Method for processing missing values of network log data and classifying the root cause of communication failures through the same
By employing machine learning and deep learning techniques to process missing values in network log data, the method addresses the limitations of existing RCA methods, enabling accurate classification of communication defects and complete dataset analysis.
Patent Information
- Application Number
- JP2023190619
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2022-11-10
- Filing Date
- 2023-11-08
- Publication Date
- 2024-11-20
- Estimated Expiration
- 2043-11-08
AI Technical Summary
Existing root cause analysis (RCA) methods for network communication defects are limited by missing values in network log data, leading to incomplete datasets and biased analyses, and human-configured logic may fail to identify the root cause accurately.
A method using machine learning and deep learning to process missing values in network log data by separating data into non-time series and time series, applying Gaussian mixture models and autoregressive tensor completion to replace missing values, and training neural networks for classification.
This approach enables accurate classification of communication defects by compensating for missing data, reducing the number of unidentified root causes, and providing a complete dataset for effective network optimization.
Smart Images

Figure 0007587662000003 
Figure 0007587662000004 
Figure 0007587662000005
Abstract
Description
[Technical field]
[0001] The present invention relates to a method for processing missing values of network log data and classifying the root cause of communication defects through the same, and more particularly, to a method for processing missing values of network log data and classifying the root cause of communication defects through the same, which is capable of analyzing the root cause of various defects occurring in a network environment by utilizing machine learning or deep learning technology in a state in which a complete data set is secured by obtaining an appropriate substitute value according to the characteristics of a parameter where a missing value exists. [Background technology]
[0002] Root Cause Analysis (RCA) identifies the root cause of network communication faults and provides network operators with the information they need to optimize their networks. Existing RCA technology measures various network parameters during a call and identifies the root cause of the fault through logic set by experts based on the correlation between the measured values and various parameters, so it has many limitations.
[0003] Specifically, existing root cause identification methods compare parameter values measured during the call section (call time) where a defect occurred with a manually set logic threshold to identify the root cause of the defect. First, defects are classified into broad RCA categories, and then each RCA category is divided into reason groups, and the measured data is compared with the corresponding logic threshold to subdivide the defect into the corresponding root category. These existing methods have the following limitations:
[0004] First, there are likely to be missing values in the parameter data that represents the network environment due to various causes within the communication network, and there are likely to be limitations in the logic because various cases are not fully taken into account.
[0005] In addition, if only limited data is used in the process of setting up the logic and in the process of making judgments through the logic, the root cause of the defect cannot be identified and it may end up being judged as "unknown." For this reason, a method is needed to supplement the limitations of existing logic by considering various cases even when complete data is not given.
[0006] In addition, the logic set by a human may be limited to the viewpoint of an expert who sees the network environment where the defect occurred. In the existing logic, in addition to the parameter relationships and thresholds specified as each root cause, there may be other factors that can determine the root, or the thresholds may need to be adjusted. For example, a case where the root was determined to be "unknown" in the existing logic may be classified into the closest root through appropriate threshold adjustment.
[0007] In addition, in order to classify a specific root using the designed logic, the measured value of the network parameter cannot be used as it is, and the data must be converted into an order type or category type. In this case, if the data value is used as it is, it may not be possible to recognize the interrelationship between parameters that can be perceived as the cause of a specific root.
[0008] To complement this, technology has been presented that uses machine learning and deep learning based on massive amounts of measurement data to recognize correlations between various parameters and identify the root cause of defects.
[0009] Other solutions have been presented to solve the problem of incomplete data sets that are realistically collected in the field. For example, the problem of missing values in a data set is a major limitation that often occurs and causes many difficulties in classifying the root of the defect.
[0010] Specifically, the network communication log data used in RCA contains missing values due to the inability to measure certain parameters for various technical or environmental reasons, and there are three main types of missing values. The most widespread type of missing value data is Missing at Random (MAR). This type is when the proportion of missing value data in the total is related to the observed data in the dataset. In other words, when a missing value exists for a specific parameter, it is related to other parameters in the dataset.
[0011] The opposite of missing at random is missing not at random (MNAR), which means that missing values are not random, but rather that the missing parameter values are related to the reason they are missing. The final category, missing completely at random (MCAR), is when the overall proportion of missing values is unrelated to the rest of the data. This category is when any parameter is randomly missing across the entire dataset, and data are missing with a similar distribution regardless of the type and value of the parameter.
[0012] On the other hand, in order to apply machine learning or deep learning algorithms to RCA analysis, it is necessary to perform appropriate processing of missing data. In general, there are two methods for handling missing values. One is to remove the incomplete parameters or samples with missing values, or to replace the missing values with appropriate values. In the case of MCAR, the parameters or data samples with missing values can be deleted to ensure a complete data set, while in the case of MAR or MNAR, replacement of missing values is required.
[0013] As such, in the case of MAR or NMAR, the method of removing parameters or samples has the risk of causing partial loss of information in the data, which can lead to biased analysis results, so various methods have been developed to estimate missing values based on observed data in the dataset. [Prior art documents] [Patent documents]
[0014] [Patent Document 1] Korean Patent Publication No. 10-2021-101557 (Name of invention: Method and device for estimating missing intervals in oil well log data) [Patent Document 2] Korean Patent No. 10-2059112 (Name of invention: Internet stream data quality measurement index and profiling method and system thereof) [Patent Document 3] Korean Patent Publication No. 10-2015-38905 (Title of invention: Data pre-processing device and method) Summary of the Invention [Problem to be solved by the invention]
[0015] The present invention has been made to solve the above-mentioned problems, and aims to provide a method for processing missing values of network log data and classifying the root causes of communication defects thereby, which enables the analysis of the root causes of various defects occurring in a network environment by utilizing machine learning and deep learning technologies while securing a complete data set by finding appropriate substitute values according to the characteristics of parameters where missing values exist. [Means for solving the problem]
[0016] In order to achieve the above-mentioned object, the method for processing missing values of network log data of the present invention includes the steps of (a) separating network communication log data collected during a call section into non-time series data and time series data, (b) substituting missing values according to specific parameter types of the non-time series data and the time series data, (c) substituting missing values for non-time series data among parameters other than the parameters processed in (b) through a Gaussian Mixture Model (GMM), and (d) substituting missing values for time series data among parameters other than the parameters processed in (b) through a joint approach.
[0017] In the above-mentioned configuration, a new parameter is generated to represent the serving PCI (Serving Physical Cell ID) parameter, and is represented as 1 if data exists in the serving PCI, and as 0 if the data is missing.
[0018] For RF-related parameters, including Reference Signals Received Power (RSRP), Reference Signal Received Quality (RSRQ) and Signal to Interference plus Noise Ratio (SINR), data samples with missing values are removed, if any.
[0019] If a missing value in a PDCP / RLC summary statistics (Packet Data Convergence Protocol / Radio Link Control statistics summary) parameter has the same meaning as 0, the missing value is replaced with 0.
[0020] Parameters other than the parameters processed in step (b) include PCell Estimated Distance, KPI PCell PDSCH & PUSCH BLER[%], PHY R-BLER Info PDSCH Total Info DL R-BLER[%], and Total Info PHY R-BLER Info PUSCH UL R-BLER[%].
[0021] Step (c) is to convert non-time series data into a dataset with no missing values (D y : Dataset with no missing values) and dataset with missing values (D n : The missing data set (c1) and the non-missing data set (D y ) using a GMM-based EM algorithm, and assigning each instance to the cluster with the closest Euclidean distance to the center of each cluster (c2). Based on the clustering results of (c2), n ), and then the method includes a step of (c3) finding a cluster for each instance, a step of (c4) finding a complete instance that is closest to the instance having the missing value based on Euclidean distance in each cluster found in the step (c3), and a step of (c5) replacing the missing value with the average value of the complete instance found in the step (c4).
[0022] In step (d), the missing values are replaced in a direction that minimizes the two loss values through an iterative training process, so as to minimize the loss value in the process of reconstructing existing time series data having missing values and the loss value generated when the missing values are randomly replaced.
[0023] The method includes the steps of (d1) generating virtual missing values by artificially masking a predetermined ratio of data at random in existing time series data having missing values, (d2) calculating a replacement loss value between the artificially generated missing values and replacement values using MAE (mean absolute error) after a replacement model replaces all the missing values, and (d3) reconstructing observed data present in the existing data through a modeling process and calculating the difference between the observed data and the reconstructed data using MAE to obtain a loss value.
[0024] In order to recognize the artificially generated missing values, we utilize an indicating mask that represents the artificially masked values as 1 and all other values as 0.
[0025] The replacement model is Low-Rank Autoregressive Tensor Completion (LATC), which is based on an autoregressive model that can take into account the overall data flow of time-series data. In this process, multivariate time-series data is converted into a three-dimensional tensor and applied to the autoregressive model.
[0026] As mentioned above, a fully connected neural network (FCNN) is trained on non-time series data with missing values replaced to obtain a model with optimal parameters, and the source of defects in the non-time series data is then classified through this.
[0027] As mentioned above, a 1D-CNN (1-Dimensional Convolutional Neural Network) is trained on the time series data with missing values replaced to obtain a model with optimal parameters, and the source of defects in the time series data is then classified through this. Effect of the Invention
[0028] According to the method for processing missing values of network log data and classifying the root cause of communication defects through the same of the present invention, by carrying out appropriate data pre-processing on network log data to complement its limitations and identifying various root causes of communication defects through a deep learning algorithm that has learned from a huge amount of data, it is possible to significantly reduce cases in which the root cause cannot be clearly identified in network log data that has been determined to be defective due to a high proportion of missing values in the past. [Brief description of the drawings]
[0029] [Figure 1] 1 is an overall flowchart of a method for processing missing values of network log data and classifying the root of communication failures therethrough according to the present invention; [Diagram 2] 1 is a flowchart illustrating a process of imputing missing values in non-time series data according to the method of the present invention. [Diagram 3] 1 is a sequence chart illustrating a process of replacing missing values of time-series data based on a joint approach according to the method of the present invention. [Figure 4] 4 is a flowchart illustrating a non-time series data learning process in the method of processing missing values of network log data and classifying defects based thereon according to the present invention; [Diagram 5] 4 is a flowchart illustrating a time series data learning process in the method for processing missing values of network log data and classifying the root of a communication fault through the missing values according to the present invention; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0030] Hereinafter, a preferred embodiment of the method for processing missing values of network log data and classifying the root cause of communication failures therethrough will be described in detail with reference to the accompanying drawings.
[0031] The present invention aims to solve the underlying classification problem by filling in the missing value problem of a data set that measures a network environment and learning a deep learning algorithm to grasp the multi-dimensional relationship between parameters. To solve the missing value problem, the missing values are replaced with appropriate values that match the characteristics of the data set or each parameter.
[0032] Then, a deep learning algorithm is trained on the improved data set using a supervised learning method, and a multi-class classification model that classifies the roots through the learned algorithm can be obtained. When trained on a large amount of data and learning a wide variety of patterns, the model recognizes multidimensional relationships between various network parameters, and based on this, it can identify the roots of defects that were difficult to classify using existing logic.
[0033] The terms used in the present invention will be briefly explained below.
[0034] Imputation: This refers to replacing missing values. In order to use data in machine learning, missing values must be replaced with specific values. The type of missing value can be determined based on the cause of the missing value in a dataset or the weight of the missing values, and the characteristics of such missing values must be understood to set an appropriate replacement value.
[0035] Gaussian mixture: A mixture with k Gaussian distributions. It can be used to analyze the characteristics of a data set that can be classified into k clusters. Each Gaussian distribution is expressed by the mean, which indicates the center of the cluster, the covariance, which indicates the width of the distribution, and the mixing probability, which indicates how large or small the Gaussian function is.
[0036] Gaussian Mixture Model (GMM): It is one of the machine learning algorithms that can classify data into categories based on the probability distribution of the data. It is possible to find the optimal GMM parameters through the Expectation-Maximization (EM) algorithm, which is an iterative method consisting of expectation and maximization.
[0037] FCNN (Fully Connected Neural Network): A neural network in which the layers that make up the neural network are connected in a dense manner. In other words, it is a neural network in which all the neurons that make up each layer are connected.
[0038] 1D-CNN (1-Dimensional Convolutional Neural Network): A neural network that extracts features from data through a one-dimensional convolutional layer. It is suitable for extracting data characteristics when the data the neural network learns is one-dimensional, such as sequences or time-series data like natural language.
[0039] FIG. 1 is an overall flowchart of the method for processing missing values in network log data and classifying the root cause of communication defects through the same according to the present invention, showing the overall process of completing a data set through preprocessing and training a supervised learning algorithm for root classification.
[0040] As shown in FIG. 1, according to the method for processing missing values of network log data and classifying the root cause of communication faults through the same of the present invention, first, in step S100, network log data of a call section (call time) used to identify the root cause of a fault is divided into non-time series data and time series data.
[0041] For non-time series data, one basis is a function of a set of features, F={feature 1 , feature 2 , , feature n On the other hand, in the case of time series data, the feature that distinguishes one root is a sequence of the same length as the time interval. Examples of non-time series data and time series data are shown in Table 1 and Table 2 below, respectively.
[0042] [Table 1]
[0043] [Table 2]
[0044] Next, in step S200, the missing values are processed based on appropriate grounds for each parameter of the non-time series data and the time series data.
[0045] For example, since the serving physical cell ID (PCI) parameter is meaningful only in terms of the presence or absence of a value rather than the value itself, a new parameter is created to represent this parameter, and is represented as 1 when data is present in the serving PCI and 0 when data is missing.
[0046] RF-related parameters such as RSRP (Reference Signals Received Power), RSRQ (Reference Signal Received Quality) or SINR (Signal to Interference plus Noise Ratio) that indicate the state of the communication environment are sensitive parameters that represent the communication environment. If there are any data samples with missing values, these are removed.
[0047] Substitute a missing value with 0 when the missing value has the same meaning as 0, such as PDCP / RLC summary statistics (Packet Data Convergence Protocol / Radio Link Control statistics summary).
[0048] Furthermore, parameters that do not fall under the above-mentioned parameter-specific conditions (e.g., PCell Estimated Distance, KPI PCell PDSCH & PUSCH BLER [%], PHY R-BLER Info PDSCH Total Info DL R-BLER [%], Total Info PHY R-BLER Info PUSCH UL R-BLER [%], etc.) are divided into non-time series data and time series data as in steps S300 and S500, and respective alternative techniques are applied.
[0049] First, for non-time series data, substitute values are generated based on observed data in a dataset, for example, by searching for substitute values for missing values based on clustering of observed data through a Gaussian Mixture Model (GMM). Figure 2 is a flowchart for explaining the process of substituting missing values for non-time series data in the method of the present invention.
[0050] First, in step S310, the data set is divided into a data set with no missing values (D y : Dataset with no missing values) and dataset with missing values (D n : Missing data set).
[0051] Next, in step S320, a missing data set (D y ) is clustered using the GMM-based EM algorithm, the center of each cluster is found, and each instance is assigned to the cluster with the closest Euclidean distance.
[0052] Next, in step S330, a missing data set (D n ), then find the clusters for each instance.
[0053] Next, in step S340, a complete instance (an observed instance without missing values) that is closest to the instance with missing values is searched for in each cluster based on the Euclidean distance.
[0054] Finally, in step S350, the missing values are replaced with the mean of the complete instances found in step S340.
[0055] Returning to FIG. 1 again, in step S500, time series data is replaced based on a joint approach, and the missing values are replaced in a direction that minimizes the two loss values through an iterative training process in order to minimize the loss values that occur when randomly replacing the loss values and missing values in the process of “reconstructing” existing data.
[0056] FIG. 3 is a sequence chart for explaining the process of replacing missing values of time series data based on a joint approach in the present invention, which is composed of two training methods.
[0057] First, a certain percentage of existing time series data containing missing values is randomly masked, for example 20%, to generate virtual missing values. In Figure 3, data that exists in the existing data but has been artificially masked to become missing is shown in green (thick line). The substitution model substitutes all missing values, at which point the substitution loss rate between the artificially generated missing values and the substitute values is calculated using MAE (mean absolute error) (Process A). To recognize artificially masked values, an indicating mask is used, which represents values masked by the artificial mask as 1 and all other values as 0.
[0058] The observation data present in the existing data is reconstructed through modeling processing. The difference between the original observation data and the reconstructed data is calculated using MAE to obtain the loss value (Process B).
[0059] The roles of Process A and Process B are different. Process A guides the model to find the most appropriate value for the missing value, while Process B guides the model to find the substitute value to converge to the distribution of existing observation data. As the "Imputation Model" shown in blue in Figure 3, various models that substitute for missing values in a time series can be applied.
[0060] For example, Low-Rank Autoregressive Tensor Completion (LATC), which is based on an autoregressive model that can take into account the overall data flow of time series data, can be applied. In this process, multivariate time series data is converted into a three-dimensional tensor and then applied to the autoregressive model.
[0061] Next, with the data set having been pre-processed through steps S300 and S500, i.e., with missing value issues imputed, a deep learning algorithm is used in steps S400 and S600 to classify the root of the defect.
[0062] Specifically, deep learning algorithms suitable for the data types of non-time series and time series datasets are trained, and each dataset is divided into training, validation, and test, and the training and validation of the algorithms are carried out on the training and validation datasets while multiple epochs are repeated. Through repeated training and validation, the hyperparameters of the neural network are tuned to ensure a model with optimal parameters.
[0063] The final classification performance of the model is confirmed by classification on test data and using indicators such as accuracy, recall, and precision. The deep learning neural networks adapted to each dataset are as follows:
[0064] 4 is a flow chart for explaining a non-time series data learning process in the method for processing missing values of network log data and classifying the root of communication faults through the missing values according to the present invention. As shown in FIG. 4, first, FCNN (Fully Connected Neural Network) is trained with non-time series data. FCNN is a type of ANN (artificial neural network), and like MLP (multi-layer perceptron), multiple hidden layers are formed between the input and output of the neural network, and the neurons that make up each layer are all connected.
[0065] In order to prevent overfitting in such a dense neural network, dropout and batch normalization are added to the hidden layer to form a neural network. In each hidden layer, weights and biases are set for the input or data transmitted in the previous layer, and the data is scaled through the activation function before being transmitted to the output or next layer.
[0066] The final output is the probability of each target class obtained through a softmax function. Since this is a multi-class classification problem, the difference between the model's prediction and the target class (i.e., the correct answer) is obtained through a CrossEntropyLoss function, and training proceeds in the direction of minimizing the loss through the Adam optimizer. In this process, the training ends when the specified number of epochs is reached and the minimum loss value is confirmed continuously for the specified number of iterations.
[0067] 5 is a flow chart for explaining a time series data learning process in the method of processing missing values of network log data and classifying the root of communication faults through the same according to the present invention. As shown in FIG. 5, 1D-CNN (1-Dimensional Convolutional Neural Network) is trained on time series data. The method of realizing 1D-CNN time series classification is realized in two stages. In the first stage, the difference of the time series to be classified is measured and the data is converted into feature vectors.
[0068] Then, in the second stage, the classification task is performed, where the algorithms that can be used are very diverse, ranging from the main machine learning algorithms k-nearest neighbors and SVM algorithms to deep learning neural networks.
[0069] The general structure of the 1D-CNN classification model is as follows:
[0070] It goes through a convolutional layer that extracts patterns in the time series data as features, and a max pooling stage that compresses the extracted features into a vector. The extracted and refined feature vector is then input into a fully connected layer where classification work can be performed. Appropriate weights and bias values are secured through the fully connected layer, and the probability of the root class to be classified is secured as the output through the final softmax function. The 'fully connected' part proceeds in the same way as the FCNN. Training proceeds in the direction of minimizing the difference between the model's prediction and the target through the cross entropy error function and Adam optimization.
[0071] Although the missing value processing of network log data and the method for classifying the root of communication failures through the same have been described in detail above, the present invention is not limited to the above-described embodiments and can be implemented in various modified forms within the scope permitted by the technical concept of the present invention.
Claims
1. (a) separating network communication log data collected during a call section into non-time series data and time series data; (b) replacing missing values according to the type of specific parameters of the non-time series data and the time series data; (c) replacing missing values of non-time series data among parameters other than those processed in (b) through a Gaussian mixture model (GMM); (d) replacing missing values of time series data among parameters other than those processed in (b) through a joint approach; A method for handling missing values in network log data, comprising:
2. The method for processing missing values of network log data according to claim 1, characterized in that a new parameter is generated to represent a serving PCI (Serving Physical Cell ID) parameter, and is represented as 1 when data exists in the serving PCI, and as 0 when data is missing.
3. 2. The method of claim 1, further comprising removing any data samples having missing values for RF-related parameters including Reference Signals Received Power (RSRP), Reference Signal Received Quality (RSRQ), and Signal to Interference plus Noise Ratio (SINR).
4. 2. The method of claim 1, further comprising replacing a missing value with 0 in a PDCP / RLC summary statistics parameter when the missing value has the same meaning as 0.
5. 2. The method of claim 1, wherein the parameters other than the parameters processed in step (b) include PCell Estimated Distance, KPI PCell PDSCH & PUSCH BLER [%], PHY R-BLER Info PDSCH Total Info DL R-BLER [%], and Total Info PHY R-BLER Info PUSCH UL R-BLER [%].
6. Step (c) is to convert the non-time series data into a dataset with no missing values (D y : Dataset with no missing values) and dataset with missing values (D n (c1) dividing the data into a dataset with missing values; Missing data set (D y (c2) clustering the images using a GMM-based EM algorithm, finding the center of each cluster, and assigning each instance to the cluster with the closest Euclidean distance; Based on the clustering results of step (c2), the missing data set (D n (c3) finding a cluster for each instance after clustering the instances; (c4) searching for a complete instance that is closest to the instance having the missing value based on Euclidean distance in each cluster searched in (c3); (c5) replacing the missing values with the average value of the complete instances found in (c4); 2. The method of claim 1, further comprising:
7. 2. The method of claim 1, wherein step (d) replaces the missing values in a direction that minimizes two loss values in a process of reconstructing existing time series data having missing values and a loss value generated when the missing values are randomly replaced through an iterative training process to minimize the two loss values.
8. A step of generating virtual missing values by randomly masking a predetermined ratio of data in existing time series data having missing values (d1); After the substitution model substitutes all the missing values, a substitution loss value between the missing values and the substitution values artificially generated by the mean absolute error (MAE) is calculated (d2); (d3) reconstructing the observed data present in the existing data through a modeling process, and calculating the difference between the observed data and the reconstructed data using MAE to obtain a loss value; 8. The method for processing missing values of network log data according to claim 7, comprising:
9. The method for processing missing values of network log data according to claim 8, characterized in that an indicating mask is utilized to recognize the artificially generated missing values, which represents artificially masked values as 1 and all other values as 0.
10. 10. The method of claim 9, wherein the substitution model is a Low-Rank Autoregressive Tensor Completion (LATC) based on an autoregressive model that can consider an overall data flow of the time series data, and in this process, multivariate time series data is converted into a three-dimensional tensor and applied to the autoregressive model.
11. A method for classifying the root cause of defects through missing value processing of network log data, comprising: learning a Fully Connected Neural Network (FCN) with non-time series data in which missing values have been replaced by the method of claim 6 to obtain a model with optimal parameters; and then classifying the root cause of defects in the non-time series data through the model.
12. A method for classifying the cause of defects through missing value processing of network log data, comprising: learning a 1D-CNN (1-Dimensional Convolutional Neural Network) on time series data in which missing values have been replaced by the method of claim 7 to obtain a model having optimal parameters; and then classifying the cause of defects in the time series data through the model.
Citation Information
Patent Citations
Anomaly detection method and device
CN112188534A
Risk prediction device for identifying risk factor
JP2009237914A
Anomaly detection method and device
JP2022539578A
Apparatus and method for preprocessinig data
KR1020150038905A
Method and apparatus for estimating missing section of oil well logging data
KR1020210101557A
