Fault diagnosis method and system for wireless network

Abstract

The invention discloses a fault diagnosis method and system for a wireless network, and the method comprises the steps: obtaining graph data of a current wireless network, the graph data comprising a correlation matrix and an adjacent matrix between nodes; converting the graph data into a tree matrix according to the correlation matrix and the adjacent matrix; performing spatial convolution operation on the tree matrix to obtain a spatial feature matrix of the tree matrix; calculating a weight matrix based on the spatial feature matrix of the tree matrix; performing time convolution operation on the weight matrix to obtain a time sequence characteristic matrix; and performing full connection processing and activation operation on the time sequence characteristic matrix to obtain a final node fault classification result. According to the method, the problem of a local path existing in graph convolution is solved, space convolution is applied to wireless network fault detection, convolution can be performed on different paths and layers, the possibility of information loss is reduced, the spatial correlation obtaining capability is enhanced, and faults in a wireless network can be determined more quickly and accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of network fault diagnosis, and in particular to a wireless network fault diagnosis method and system. Background Art

[0002] Wireless sensor networks are composed of a series of interconnected sensor nodes. They use distributed technology, wireless communication, and computer networks to achieve multiple functions such as data monitoring. At present, WSN plays an important role in many fields, from basic data monitoring to danger warning and traffic flow analysis in smart cities, such as real-time traffic flow monitoring and pedestrian trajectory recording, which fully demonstrates its broad application prospects. In order to ensure the stability and user experience of wireless networks, fault diagnosis is a key link. With the widespread application of wireless networks, their complexity is also increasing. Current fault diagnosis methods cover a variety of technologies such as data analysis, machine learning, artificial intelligence, and distributed computing. Among them, machine learning and artificial intelligence methods are particularly representative.

[0003] There are certain defects in various traditional fault diagnosis methods. For example, in the method based on time series analysis, this method focuses on low-dimensional data and cannot be applied to high-dimensional data. The time series analysis method is more detailed in fault division and has higher detection efficiency, but it cannot capture the dynamic change characteristics in the process of wireless network information transmission. For another example, in the method based on graph data analysis, when processing graph data, its adjacency matrix can only reflect whether there is a relationship or not. When the relationship between nodes is strong or weak, the adjacency matrix alone cannot accurately represent the relationship between nodes. This will affect the acquisition of features in the data and thus affect the accuracy of fault diagnosis.

[0004] Therefore, how to overcome the shortcomings of the above-mentioned prior art solutions has high research value. Summary of the invention

[0005] In view of this, the present invention proposes a wireless network fault diagnosis method and system, which can transform the original graph structure into a hierarchical tree structure by constructing a tree matrix, so as to improve the efficiency and accuracy of fault diagnosis between nodes. The present invention provides the following technical solutions:

[0006] A method for diagnosing a wireless network fault, comprising:

[0007] Obtain graph data of the current wireless network, the graph data including a correlation matrix and an adjacency matrix between nodes;

[0008] Convert graph data into tree matrix based on correlation matrix and adjacency matrix;

[0009] Perform spatial convolution operation on the tree matrix to obtain its spatial feature matrix;

[0010] Calculate the weight matrix based on the spatial feature matrix of the tree matrix;

[0011] Perform a time convolution operation on the weight matrix to obtain a time series feature matrix;

[0012] The time series feature matrix is ​​fully connected and activated to obtain the final node fault classification result.

[0013] Further, the method of converting graph data into a tree matrix according to the correlation matrix and the adjacency matrix includes:

[0014] Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures;

[0015] Respectively obtain all path data starting from any specified node in the tree structure, and expand the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in the matrix;

[0016] Taking each node of the tree structure as a root node, generating a corresponding one-dimensional tree matrix, so as to construct a spatial tree matrix composed of multiple one-dimensional tree matrices;

[0017] The node values ​​of the spatial tree matrix are converted into the magnitude of the correlation between nodes according to the correlation matrix to complete the transformation of the tree matrix.

[0018] Further, the method of performing a spatial convolution operation on the tree matrix to obtain its feature matrix includes:

[0019] The spatial convolution includes path convolution and layer convolution;

[0020] The path convolution method includes:

[0021] Each column of the tree matrix serves as a local path in the path tree convolution;

[0022] A convolution operation is performed upward from the bottom end of each local path of each one-dimensional tree matrix, and a feature matrix that aggregates the local features of each node is obtained;

[0023] Perform circular convolution operations on the feature matrix column by column to obtain high-level feature matrices of all local paths, and connect multiple high-level feature matrices to obtain the path feature matrix of each one-dimensional tree matrix;

[0024] The hierarchical convolution method includes:

[0025] Starting from the leaf nodes of each one-dimensional tree matrix, a convolution operation is performed horizontally to obtain a feature matrix including the node and the level information;

[0026] Perform a circular convolution operation on the feature matrix to obtain high-level feature matrices of all local paths, and connect multiple high-level feature matrices to obtain a hierarchical feature matrix of each one-dimensional tree matrix;

[0027] The path feature matrix and the hierarchical feature matrix are fused to obtain the spatial feature matrix of the tree matrix.

[0028] Further, the method for calculating the weight matrix based on the spatial feature matrix of the tree matrix includes:

[0029] Get the timing data of the current wireless network;

[0030] Multiply the spatial feature matrix with the temporal data to obtain the weight matrix.

[0031] Further, the method of performing a time convolution operation on the weight matrix to obtain a time series feature matrix includes:

[0032] A Gaussian convolution operation is performed on the weight matrix to obtain high-dimensional time feature data in the time series data, thereby obtaining a time series feature matrix.

[0033] The present invention further discloses a wireless network fault diagnosis system, comprising:

[0034] A data acquisition module is used to acquire graph data of the current wireless network, where the graph data includes a correlation matrix and an adjacency matrix between nodes;

[0035] A conversion module for converting graph data into a tree matrix based on a correlation matrix and an adjacency matrix;

[0036] The convolution operation module is used to perform spatial convolution operation on the tree matrix to obtain its spatial feature matrix; it is also used to,

[0037] Calculate the weight matrix based on the spatial feature matrix of the tree matrix;

[0038] Perform a time convolution operation on the weight matrix to obtain a time series feature matrix;

[0039] The fault diagnosis module is used to perform full connection processing and activation operations on the time series feature matrix to obtain the final node fault classification result.

[0040] Furthermore, the conversion module is also used to:

[0041] Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures;

[0042] Respectively obtain all path data starting from any specified node in the tree structure, and expand the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in the matrix;

[0043] Taking each node of the tree structure as a root node, generating a corresponding one-dimensional tree matrix, so as to construct a spatial tree matrix composed of multiple one-dimensional tree matrices;

[0044] The node values ​​of the spatial tree matrix are converted into the magnitude of the correlation between nodes according to the correlation matrix to complete the transformation of the tree matrix.

[0045] The present invention further discloses a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0046] The present invention further discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the program.

[0047] The present invention further discloses a computer program product, comprising a computer program, wherein the computer program implements the above method when executed by a processor.

[0048] According to the technical solution of the present invention, the original adjacency matrix is ​​replaced by a tree matrix, and the graph structure is represented in a tree manner. The tree structure can more clearly represent all nodes that have an impact on the root node, not just the neighboring nodes, and solve the local path problem existing in the graph convolution. At the same time, applying spatial convolution to wireless network fault detection can convolve different paths and levels in the tree matrix. This method not only aggregates features at different levels but also retains important information, reducing the possibility of information loss. Moreover, the hierarchical nature of the tree matrix makes it convenient for us to set the correlation size of the nodes, enhances the ability to obtain spatial correlation, and is conducive to more rapid and accurate determination of faults in wireless networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] For the purpose of illustration and not limitation, the present invention is now described in conjunction with the embodiments of the present invention and the accompanying drawings, in which:

[0050] Figure 1 is a flowchart of a method for diagnosing a wireless network fault in an embodiment of the present invention;

[0051] Figure 2 It is a schematic diagram of the implementation architecture of the wireless network fault diagnosis method in an embodiment of the present invention;

[0052] Figure 3 It is a structural diagram of a wireless network fault diagnosis system in an embodiment of the present invention.

[0053] Figure 4 is a schematic diagram of the structure of an electronic device in an embodiment of the present invention;

[0054] Figure 5 is a schematic diagram of path convolution in an embodiment of the present invention;

[0055] Figure 6 Schematic diagram of hierarchical convolution in an embodiment of the present invention;

[0056] Figure 7 Schematic diagram of Gaussian convolution in an embodiment of the present invention;

[0057] Figure 8 It is a comparison of model accuracy under different numbers of nodes in the comparative experiment in the implementation mode of the present invention;

[0058] Fig. 9 is a schematic diagram of experimental comparison of classification accuracy of different comparison models in an embodiment of the present invention;

[0059] Fig.10 Schematic diagram of comparison of confusion matrix results of different comparison models in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the implementation mode of the present application will be clearly and completely described below in conjunction with the drawings in the implementation mode of the present application. Obviously, the described implementation mode is only a part of the implementation mode of the present application, not all the implementation modes. Based on the implementation mode in the present application, all other implementation modes obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0061] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0062] In the present application, the terms "upper", "lower", "left", "right", "front", "back", "top", "bottom", "inner", "outer", "middle", "vertical", "horizontal", "lateral", "longitudinal" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings. These terms are mainly used to better describe the present application and its implementation methods, and are not used to limit the indicated devices, elements or components to have a specific orientation, or to be constructed and operated in a specific orientation.

[0063] In addition, some of the above terms may be used to express other meanings in addition to indicating orientation or positional relationship. For example, the term "on" may also be used to express a certain dependency or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in this application can be understood according to the specific circumstances. In addition, the term "plurality" should mean two or more than two.

[0064] It should be noted that, in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other. The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0065] refer to Figure 1 and Figure 2 , a wireless network fault diagnosis method, comprising:

[0066] S100: Acquire graph data of the current wireless network, where the graph data includes a correlation matrix and an adjacency matrix between nodes.

[0067] The adjacency matrix reflects the direct connection between nodes on the way, while the correlation matrix represents the strength or similarity of the relationship between nodes, reflecting the correlation between nodes. Since the adjacency matrix can only reflect whether there is a relationship or not between nodes, when the relationship between nodes is strong or weak, the adjacency matrix alone cannot accurately represent the relationship between nodes. When the features in the data are obtained based on this, the accuracy of fault diagnosis will be greatly affected. Therefore, it is necessary to construct a structure that can hierarchically represent the hierarchy of the original graph structure to avoid spatial correlation between nodes.

[0068] S200: Convert the graph data into a tree matrix according to the correlation matrix and the adjacency matrix.

[0069] The tree matrix can transform the original complex graph structure into a hierarchical tree structure to better capture the spatial correlation between nodes. The tree matrix node contains the correlation between the node and its parent node. Specifically, this embodiment discloses a method for constructing a tree matrix, which specifically includes:

[0070] S210: Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures.

[0071] If there are loops in the graph structure converted from the adjacency matrix, then there will also be loops in the result after breadth-first traversal. Therefore, the one-dimensional tree matrix will have problems in representing the original graph structure. Therefore, a spatial tree matrix is ​​used to represent the original tree structure. Due to the characteristics of breadth-first traversal, the tree structure result can be directly obtained. At the same time, the obtained tree structure result is divided into different subtrees and the maximum number of layers, number of nodes, and number of child nodes are calculated.

[0072] S220: respectively obtaining all path data starting from any designated node in the tree structure, and expanding the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in a matrix.

[0073] S230: taking each node of the tree structure as a root node, generating a corresponding one-dimensional tree matrix, so as to construct a spatial tree matrix composed of multiple one-dimensional tree matrices;

[0074] S240: Convert the node values ​​of the spatial tree matrix into the magnitude of the correlation between the nodes according to the correlation matrix, so as to complete the conversion of the tree matrix.

[0075] The fault diagnosis model constructed in this embodiment is a tree path convolutional network, and the model uses a combination of spatial convolution and temporal convolution to complete the fault diagnosis of the wireless network. Specifically:

[0076] S300: Perform a spatial convolution operation on the tree matrix to obtain its spatial feature matrix.

[0077] Specific methods include:

[0078] Spatial convolution includes path convolution and layer convolution;

[0079] refer to Figure 5 , the path convolution methods include:

[0080] S310: using each column of the tree matrix as a local path in the path tree convolution;

[0081] S320: performing a convolution operation upward from the bottom end of each local path of each one-dimensional tree matrix, and obtaining a feature matrix that aggregates local features of each node;

[0082] This implementation takes a 3x4 matrix as an example. The first row is the root node of the tree matrix, and each column is a local path. The shaded part in the figure is the convolution kernel of the current convolution operation. The path convolution process is only performed between the current node and its parent node each time, ensuring that more accurate information can be obtained during convolution. At the same time, the implicit relationship between nodes can be captured, and a high-level representation of the nodes in the current path can be established to reduce the complexity of the network.

[0083] S330: Perform circular convolution operations on the feature matrix by column to obtain high-level feature matrices of all local paths, and connect multiple high-level feature matrices to obtain a path feature matrix of each one-dimensional tree matrix.

[0084] The convolution process of the local path is performed independently, and each path does not interfere with each other, making the convolution feature more accurate. Compared with the method of convolving the entire local path at once, this convolution method can better retain the information of deep nodes. During the overall convolution, the information of all nodes will be processed at the same time, and the deep node information may be partially ignored. However, this path convolution is performed step by step, which can better retain the information of these nodes. To a certain extent, information loss is avoided, and the characteristics of each path are deeply learned.

[0085] refer to Figure 6 , the methods of hierarchical convolution include:

[0086] S340: Starting from the leaf nodes of each one-dimensional tree matrix, a convolution operation is performed horizontally to obtain a feature matrix including node and level information.

[0087] This embodiment uses Figure 6 Take the tree structure and tree matrix in as an example, where f represents the hierarchical features obtained by the convolution process.

[0088] During the generation process of the tree matrix, the root node will be copied multiple times to ensure that each node has a path to the root node, so that the information of the root node can be obtained multiple times during convolution. Similarly, the nodes connected to the root node will also be convolved multiple times. The convolution kernel performs horizontal translation calculations according to the row priority principle, so that the information of the node and the level can be obtained during the operation. Because the root node and the nodes connected to the root node are convolved multiple times, they retain more information and reduce the information loss during convolution. The hierarchical features obtained in this way can not only include the features of each level, but also the hidden information between nodes. The nodes in the wireless network have different effects on the root node according to the different levels. The use of hierarchical convolution can more comprehensively capture the relationship between the node and the global state to obtain the position and features of the node at each level.

[0089] S350: performing a circular convolution operation on the feature matrix to obtain high-level feature matrices of all local paths, and connecting multiple high-level feature matrices to obtain a hierarchical feature matrix of each one-dimensional tree matrix;

[0090] S360: Fuse the path feature matrix and the hierarchical feature matrix to obtain a spatial feature matrix of the tree matrix.

[0091] After the spatial convolution operation is completed and the spatial feature matrix is ​​obtained, it is necessary to add temporal convolution to obtain the temporal features in the time series data. Specifically:

[0092] S400: Calculate a weight matrix based on the spatial feature matrix of the tree matrix. Specifically including:

[0093] Get the timing data of the current wireless network;

[0094] Multiply the spatial feature matrix with the temporal data to obtain the weight matrix.

[0095] In this embodiment, before the weight matrix is ​​calculated, the graph data of the wireless network needs to be processed. Since each root cause has specific feature data corresponding to it, all data cannot be put into the classifier at once. Therefore, the first step after obtaining the graph data is to extract specific features. For example, root cause 1 uses 39 features, root cause 2 and root cause 3 use 18 features, and these three root causes have a few repeated features. Then, after the features are divided, data processing is performed according to the input format required by the classifier. For example, it is found through observation that the original data set has problems such as missing sample data and insufficient sample size, so the input data must be processed first. For example, the characteristic values ​​of the training set data and the test set data are quite different, and the characteristic value range is distributed between 0-700, which is not conducive to model training. Therefore, z-score normalization is used to narrow the characteristic value range and reduce the data scale. After normalization, the original data is mapped to a standard normal distribution centered on 0, which can accelerate the training of the model. Secondly, since the label categories in the data set are unbalanced, this will affect the classification effect, so Borderline SMOTE is used for oversampling. Borderline SMOTE can increase the amount of data by synthesizing samples, which can improve the model classification effect while solving the data imbalance. Furthermore, in traditional neural networks, as the number of network layers increases, the performance of the model may degenerate, that is, the deeper network performs worse than the shallow network. Therefore, the residual network (Residual Network, ResNet) is introduced in this embodiment, which is a deep learning network structure that solves the gradient disappearance and degradation problems in deep neural networks by introducing skip connections. The principle of the skip connection is to directly pass the input to the subsequent layer after bypassing several layers, so that the network can learn the residual. Through such a design, the gradient is allowed to be more effectively propagated in the deep network, avoiding the problems of gradient disappearance and gradient explosion.

[0096] refer to Figure 6 , S500: performing a time convolution operation on the weight matrix to obtain a time series feature matrix;

[0097] The temporal convolution process is the process of obtaining the time series feature matrix through Gaussian convolution operation.

[0098] The specific operations are:

[0099] refer to Figure 7 , perform Gaussian convolution operation on the weight matrix to obtain high-dimensional time feature data in the time series data, thereby obtaining the time feature matrix.

[0100] The time series data in wireless networks are recorded at certain time intervals. There is continuity between the data, and the occurrence of faults is usually reflected in the fluctuation of a certain period of time series data. Therefore, the Gaussian convolution method can be used to extract high-dimensional time features in the data.

[0101] The probability density function formula of Gaussian distribution is expressed as:

[0102]

[0103] For example, where σ is the standard deviation, the value is 0.7, x is the variable, and μ is the mean.

[0104] After normalizing the Gaussian convolution, we can get the formula for Gaussian convolution:

[0105] M i+1 =G n *f(M i );

[0106] Where Mi is the spatial feature matrix obtained after spatial convolution, G n is the convolution kernel of the current Gaussian convolution, f(M i ) is the output of the activation function.

[0107] After the convolution is completed, it is normalized, passed through the ReLu layer, and finally regularized by Dropout. This process is repeated many times to capture the temporal features in the data and form a temporal feature matrix.

[0108] S600: Perform full connection and activation operations on the time series feature matrix to obtain a final node fault classification result.

[0109] The wireless network data used in this embodiment is from the ICASSP Communication Network Intelligent Operation and Maintenance Competition. The purpose of the data research is to find out the root cause of the low value of feature0 in the wireless network data. Each feature represents a key performance indicator (KPI). The training data contains a total of 2984 samples, which come from different 5G road test segments. In order to simulate the real situation, only 45% of the data are labeled with root causes, and the rest of the data are unlabeled. The node records data using time slice analysis, and a time slice may contain one or more root causes. The feature data in the node is diversified and heterogeneous data, which may be continuous values ​​or discrete values. The data collection granularity is also different. Most data is recorded once every one second, but some data intervals will reach tens of seconds, and in order to simulate the real situation, all data are not fixed length, and the number of data records ranges from a few to thousands.

[0110] The following is the experimental part:

[0111] In the experimental part, the four indicators of accuracy, precision, recall and F1 score are used to evaluate the classification effect of the fault diagnosis model in this implementation. Accuracy is the most commonly used evaluation value indicator and the most intuitive indicator to reflect the quality of a classification model. Accuracy is the proportion of all correct samples to the total samples, which can show the overall performance of the model, which is simple and intuitive. Precision is the proportion of the number of positive examples that the model can correctly classify as positive examples to the number of all positive examples classified. This indicator can detect the proportion of faults detected by the model that actually occur, and can check the accuracy of the model's prediction of positive examples, which is helpful for model improvement. Recall is the proportion of correct samples to actual correct samples. In wireless networks, it is to check its ability to detect faults. The recall rate can show the model's ability to identify actual positive samples. The F1 score takes into account both accuracy and recall, which can comprehensively reflect the model's capabilities and is a more comprehensive indicator.

[0112] This experiment will compare with some common machine learning and deep learning algorithms. The specific comparison models and parameter settings are as follows:

[0113] ACO-GCN: The ant colony algorithm and the combination of GCN and GRU are used for diagnosis. The initial step size is 1 and the compensation attenuation factor is 0.96.

[0114] GAT+LSTM: For the data, LSTM is first used to extract temporal information, and then GAT is used to extract spatial information. The dropout rate is 0.2 and the attention unit is 8.

[0115] GCN: A method for convolution of graph structures. The convolution process can consider the neighbor node information of the node. A three-layer GCN is used for wireless network fault diagnosis. Self-loop is added to the training to consider its own information. Weight decay is added to prevent overfitting, and the parameter is set to 5e-4.

[0116] GraphSAGE: An improved method of graph convolution, which uses random discarding to avoid the time-consuming problem caused by too many neighbor nodes. It uses three layers of convolution to help learn complex nodes. The number of neighbor sampling is set to 10.

[0117] ERT: Similar to random forest, it uses random feature subsets for splitting and then votes to decide the classification. The number of trees is set to 30, the quality criterion is the Gini coefficient, and the minimum number of samples required to split a node is 2

[0118] Quantitative experiments:

[0119] In this section, we will conduct a comparative experiment between the tree path convolutional network and the set comparison model. Before the experiment, we must first set the tree matrix. In order to better compare, all experiments will include ordinary tree path convolutional networks and tree path convolutional networks with residual networks. In the tree path convolutional network model, the tree matrix size needs to be defined according to the task requirements. If the experiment is not conducted, the processing efficiency of the model will be affected. Figure 8 The figure shows the detection accuracy of the tree path convolutional network with different numbers of nodes. It can be seen that the model cannot achieve a high accuracy when the number of nodes is small. At this time, too few features can be obtained to accurately judge the fault. When the number of nodes exceeds 50, the accuracy shows a downward trend. The probability of this situation will increase as the number of nodes increases, which will limit the transmission of information and lead to a decrease in accuracy. Figure 8 It can be seen that the accuracy reaches the highest 91.83% when the number of nodes is 50, so we choose a tree matrix with a size of 50 nodes.

[0120] Fig. 9The comparison results of tree path convolution network and other comparison models in terms of accuracy, precision and other indicators. In terms of accuracy, the path tree convolution is the highest among these models. In terms of precision and recall, the tree path convolution network has reached a very high value, which shows that compared with other graph-related models, it has fewer misjudgments and can better distinguish positive examples. The GAT+LSTM method uses LSTM to obtain the time characteristics of the node. At the same time, GAT can set different weights for all neighbor nodes to distinguish the influence of different nodes. Therefore, the GAT+LSTM method is the highest among the graph-related methods. However, the difference in accuracy is 1.83%, which does not exceed the path tree convolution. ACO-GCN has improved in accuracy compared to GCN, but it does not exceed the GAT+LSTM method because it still uses the graph convolution method in terms of spatial features. SVM usually processes structured data, and the precision and recall rates for this wireless network dataset are not high. Although ERT can capture nonlinear relationships and can adapt to a variety of different types of data, it cannot analyze graph structured data, so the overall F1 score is low. In general, the F1 score of the tree path convolutional network is the highest, which is 1.86% and 2.03% higher than that of GAT-LSTM and ACO-GCN, which are also spatiotemporal graph convolution methods, respectively, proving that its comprehensive ability is superior to these comparison models.

[0121] Qualitative experiments and analysis:

[0122] like Fig.10 This is the confusion matrix comparison between the model and other comparison models. Rootcase1 is represented by R1, and so on. The confusion matrix can reflect the comparison between the predicted results and the actual results. The darker the color on the diagonal, the closer the predicted value is to the actual value. In addition, the confusion matrix can more intuitively display the classification of specific root causes, which helps to improve the efficiency of root cause analysis. Figure 7The classification results of path tree convolution are generally better than those of other comparison models. R2 has the most misclassifications. This is because there are only four data points in the training data where R2 occurs alone, and the rest are where R2 and R3 occur together. This will cause the model to have insufficient data to distinguish the cases where R2 occurs alone, resulting in misclassification of the results. GAT+LSTM is roughly the same as path tree convolution in the case of a single root cause, but the number of misclassifications increases in the case of multiple root causes. In the process of spatial convolution, the GAT+LSTM method cannot consider the influence of more distant nodes, which leads to its overall performance being inferior to path tree convolution. In particular, when R1 and R3 occur at the same time, the shared features weaken the classification effect on edge nodes such as R3, so the classification results in the confusion matrix are 6% worse. Compared with GCN, ACO-GCN has an additional time convolution part, so it is worse than GCN in overall classification effect. GAT+LSTM can set the size of the correlation, while ACO-GCN cannot, which leads to its lower accuracy in root cause classification than GAT+LSTM. GCN's classification effect is relatively good in the case of single root causes, but it cannot accurately distinguish the root causes in complex situations with multiple root causes. This is because the basic graph convolution cannot take into account the correlation and time characteristics in the nodes, and its diagnostic ability is limited in complex topological environments. GraphSAGE randomly discards nodes during the convolution process, further exacerbating the problem of insufficient training data, resulting in inaccurate classification results in the case of multiple root causes. That is, since random trees cannot utilize the graph structure in wireless networks, their confusion matrix performance is generally lower than that of graph-related methods.

[0123] Ablation experiment:

[0124] Path tree convolution is a spatiotemporal hybrid model that extracts spatial features and temporal features from wireless network data for convolution. In order to further study the impact of temporal and spatial features in the model on the model's diagnostic capabilities, we removed the temporal feature extraction part and the spatial feature extraction part of the model and compared the experimental results. As shown in the following table:

[0125] Table 1. Comparison of ablation experiments

[0126]

[0127] From the data in Table 1, it can be seen that when only the spatial feature extraction part is retained, the overall fault diagnosis ability of the model does not decrease much. This is because there is a transmission effect of faults between nodes in the wireless network data. The path convolution and hierarchical convolution contained in the spatial feature extraction part of the model can capture these faults. Since the faults affect each layer upward, the corresponding tree structure is that the faults of the leaf nodes are transmitted to the root node, which is the same as the way of extracting features by path convolution, so it will not have much impact on the overall performance. However, when only the time feature extraction part is retained, the accuracy and precision rate and other indicators are greatly reduced. This is because the model only convolves the node time series data and ignores the influence of space. This will increase the probability of misjudging normal nodes as faulty nodes, affecting the model's ability to diagnose faults. Experiments have shown that for wireless network data with complex topology structures, the method of using time correlation plus spatial correlation can achieve better diagnosis results.

[0128] In summary, when using graph convolution for wireless network fault diagnosis, it is found that graph convolution has local path and hierarchical limitations. It can only capture the characteristics of neighboring nodes of a node and cannot consider the impact of more distant nodes on the node. The path between nodes is too complex and has no hierarchy, and it is impossible to study the size of node correlation. In view of these limitations, this embodiment proposes a method for wireless network fault diagnosis using a tree path convolution network. The tree matrix can be used to transform the original graph structure into a hierarchical tree structure, and the tree path convolution network can also ensure that the spatial relationship between nodes can be captured more comprehensively. The tree matrix replaces the original adjacency matrix, and the graph structure is represented in the form of a tree. The influence between nodes in the wireless network will be transitive, and the tree structure can more clearly represent all nodes that have an impact on the root node, not just the neighboring nodes. This solves the local path problem in graph convolution. The tree path convolutional network is applied to wireless network fault detection. The tree path convolutional network can convolve different paths and levels in the tree matrix. This method aggregates features at different levels while retaining important information, reducing the possibility of information loss. In addition, the hierarchical nature of the tree matrix makes it easier for us to set the correlation size of nodes, enhancing the ability to obtain spatial correlation. The tree path convolutional network is compared with other comparison models on the competition data set. Experiments show that the tree path convolutional network has an accuracy improvement of 2.98% compared to graph convolution. It also performs better in terms of precision and recall compared to other comparison models. In the confusion matrix, the path tree convolution also demonstrates its fault diagnosis ability in the case of multiple causes.

[0129] refer to Figure 3 This embodiment further discloses a wireless network fault diagnosis system, including:

[0130] The data acquisition module 21 is used to acquire graph data of the current wireless network, where the graph data includes a correlation matrix and an adjacency matrix between nodes.

[0131] A conversion module 22 is used to convert the graph data into a tree matrix according to the correlation matrix and the adjacency matrix; and is also used to,

[0132] Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures;

[0133] Respectively obtain all path data starting from any specified node in the tree structure, and expand the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in the matrix;

[0134] Taking each node of the tree structure as a root node, generating a corresponding one-dimensional tree matrix, so as to construct a spatial tree matrix composed of multiple one-dimensional tree matrices;

[0135] The node values ​​of the spatial tree matrix are converted into the magnitude of the correlation between nodes according to the correlation matrix to complete the transformation of the tree matrix.

[0136] The convolution operation module 23 is used to perform a spatial convolution operation on the tree matrix to obtain its spatial feature matrix; and is also used to:

[0137] Calculate the weight matrix based on the spatial feature matrix of the tree matrix;

[0138] Perform a temporal convolution operation on the weight matrix to obtain a temporal feature matrix.

[0139] The fault diagnosis module 24 is used to perform full connection processing and activation operations on the time series feature matrix to obtain the final node fault classification result.

[0140] Figure 4 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 4 As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0141] The processor 501 and the memory 502 communicate with each other via the bus 503 ; the processor 501 is used to call program instructions in the memory 502 to execute the methods provided by the above-mentioned method implementation methods.

[0142] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above-mentioned method embodiments.

[0143] This embodiment further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the computer program implements the above-mentioned method for diagnosing faults in a wireless network.

[0144] A person skilled in the art can understand that all or part of the steps for implementing the above-mentioned method implementation method can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes the steps of the above-mentioned method implementation method; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0145] The device implementation described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present implementation scheme. Those of ordinary skill in the art may understand and implement it without creative effort.

[0146] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each implementation method or some parts of the implementation method.

[0147] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for diagnosing a wireless network fault, characterized in that: include: Obtain graph data of the current wireless network, the graph data including a correlation matrix and an adjacency matrix between nodes; Convert graph data into tree matrix based on correlation matrix and adjacency matrix; Perform spatial convolution operation on the tree matrix to obtain its spatial feature matrix; Calculate the weight matrix based on the spatial feature matrix of the tree matrix; Perform a time convolution operation on the weight matrix to obtain a time series feature matrix; The time series feature matrix is ​​fully connected and activated to obtain the final node fault classification result.

2. The wireless network fault diagnosis method according to claim 1, characterized in that: The method for converting graph data into a tree matrix according to a correlation matrix and an adjacency matrix comprises: Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures; Respectively obtain all path data starting from any specified node in the tree structure, and expand the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in the matrix; Taking each node of the tree structure as a root node, generating a corresponding one-dimensional tree matrix, so as to construct a spatial tree matrix composed of multiple one-dimensional tree matrices; The node values ​​of the spatial tree matrix are converted into the magnitude of the correlation between nodes according to the correlation matrix to complete the transformation of the tree matrix.

3. The wireless network fault diagnosis method according to claim 1, characterized in that: The tree matrix is ​​subjected to a spatial convolution operation to obtain its feature matrix. Specific methods include: The spatial convolution includes path convolution and layer convolution; The path convolution method includes: Take each column of the tree matrix as a local path in the path tree convolution; A convolution operation is performed upward from the bottom end of each local path of each one-dimensional tree matrix, and a feature matrix that aggregates the local features of each node is obtained; Perform circular convolution operations on the feature matrix by column to obtain high-level feature matrices of all local paths, and connect multiple high-level feature matrices to obtain the path feature matrix of each one-dimensional tree matrix; The hierarchical convolution method includes: Starting from each leaf node of each one-dimensional tree matrix, a convolution operation is performed horizontally to obtain a feature matrix including the node and the level information; Perform a circular convolution operation on the feature matrix to obtain high-level feature matrices of all local paths, and connect multiple high-level feature matrices to obtain a hierarchical feature matrix of each one-dimensional tree matrix; The path feature matrix and the hierarchical feature matrix are fused to obtain the spatial feature matrix of the tree matrix.

4. The wireless network fault diagnosis method according to claim 1, characterized in that: The method for calculating the weight matrix based on the spatial feature matrix of the tree matrix includes: Get the timing data of the current wireless network; Multiply the spatial feature matrix with the temporal data to obtain the weight matrix.

5. The wireless network fault diagnosis method according to claim 1, characterized in that: The method of performing a time convolution operation on the weight matrix to obtain a time series feature matrix includes: A Gaussian convolution operation is performed on the weight matrix to obtain high-dimensional time feature data in the time series data, thereby obtaining a time series feature matrix.

6. A wireless network fault diagnosis system, characterized in that: include: A data acquisition module is used to acquire graph data of the current wireless network, where the graph data includes a correlation matrix and an adjacency matrix between nodes; A conversion module for converting graph data into a tree matrix based on a correlation matrix and an adjacency matrix; A convolution operation module, used for performing a spatial convolution operation on the tree matrix to obtain its spatial feature matrix; Also used for Calculate the weight matrix based on the spatial feature matrix of the tree matrix; Perform a time convolution operation on the weight matrix to obtain a time series feature matrix; The fault diagnosis module is used to perform full connection processing and activation operations on the time series feature matrix to obtain the final node fault classification result.

7. The wireless network fault diagnosis system according to claim 6, characterized in that: The conversion module is also used for: Perform a breadth-first traversal operation on the graph data according to the adjacency matrix to obtain multiple tree structures; Respectively obtain all path data starting from any specified node in the tree structure, and expand the root node according to the path data to ensure that the lengths of multiple tree structures are unified when represented in the matrix; Each node of the tree structure is taken as a root node to generate a corresponding one-dimensional tree matrix to construct a spatial tree matrix composed of multiple one-dimensional tree matrices; The node values ​​of the spatial tree matrix are converted into the magnitude of the correlation between nodes according to the correlation matrix to complete the transformation of the tree matrix.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method described in any one of claims 1 to 5 is implemented.

10. A computer program product, comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images