step S1.2, XGBoost continuously increases the feature selection threshold according to the importance score, retains the feature parameters with scores higher than the threshold, and discards those with scores lower than the threshold, thereby obtaining the accuracy of the XGBoost model under different feature combinations.

[0071] Step S1.3: Weigh the model accuracy and the number of features to obtain the optimal subset of network feature parameters.

[0072] Step S2, inputting the preprocessed data set in step S1 into the generative adversarial network to generate a simulated data set with labels representing different network states, and combining it with the preprocessed data set in step S1;

[0073] Step S2.1: The selected generative adversarial network model is WGAN-GP, and the specific optimization objective is as follows:

[0074]

[0075] D(s) is a scalar representing the probability that s comes from the real data distribution rather than p g where represents the distribution that the data generated by the generator follows, is the distribution that the real data follows, which refers to the data distribution under different network states here. is sampled from the entire dataset composed of real data and generated data. Using ∈ that follows a uniform distribution between [0, 1], and then randomly interpolating and sampling on the line connecting and to obtain As a penalty term, in the penalty term, it is hoped that is closer to 1, and the penalty is less. λ is the penalty parameter.

[0076] Step S2.2: The specific steps for generating a simulated dataset based on WGAN-GP are as follows:

[0077] (1) Use two fully connected neural networks to form the generator and discriminator respectively.

[0078] (2) Train the generator so that the generator imitates the small amount of labeled datasets with different network states collected in the heterogeneous wireless network environment respectively to generate simulated data.

[0079] (3) Input the small amount of labeled network state datasets collected from the heterogeneous wireless network environment and the simulated data generated by the generator into the discriminator respectively for discriminator training.

[0080] (4) Alternately and iteratively train the generator and discriminator until balance is finally achieved and the model converges.

[0081] (5) Generate a simulated dataset with labels representing different network states.

[0082] Step S3: Use expert knowledge to discretize the KPIs. According to the discretized KPI attributes, train a Naive Bayes classifier by reasonably partitioning the training dataset, and use the trained Naive Bayes model to classify the remaining data to obtain a pre-diagnosis result. Use the pre-diagnosis result set to construct a topological association graph, that is, an adjacency matrix;

[0083] Step S3.1: According to the discretized KPI attributes, train a Naive Bayes classifier by reasonably planning the training dataset;

[0084] Step S3.2: Classify the remaining data by using the trained Naive Bayes model combined with expert knowledge to obtain a pre-diagnosis result label set where N represents the total number of samples in the dataset after expansion by WGAN-GP. The specific classification steps are as follows:

[0085] Select the network fault category with the maximum posterior probability as the network fault h * (X) suffered by the current network, that is:

[0086]

[0087] To avoid underflow errors, convert the above formula to logarithmic form:

[0088]

[0089] In this chapter, the Laplace smoothing method is used to estimate the prior probability P(y i ) and the conditional probability P(x j |y i ), that is:

[0090]

[0091]

[0092] where D t represents the total number of samples contained in the training dataset; is the total number of samples in the training set in the case of network fault y i ; represents the total number of samples in the training set in the case of network fault y i and the value of the j-th KPI parameter is x j ; L is the total number of network fault categories defined previously; S j is the number of all possible values of the j-th KPI.

[0093] Step S3.2: Construct a topological association graph, that is, an adjacency matrix A, based on the pre-diagnosis result set The specific content is as follows:

[0094] In the pre-diagnosis result set , the data diagnosed as having the same network fault type are connected to each other in the graph, while the data with different network fault types are not connected to each other, that is:

[0095]

[0096] Step S4: Based on the obtained topological association graph, improve the GCN on the basis of the original GCN model, and use the improved GCN to obtain the final network fault diagnosis result.

[0097] In the theory of the original GCN model, the matrix is directly obtained by adding the adjacency matrix A of the dataset and the identity matrix of the same size, that is:

[0098]

[0099] Improve the original GCN by adding a weight coefficient λ. λ is used to control the influence of the pre-diagnosis prior knowledge obtained by Naive Bayes and the scale of the training dataset on the model accuracy respectively

[0100]

[0101] Therefore, the propagation formula of the improved GCN model is:

[0102]

[0103] Among them, is the degree matrix of the matrix Each element on its main diagonal is obtained by summing all the elements in the corresponding row of the matrix So is completely dependent on The output of the neural network is completed through the forward excitation propagation of the input, while the weight W (l) is updated according to the error backpropagation using batch gradient descent.

[0104] To illustrate the effectiveness of the method proposed in the present invention, an example is given below.

[0105] Step S1: Collect a labeled network state dataset from a dense heterogeneous cellular network environment and perform data preprocessing. The specific framework is as Figure 4 shown. Select the optimal subset from the dataset through the XGBoost algorithm;

[0106] First, obtain the scores of each feature through the feature importance ranking function of XGBoost, and then perform a descending order sorting. The feature importance ranking is as Figure 3As shown below, feature screening is then carried out. XGBoost will continuously increase the feature selection threshold according to the obtained importance scores of network parameters, retain the feature parameters with importance scores higher than this threshold, and discard the others, so as to obtain the accuracy of the XGBoost model under different feature combinations. Finally, the model accuracy and the number of features are weighed to obtain the optimal subset of network feature parameters. The diagnostic accuracies of the XGBoost model under different numbers of features are shown in Table 1.

[0107] Table 1 Diagnostic accuracies of the model under different numbers of features

[0108] Feature selection threshold Number of features Accuracy rate 190 14 86.54％ 199 13 86.76％ 209 12 86.60％ 230 11 86.52％ 253 10 86.12％ 266 9 86.30％ 281 8 86.70％ 284 7 85.44％ 302 6 84.35％ 302 5 83.39％ 331 4 81.53％ 376 3 81.17％ 449 2 79.60％ 715 1 77.01％

[0109] It can be seen that when 8 features are selected, the model can obtain better diagnostic accuracy and achieve the purpose of feature screening. Therefore, select Figure 3 the top 8 KPI parameters in the ranking as the KPI parameters after feature screening.

[0110] Step S2: Input the preprocessed dataset in Step S1 into the generative adversarial network to generate a simulated dataset with labels representing different network states. The framework of the generative adversarial network is as Figure 5 shown. And summarize the generated simulated dataset with the preprocessed dataset in Step S1;

[0111] The real network dataset collected in this example has a total of 817 labeled data with labels, including 8 different categories of network fault categories. By using WGAN-GP, the scale of the original real dataset is expanded to about three times the original, including a total of 2657 labeled data. It should be noted that when expanding the data under each network fault category, try to make the proportion of the number of samples under each category in the total number of samples the same, so as to make the category sample distribution more uniform. Combine the generated simulated data with the real data in the original dataset to obtain the expanded dataset. The expanded dataset is shown in Table 2.

[0112] Table 2 Data distribution of the dataset after expansion using WGAN-GP

[0113] Serial number Fault type Number of samples 1 In-building distribution leakage 347 2 Abnormal measurement threshold 342 3 Large site spacing 239 4 Modulo-3 interference 356 5 Abnormal handover threshold 300 6 Pilot pollution 214 7 Overlapping coverage 413 8 Missing neighboring cell 446

[0114] Step S3: Use expert knowledge to discretize the KPI. According to the discretized KPI attributes, train the naive Bayes classifier by reasonably dividing the training dataset, and use the trained naive Bayes model to classify the remaining data to obtain the preliminary diagnosis results. Use the preliminary diagnosis result set to construct a topological association graph, that is, an adjacency matrix;

[0115] Discretize the KPI feature attributes in the dataset according to the rules in Table 3. This is beneficial for calculating the likelihood function based on the occurrence frequencies of KPI values under different network fault states in the training dataset through statistical counting, and finally obtaining the pre-diagnosis classification results.

[0116] Table 3 KPI Discretization Rules

[0117]

[0118]

[0119] Use the trained Naive Bayes model combined with expert knowledge to classify the remaining data and obtain the pre-diagnosis result label set. Where N represents the total number of samples in the dataset after expansion by WGAN-GP.

[0120] Based on the pre-diagnosis result set Construct a topological association graph, that is, an adjacency matrix A. The specific content is as follows:

[0121] In the pre-diagnosis result set Data diagnosed as having the same network fault type are connected to each other in the graph, while data with different network fault types are not connected to each other, that is:

[0122]

[0123] Step S4: According to the obtained topological association graph, improve the GCN on the basis of the original GCN model, and use the improved GCN to obtain the final network fault diagnosis result.

[0124] In the theory of the original GCN model, the matrix is directly obtained by adding the adjacency matrix A of the dataset and the identity matrix of the same size, that is:

[0125]

[0126] Improve the original GCN by adding a weight coefficient λ. λ is used to control the influence of the pre-diagnosis prior knowledge obtained from Naive Bayes and the scale of the training dataset on the model accuracy respectively.

[0127]

[0128] Therefore, the propagation formula of the improved GCN model is:

[0129]

[0130] Among them, is the matrix The degree matrix, where each element on the main diagonal is obtained by summing all elements in the corresponding row of the matrix , so is formed entirely depending on The output of the neural network is completed through forward excitation propagation of the input, while the weights W (l) are updated according to backpropagation of error using batch gradient descent. Among them, λ is a weight coefficient positively correlated with the size of the training set, and in this example, it is specifically defined as λ = 1 + re r , where r represents the proportion of the labeled training set in the total dataset size.

[0131] The topological association graph of the data obtained according to the pre-diagnosis results has good characteristics, solving the problem of selecting the number of graph convolutional layers in the GCN structure. Therefore, in this example, the hidden layer depth of all GCN models is set to 2, the learning rate is set to 0.01, the probability of the dropout layer is set to 0.25, the maximum number of iterations for training the neural network is 200, and the L2 regularization parameter is set to 1×10 -5 . The output of the neural network is completed through forward excitation propagation of the input, while the weights W (0) and W (1) are updated according to backpropagation of error using batch gradient descent.

[0132] Figure 6 Figure [ID] shows a GCN model, which mainly includes two graph convolutional layers. For the convenience of explanation, the 0th graph convolutional layer in the actual GCN model is called the 1st graph convolutional layer, and so on.

[0133] First, calculate where represents the normalized symmetric adjacency matrix. Next, perform a weighting operation by multiplying with the trainable weight matrix W (0) to obtain a new set of node features Finally, select an activation function for the new feature matrix to obtain the output feature matrix H (1) of the first graph convolutional layer, that is, the new node feature representation learned by the first graph convolutional layer:

[0134]

[0135] Since stacking multiple graph convolutional layers can aggregate the feature attribute information of neighboring nodes in higher-order neighborhoods. Therefore, the output H (1) of the previous graph convolutional layer is used as the input of the second graph convolutional layer. After passing through the second graph convolutional layer, another set of node features

[0136] Finally, the feature matrix is input into the SoftMax activation function for processing, and the finally output node feature matrix is:

[0137]

[0138] where W (1) is the weight matrix of the second graph convolutional layer. The SoftMax activation function needs to be applied to each row of the feature matrix .

[0139] The output of the graph convolutional neural network is a node feature matrix where c is the number of predefined network fault categories. For the output result matrix Z = [Z 1 , Z 2 ,..., Z n , its form is similar to the label matrix Y. Each row vector Z i (1 ≤ i ≤ n) in Z corresponds to the predicted final network fault category of the sample node x i in the original dataset. Specifically, for the row vector Z i = [Z i,1 , Z i,2 ,..., Z i,c , the predicted label of the sample node x i is

[0140] During the GCN training process, finally, the cross-entropy loss function needs to be calculated using the labeled samples in the training set, and the error is backpropagated. According to the gradient descent method, the weights of the weight matrices in each graph convolutional layer are optimized.

[0141]

[0142] 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.

[0143] So far, by using the improved GCN model, the final network fault diagnosis task is completed.

[0144] The above are only the preferred embodiments of the present invention. It should be noted that: for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A cellular network fault diagnosis method based on knowledge and data fusion, It is characterized in that The following steps are involved: Step S1, collecting a network status data set with labels from a dense heterogeneous cellular network environment, and selecting an optimal subset from the data set through the XGBoost algorithm; Step S2, inputting the network state data set with labels preprocessed in step S1 into the generative adversarial network to obtain simulated data sets with labels representing different network states, and combining them with the data set preprocessed in step S1; Step S3: Use the KPI discretization rule to discretize the optimal feature subset obtained by the XGBoost algorithm in step S1, train the naive Bayes classifier according to the discretized KPI attributes by reasonably dividing the training data set, and use the trained naive Bayes model to classify the remaining data to obtain the pre-diagnosis result; Use the pre-diagnosis result set to construct a topological association graph, i.e., an adjacency matrix; Step S4. Based on the obtained topological association graph, improve the GCN on the basis of the original GCN model, and use the improved GCN to obtain the final network fault diagnosis result, which specifically includes first converting the feature parameter vector of the data samples in the dataset into a feature matrix X of n×d 0 dimensions, then constructing an adjacency matrix A of n×n dimensions according to the similarity between nodes, and taking X and A as the inputs of the GCN; input=(X,A) The forward excitation propagation formula defined in GCN is: The original GCN is improved by increasing the weight coefficient λ, which is used to control the impact of the pre-diagnosis prior knowledge derived from naive Bayes and the size of the training data set on the model accuracy. The propagation formula of the improved GCN model is: where σ is the activation function, is the degree matrix of the matrix and each element on its main diagonal is obtained by summing all the elements in the corresponding row of the matrix ; 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; The output of the graph convolutional neural network is a node feature matrix where c is the number of predefined network fault categories; for the output result matrix Z = [Z 1 , Z 2 ,..., Z n , each row vector Z i (1 ≤ i ≤ n) in Z corresponds to the predicted final network fault category of the sample node x i in the original dataset; specifically, for the row vector Z i = [Z i,1 , Z i,2 ,..., Z i,c , the predicted label of the sample node x i is In the GCN training process, the cross entropy loss function is calculated through the labeled samples in the training set, and the error is propagated backwards. The weights of the weight matrix in each graph convolution layer are optimized according to the gradient descent method. Among them, l refers to the number of labeled samples, c is the total number of network state categories, and Y is the label matrix of the node.

2. A cellular network fault diagnosis method based on knowledge and data fusion according to claim 1, It is characterized in that The specific method of selecting the optimal subset from the data set by the XGBoost algorithm in step S1 is as follows: Step S1.1, obtain the importance score of each feature through the feature importance sorting function of XGBoost, and sort them in descending order; Step S1.2, XGBoost continuously increases the feature selection threshold according to the importance score, retains the feature parameters with scores higher than the threshold, and discards them otherwise, 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 subset of network feature parameters.

3. A cellular network fault diagnosis method based on knowledge and data fusion according to claim 1, It is characterized in that The specific steps of obtaining the simulated data set with labels representing different network states in step S2 are: Step S2.1: The selected generative adversarial network model is WGAN-GP, and the optimization objectives are: D(s) is a scalar representing that s comes from the true data distribution; where represents the distribution followed by the data generated by the generator, is the distribution followed by the true data; is sampled from the entire dataset composed of the true data and the generated data. Using ∈ that follows a uniform distribution between [0, 1], and then and randomly interpolating and sampling on the line connecting them to obtain as a penalty term. In the penalty term, the closer it is to 1, the less the penalty. λ is the penalty parameter; Step S2.2: The specific steps of generating a simulated data set based on WGAN-GP are as follows: (1) Two fully connected neural networks are used to form the generator and the discriminator respectively; (2) training the generator to simulate different network status data sets with labels collected in heterogeneous wireless network environments to generate simulated data; (3) Input the labeled network state data set collected from the heterogeneous wireless network environment and the simulated data generated by the generator into the discriminator respectively for the training of the discriminator; (4) Conduct alternating iterative training on the generator and the discriminator until balance is finally achieved and the model converges; (5) Generate a simulated data set with labels representing different network states.

4. A method for diagnosing cellular network faults based on knowledge and data fusion according to claim 1, characterized in that, the specific steps for obtaining the preliminary diagnosis result in step S3 are as follows: Step S3.1: Train a naive Bayes classifier with the training data set according to the discretized KPI attributes; Step S3.2, Rule-based network fault diagnosis method. Assume that the network fault set C = {C 1 , C 2 ,..., C M , UN}, where the first M records represent M types of network faults stored in the historical database, UN represents the unrecognized network fault, and S = [KPI 1 , KPI 2 ,..., KPI n is the characteristic attribute vector used to describe the network state obtained according to the values of each KPI. The process of fault diagnosis is shown as follows: IF KPI 1 >TH 1 AND KPI 2 <TH 2 ...AND KPI n >TH n ,THEN D(cell)=C i Among them, TH i is the status division threshold of the preset i-th KPI, used to indicate whether the KPI is normal, and D(cell) represents the fault diagnosis result of the cell; when all conditions of a certain rule are met, the method will output the fault cause C corresponding to the status S i ; however, if no rule matches, the output is UN, indicating an unforeseen fault; Step S3.3: Classify the remaining data using the trained Naive Bayes model to obtain a pre-diagnosis result label set where N represents the total number of samples in the dataset after expansion by WGAN-GP; the specific classification steps are as follows: Select the network fault category with the maximum posterior probability as the network fault h suffered by the current network * (X), that is: Convert to logarithmic form: Estimate the prior probability P(y i ) and the conditional probability P(x j |y i ) by the Laplace smoothing method, that is: Among them, D t represents the total number of samples contained in the training dataset; is the total number of samples in the training set under the network fault y i condition; represents the total number of samples in the training set under the network fault y i condition and the value of the j-th KPI parameter is x j ; L is the total number of previously defined network fault categories; S j is the number of all possible values of the j-th KPI.

5. A method for diagnosing cellular network faults based on knowledge and data fusion according to claim 1, characterized in that, Based on the pre-diagnosis result set to construct a topological association graph, i.e., an adjacency matrix A, the specific content is as follows: In the pre-diagnosis result set data diagnosed with the same network fault type are connected to each other in the figure, while data with different network fault types are not connected to each other, that is:

Images