KR20210124811A - Apparatus and method for generating training data for network failure diagnosis - Google Patents

Abstract

Disclosed are a training data generation apparatus and a training data generation method, capable of generating training data to increase accuracy and stably increase a recall factor while reducing bias of an artificial intelligence engine for diagnosing a network failure. According to the present invention, the training data generation apparatus for generating training data for diagnosing a network failure includes: a preprocessing unit for performing classification and labeling operations for each failure type on alert data collected from network devices so as to vectorize the alert data into a predetermined dimension; a first data increasing unit for selecting insufficient data that meets a predetermined criterion among the vectorized data, and interpolating the selected insufficient data to increase the insufficient data; a feature extraction unit for receiving the vectorized data in which the insufficient data is increased as input data, extracting features of the input data and clustering features for each failure type so as to generate compressed data, and performing training to minimize a difference between the input data and output data obtained by restoring the compressed data; and a second data increasing unit for adding predetermined data to a cluster region corresponding to a failure type selected by a user among regions of the compressed data.

Description

Apparatus and method for generating training data for network failure diagnosis

The present invention relates to a network failure diagnosis technology using artificial intelligence, and more particularly, to a learning data generating apparatus and method for generating learning data to be used for learning an artificial intelligence for diagnosing a network failure.

In the case of a conventional network failure diagnosis system, since the presence or absence of a root failure is determined based on a simple rule, there is a problem in that accuracy is lowered. Accordingly, artificial intelligence is being introduced to diagnose network failures. Network failure diagnosis by artificial intelligence determines the presence or absence of a root failure based on empirically learned data, so when the learning data is insufficient, the accuracy is lowered, and there is a problem that an incomplete failure diagnosis model is generated.

When the training data is insufficient, the AI engine learns using the unbalanced training data, causing a problem in that the AI engine is biased to the type of network failure data composed of a large number of groups. If the AI engine that has learned this bias is applied to the actual field, it is highly likely that data belonging to a small number of failure types will not be properly classified. Therefore, even if the accuracy of the artificial intelligence engine for diagnosing network failures is high, a phenomenon may occur that the recall, which is the classification performance of a network failure type (ie, class) with a small number of training data, is sharply lowered. For example, even if the overall accuracy of diagnosing network failures is as high as 90% or more, a phenomenon in which the accuracy is rapidly lowered may occur for types of network failures that do not occur frequently.

Accordingly, it may be considered to adjust the training data, but it is difficult to balance the large amount of data previously collected. In particular, the conventional network failure diagnosis system balances the data by reducing the number of relatively large number of failure type data, rather than increasing the insufficient failure type data to balance the learning data. However, this method can reduce the learning bias of the AI engine by balancing the learning data for each type of disability, but there is a problem in that useful information is lost in the process of removing the data.

Instead of reducing the number of data, you may consider increasing the data of the scarce type of failure. In the field of artificial intelligence, as a method of increasing insufficient data, manipulation such as moving up, down, left and right, symmetrical movement, and rotation of data or transforming a small number of data to generate new synthetic data is used. However, the alarm data collected when a network failure occurs does not explicitly reveal the correlation between each variable unlike the pixel-shaped image, and it is difficult to simply move the data up, down, left and right, or to express it through manipulations such as symmetric movement or rotation. . In addition, it is difficult to selectively generate the type of data desired by the operator in order to increase the insufficient training data. Therefore, it is difficult to guarantee the generalized performance of the artificial intelligence engine for diagnosing network failures when the existing learning data generation and modulation technology is applied.

The present invention has been proposed to solve the above problems, and while reducing the bias of an artificial intelligence engine for diagnosing network failures, an apparatus and method for generating learning data for generating learning data to increase accuracy and stably increase recall Its purpose is to provide

The apparatus for generating learning data for generating learning data for diagnosing network failures according to an embodiment includes a preprocessing unit for vectorizing alert data collected from network devices into a predetermined dimension by performing classification and labeling operations for each failure type. ; a first data increasing unit that selects insufficient data that meets a predetermined criterion from among the vectorized data and increases the insufficient data by interpolating the selected insufficient data; Learning to receive vectorized data with increased insufficient data as input data, extract features of the input data and cluster them by type of failure to generate compressed data, and minimize the difference between the output data from which the compressed data is restored and the input data a feature extraction unit that performs and a second data increase unit for adding predetermined data to a cluster area corresponding to a type of disability selected by a user among the areas of the compressed data.

The first data increasing unit may increase the insufficient data in each of the inner region and the outer region of the insufficient data by interpolating the selected insufficient data.

The first data increasing unit adds data to an inner division point of a line segment connecting two insufficient data in increasing the insufficient data in the inner region, and increases the insufficient data in the outer region by adding data to the line segment connecting the two insufficient data You can add data to the outer point.

The first data increasing unit may adjust the inner region and the outer region by applying a weight.

The feature extraction unit estimates the average and variance of all data by expressing the probability of occurrence for each element element of the vectorized data in which the lack of data is increased as the sum of the probability density function for each type of disorder, and using the estimated mean and variance Compressed data can be created.

The second data increasing unit may add predetermined data to an internal division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected failure type.

A method for generating learning data for diagnosing a network failure in the apparatus for generating learning data according to an embodiment includes performing classification and labeling operations on alert data collected from network devices for each failure type and vectorizing it into a predetermined dimension step; selecting insufficient data that meets a predetermined criterion from among the vectorized data and increasing the insufficient data by interpolating the selected insufficient data; Learning to receive vectorized data with increased insufficient data as input data, extract features of the input data and cluster them by type of failure to generate compressed data, and minimize the difference between the output data from which the compressed data is restored and the input data performing the steps; and adding predetermined data to a cluster area corresponding to the type of disability selected by the user among the areas of the compressed data.

The increasing of the insufficient data may include increasing the insufficient data in each of the inner region and the outer region of the insufficient data by interpolating the selected insufficient data.

The step of increasing the insufficient data includes adding data to the inner division point of a line segment connecting the two insufficient data in increasing the insufficient data in the inner region, and connecting the two insufficient data in increasing the insufficient data in the outer region. Data can be added to the outer division of the line segment.

The increasing of the insufficient data may include adjusting the inner region and the outer region by applying a weight.

The step of performing the learning comprises estimating the average and variance of all data by expressing the occurrence probability for each element element of the vectorized data in which the insufficient data is increased as the sum of the probability density functions for each type of disorder, and the estimated mean and variance can be used to generate compressed data.

In the adding of the data, predetermined data may be added to an internal division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected disability type.

According to the present invention, it is possible to reduce the bias of the AI engine by increasing the insufficient data among the training data of the AI engine for diagnosing network failure, and to increase not only accuracy but also recall rate limitedly. . In particular, rather than simply transforming insufficient data or synthesizing arbitrary types of data from compressed data through dimensionality reduction, the AI engine based on existing imbalanced learning data creates insufficient data by considering the difference in distribution between real and synthetic data. can prevent overfitting.

In addition, in the present invention, in extracting and compressing features from vectorized data through an encoder of an autoencoder, a boundary that can distinguish types of failure within the Euclidean distance is formed, and based on this, the user selects an area lacking data. Then, it is possible to generate balanced learning data for all disability types by transcribed compressed information of the corresponding disability type in the selected region.

1 is a schematic configuration diagram of an apparatus for generating learning data according to an embodiment.
2 is a diagram illustrating an example of a table of alert data according to an embodiment of the present invention.
3 is a diagram illustrating an example of increasing data in an inner area and an outer area of insufficient data according to an embodiment of the present invention.
4 is a diagram comparing conventional compressed data and compressed data according to an embodiment of the present invention.
5 is a flowchart illustrating a method of generating learning data according to an embodiment of the present invention.

The above-described objects, features, and advantages will become more apparent through the following detailed description in relation to the accompanying drawings, whereby those of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention. There will be. In addition, in describing the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted. Hereinafter, a preferred embodiment according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a schematic configuration diagram of an apparatus for generating learning data according to an embodiment. The training data generating apparatus 100 according to the present embodiment may include a memory, a memory controller, one or more processors (CPU), a peripheral interface, an input/output (I/O) subsystem, a display device, an input device, and a communication circuit. have. The memory may include high-speed random access memory, and may also include one or more magnetic disk storage devices, non-volatile memories such as flash memory devices, or other non-volatile semiconductor memory devices. Access to the memory by other components, such as the processor and peripheral interfaces, may be controlled by the memory controller. The memory may store various kinds of information and program instructions, and the program is executed by the processor. Peripheral interfaces connect the I/O peripherals to the processor and memory. One or more processors execute various software programs and/or sets of instructions stored in memory to perform various functions and process data in the system. The I/O subsystem provides an interface between input/output peripherals such as display devices and input devices and the peripheral interface. The communication circuit performs communication through an external port or communication by an RF signal. The communication circuitry converts electrical signals into RF signals and vice versa through which the RF signals can communicate with communication networks, other mobile gateway devices, and communication devices. As shown in FIG. 1 , the learning data generating apparatus 100 includes a preprocessing unit 110 , a first data increasing unit 120 , a feature extracting unit 130 , and a second data increasing unit 140 , and , these components may be implemented as a program, stored in a memory, and executed by a processor, or may be implemented and operated by a combination of hardware and software.

The pre-processing unit 110 receives and pre-processes alert data collected from various network devices according to the occurrence of a failure. The preprocessor 110 may receive alert data from a network management system (NMS) that collects alert data generated from various network devices. The preprocessor 110 may classify the alarm data for each failure, classify the types of alarm data of each failure, perform a labeling operation, convert it into a table form, and store it in a storage unit (eg, a database). The alarm data may include identification information and type information of a network device in which a failure has occurred, type information of an alarm, and hierarchical information of a network device in which a failure has occurred. The types of alerts include, for example, Loss of Signal (LOS), Loss of Pointer (LOP), Remote Defect Indication (RDI), and the like.

The preprocessor 110 vectorizes the alert data for each failure into a predetermined dimension based on the table of alert data generated for each failure. The preprocessor 110 receives dimensional information (eg, one-dimensional, two-dimensional, etc.) from the user, and vectorizes the alert data for each failure according to the received dimensional information. In the case of this embodiment, one-dimensional vectorization is taken as an example. 2 is a diagram illustrating an example of a table of alert data according to an embodiment of the present invention. The table shown in FIG. 2 is a table in which alert data generated in any one failure is arranged. As shown in FIG. 2 , the table includes network devices that generate alert data in the corresponding failure, types of alert data generated by each network device, and hierarchical information of each network device. The preprocessor 110 generates one-dimensional vector data by performing preprocessing such as summing the number of each type of alert data in the table. Accordingly, the one-dimensional vector data includes the type of network equipment, the type of alarm, the number of each type of alarm data, and hierarchical information.

The first data increasing unit 120 selects insufficient data that meets a predetermined criterion from among the data vectorized for each failure in the preprocessing unit 110 , and increases the insufficient data by interpolating the selected insufficient data. Preferably, the first data increasing unit 120 selects, as insufficient data, vectorized data generated from a failure of a failure type in which alarm data is relatively insufficient among a plurality of failure types. For example, vectorized data of the failure type with the smallest number is selected as insufficient data, or vectorized data of the failure type having a number less than a predetermined ratio with respect to the largest number is selected as insufficient data. For example, there are 6 disability types in total, 50 vectorized data for Type 1, 40 for Type 2, 35 for Type 3, 20 for Type 4, and 15 for Type 5. , when the sixth type is 13, 13 vectorized data of the sixth type, which has the smallest number, are selected as insufficient data. Alternatively, vectorized data of types 5 and 6, which are 30% or less compared to type 1, which has the largest number, are selected as insufficient data.

The first data increase unit 120 increases the insufficient data by interpolating the selected insufficient data to increase the insufficient data in each of the inner region and the outer region of the insufficient data. Specifically, the first data increasing unit 120, in increasing the insufficient data in the inner region, adds data to the inner division point of a line segment connecting any two insufficient data, and increases the insufficient data in the outer region. Data can be added to the outer division of the line segment connecting two arbitrary pieces of insufficient data.

이 되고, 외부 영역에 추가된 데이터는 하나의 벡터

가 된다.For example, if the selected scarce data is X ₁ =(α ₁ , β ₁ , γ ₁ , δ ₁ ,...), X ₂ =(α ₂ , β ₂ , γ ₂ , δ ₂ ,...) , the first data increasing unit 120 increases insufficient data according to Equations 1 and 2 below by using preset p, q, s, and r parameters. That is, arbitrary data is added to the inner area and the outer area of the selected insufficient data. The data added to the inner region is one vector.

, and the data added to the external area is one vector.

becomes

(Equation 1)

(Equation 2)

The above (Equation 1) is to add data to the inner division point of a line segment connecting any two insufficient data of the selected insufficient data, and the (Equation 2) is to add any two insufficient data of the selected insufficient data. This is to add data to the outer division of the line segment. The inner division point is an inner area of the selected insufficient data, and the outer division point is an outer area of the selected insufficient data.

The first data increasing unit 120 may adjust the data increase area based on the weight ε received from the user. That is, the first data increasing unit 120 may adjust the data increase region by adjusting each parameter by adding or subtracting the weight ε to the p, q, s, and r parameters. Through this, it is possible to increase the degree of asymmetry to prevent overfitting of the artificial intelligence engine for diagnosing network failures. The first data increasing unit 120 may determine the number of data to increase when adding data based on the number of generated data received from the user. That is, based on the generated number received from the user, by adjusting the number of combinations of any two data that can be combined in insufficient data, the number of data increases may be adjusted.

3 is a diagram illustrating an example of increasing data in an inner area and an outer area of insufficient data according to an embodiment of the present invention. The example of FIG. 3 shows only the data of the failure type 1 and the failure type 2, and the insufficient data is the data of the failure type 2. In the data of the failure type 2, data is added to the inner division 310 of the line segment connecting any two data, and data is added to the outer division 320 of the line segment connecting another arbitrary two data. The positions of the inner and outer points are determined by the values of the p, q, s, and r parameters and the weight ε.

Referring back to FIG. 1 , the feature extraction unit 130 receives, as input data, vectorized data including the data increased by the first data increase unit 120 , and extracts features of the input data for each type of failure. Compressed data is generated by clustering, and learning is performed to minimize the difference between the output data from which the compressed data is restored and the input data. Preferably, the feature extraction unit 130 uses an autoencoder. The autoencoder includes an encoder that extracts features from input data and compresses it, and a decoder that restores and outputs original data from the compressed data, and minimizes the difference between the output data output from the decoder and the input data. proceed with learning.

The feature extraction unit 130 estimates the average and variance of all data by expressing the probability of occurrence for each element element of each vectorized data as the sum of the probability density function for each type of disorder, and compresses it using the estimated mean and variance. create data For example, if there is one vector X ₁ =(x ₁ , x ₂ , x ₃ , x ₄ ,..., x _n ), the number of occurrences of the K failure types of _{x 1} which is a constituent element of the vector will occur. The sum of the probability density functions is as follows (Equation 3), and when the sum of the probability density functions that occur in each of the K disorder types of each element element is summed, the probability density function of the vector X ₁ is calculated, and from this, the vector The mean and variance of X _{1 are estimated.}

(Equation 3)

는 평균이고,

는 분산이며, N은 평균이

이고 분산이

인 연속확률분포이며, π_k는 k번째 장애 유형이 선택될 확률이다.. here,

is the average,

is the variance, and N is the mean

and the dispersion

is a continuous probability distribution where π _k is the probability that the kth type of failure is selected.

In general, an encoder of an autoencoder extracts features from input data and compresses them to generate compressed data. When the encoder of the conventional autoencoder compresses the input data, it is assumed that the input data follows a normal distribution with predetermined mean and variance as parameters and compresses the input data. However, it is not appropriate to compress the input data like a conventional encoder because the alarm data generated in the event of a network failure has a complex distribution rather than a normal distribution. When alert data that does not follow a normal distribution is compressed according to a normal distribution in a conventional manner, the compressed data is mixed and expressed within a Euclidean distance.

On the other hand, as in the present embodiment, the average and variance of all data are estimated by expressing the occurrence probability of each element element of each vectorized data as the sum of the probability density function for each type of disorder, and compressed data using the estimated mean and variance , the compressed data is clustered and expressed by failure type. In other words, since compressed data is clustered within the Euclidean boundary, the operator can control the generation of network failure types of data. 4 is a diagram comparing conventional compressed data and compressed data according to an embodiment of the present invention. Figure 4 (a) is a result of the conventional encoder compressing data based on a normal distribution, the data are not clustered according to the type of failure, data is mixed regardless of the type of failure. On the other hand, as shown in (b) of FIG. 4 as compressed data according to an embodiment of the present invention, data are clustered according to the type of failure.

The second data increasing unit 140 adds predetermined data to a cluster area corresponding to a type of failure selected by a user among areas of compressed data generated by the encoder of the feature extraction unit 130 . Preferably, the second data increasing unit 140 may add predetermined data to an internal division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected failure type. The second data increasing unit 140 may receive the number of data to be added from the user, or may add data by automatically determining the number based on a predetermined rule. As data is added to the cluster area corresponding to the type of impairment selected by the user, data may be finally generated in a balanced manner for each type of impairment.

When data is added to the compressed data generated by the encoder of the feature extracting unit 130 by the second data increasing unit 140 , the decoder of the feature extracting unit 130 restores and outputs the compressed data to which the data is added. do. The output data output from the decoder is used as learning data of the artificial intelligence engine for diagnosing network failures. The data added by the first data increasing unit 120 and the second data increasing unit 140 may be suitable as training data of an artificial intelligence engine for diagnosing a network failure, but may not be suitable as verification data. Accordingly, additional data among the training data output from the decoder is not used as verification data, and some of the remaining data is used as verification data.

According to the above embodiment, the learning data generating apparatus 100 may reduce the bias of the artificial intelligence engine by increasing insufficient data among the learning data of the artificial intelligence engine for diagnosing network failure, and not only accuracy but also recall rate. (recall) can also be raised to a limited extent. In particular, the training data generating apparatus 100 generates insufficient data by considering the difference in distribution between actual data and synthetic data, rather than simply transforming insufficient data or synthesizing arbitrary types of data from data compressed through dimensionality reduction. It is possible to prevent overfitting of the artificial intelligence engine due to unbalanced learning data. The learning data generating apparatus 100 of the present invention forms a boundary that can distinguish the types of failures within the Euclidean distance in extracting and compressing features from the vectorized data through the encoder of the autoencoder, and based on this, the user When a region lacking in is selected, balanced learning data is generated for all disability types by transcribed compressed information of the corresponding disability type into the selected region.

5 is a flowchart illustrating a method of generating learning data according to an embodiment of the present invention.

Referring to FIG. 5 , in step S501 , the learning data generating apparatus 100 receives and pre-processes alert data collected from various network devices according to the occurrence of a failure. Specifically, the learning data generating apparatus 100 may receive alert data from a network management system (NMS) that collects alert data generated from various network devices. The learning data generating apparatus 100 may classify the alarm data for each failure, classify the types of alarm data of each failure, perform a labeling operation, convert it into a table form, and store it in a storage unit (eg, a database). The alarm data may include identification information and type information of a network device in which a failure has occurred, type information of an alarm, and hierarchical information of a network device in which a failure has occurred. The types of alerts include, for example, Loss of Signal (LOS), Loss of Pointer (LOP), Remote Defect Indication (RDI), and the like. The learning data generating apparatus 100 vectorizes the alert data for each failure into a predetermined dimension based on a table of alert data generated for each failure. The learning data generating apparatus 100 receives dimensional information (eg, one-dimensional, two-dimensional, etc.) from a user, and vectorizes alert data for each failure according to the received dimensional information.

In step S502, the learning data generating apparatus 100 selects insufficient data that meets a predetermined criterion among vectorized data for each disability, and increases the insufficient data by interpolating the selected insufficient data. Preferably, the learning data generating apparatus 100 selects, as insufficient data, vectorized data generated from a failure of a failure type in which alarm data is relatively insufficient among a plurality of failure types. For example, vectorized data of the failure type with the smallest number is selected as insufficient data, or vectorized data of the failure type having a number less than a predetermined ratio with respect to the largest number is selected as insufficient data. When the training data generating apparatus 100 increases the insufficient data by interpolating the selected insufficient data, the insufficient data is increased in each of the inner region and the outer region of the insufficient data. Specifically, the learning data generating apparatus 100 adds data to an inner division point of a line segment connecting any two pieces of insufficient data in increasing the insufficient data in the inner region, and arbitrary in increasing the insufficient data in the outer region. Data can be added to the outer division of the line segment connecting the two insufficient data of . The training data generating apparatus 100 may adjust the data increase area based on the weight ε received from the user. That is, the training data generating apparatus 100 may adjust the data increase region by adjusting each parameter by adding or subtracting the weight ε to the p, q, s, and r parameters. Through this, it is possible to increase the degree of asymmetry to prevent overfitting of the artificial intelligence engine for diagnosing network failures. The training data generating apparatus 100 may determine the number of data to be increased when adding and increasing data based on the number of data generated from the user. That is, based on the generated number received from the user, by adjusting the number of combinations of any two data that can be combined in insufficient data, the number of data increases may be adjusted.

In step S503, the learning data generating apparatus 100 receives the vectorized data including the data increased in step S502 as input data, extracts features of the input data and clusters them by type of failure to generate compressed data, Learning to minimize the difference between the compressed data and the restored output data and the input data is performed. Preferably, the training data generating apparatus 100 uses an autoencoder. The autoencoder includes an encoder that extracts features from input data and compresses it, and a decoder that restores and outputs original data from the compressed data, and minimizes the difference between the output data output from the decoder and the input data. proceed with learning. The learning data generating apparatus 100 estimates the average and variance of all data by expressing the occurrence probability of each element element of each vectorized data as the sum of the probability density function for each type of disorder, and compresses it using the estimated mean and variance create data For example, if there is one vector X ₁ =(x ₁ , x ₂ , x ₃ , x ₄ ,..., x _n ), the number of occurrences of the K failure types of _{x 1} which is a constituent element of the vector will occur. The sum of the probability density functions is the same as above (Equation 3), and when the sum of the probability density functions that occur in each of the K failure types of each element element is summed, the probability density function of the vector X ₁ is calculated, and from this, the vector The mean and variance of X _{1 are estimated.}

In step S504, the training data generating apparatus 100 adds predetermined data to the cluster area corresponding to the type of disability selected by the user among the compressed data areas generated by the encoder in step S503. Preferably, the learning data generating apparatus 100 may add predetermined data to an internal division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected disability type. The learning data generating apparatus 100 may receive the number of data to be added from the user, or may add data by automatically determining the number based on a predetermined rule. As data is added to the cluster area corresponding to the type of impairment selected by the user, data may be finally generated in a balanced manner for each type of impairment.

According to the embodiment described with reference to FIG. 5, output data finally output from the decoder is used as learning data of the artificial intelligence engine for diagnosing network failure. The data added twice may be suitable as training data of an artificial intelligence engine for diagnosing network failures, but may not be suitable as validation data. Accordingly, additional data among the training data output from the decoder is not used as verification data, and some of the remaining data is used as verification data.

While this specification contains many features, such features should not be construed as limiting the scope of the invention or the claims. Also, features described in individual embodiments herein may be implemented in combination in a single embodiment. Conversely, various features described herein in a single embodiment may be implemented in various embodiments individually, or may be implemented in appropriate combination.

Although acts are described in the drawings in a specific order, it should not be understood that the acts are performed in the specific order as shown, or that all of the described acts are performed in a continuous order, or to obtain a desired result. . Multitasking and parallel processing can be advantageous in certain circumstances. In addition, it should be understood that the division of various system components in the above-described embodiments does not require such division in all embodiments. The program components and systems described above may generally be implemented as a package in a single software product or multiple software products.

The method of the present invention as described above may be implemented as a program and stored in a computer-readable form in a recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.). Since this process can be easily performed by a person skilled in the art to which the present invention pertains, it will not be described in detail any more.

The present invention described above, for those of ordinary skill in the art to which the present invention pertains, various substitutions, modifications and changes are possible without departing from the technical spirit of the present invention. It is not limited by the drawing.

100: learning data generating device
110: preprocessor
120: first data increase unit
130: feature extraction unit
140: second data increase unit

Claims (13)

A learning data generating apparatus for generating learning data for diagnosing network failure, comprising:
a preprocessor for vectorizing alert data collected from network devices into a predetermined dimension by performing classification and labeling operations for each failure type;
a first data increasing unit that selects insufficient data that meets a predetermined criterion from among the vectorized data and increases the insufficient data by interpolating the selected insufficient data;
Learning to receive vectorized data with increased insufficient data as input data, extract features of the input data and cluster them by type of failure to generate compressed data, and minimize the difference between the output data from which the compressed data is restored and the input data a feature extraction unit that performs and
and a second data increasing unit for adding predetermined data to a cluster area corresponding to a type of disability selected by a user among the areas of the compressed data. The method of claim 1,
The first data increasing unit,
The apparatus for generating learning data, characterized in that by interpolating the selected insufficient data, the insufficient data is increased in each of an inner region and an outer region of the insufficient data. 3. The method of claim 2,
The first data increasing unit,
In increasing the insufficient data in the inner region, data is added to the inner division point of the line segment connecting the two insufficient data,
In increasing the insufficient data in the outer region, the learning data generating apparatus, characterized in that the data is added to the outer division point of the line segment connecting the two insufficient data. 3. The method of claim 2,
The first data increasing unit,
The apparatus for generating learning data, characterized in that the inner region and the outer region are adjusted by applying a weight. The method of claim 1,
The feature extraction unit,
Expressing the occurrence probability of each element element of the vectorized data in which the insufficient data is increased as the sum of the probability density function for each type of disorder to estimate the average and variance of all data, and to generate compressed data using the estimated mean and variance Learning data generating device, characterized in that. The method of claim 1,
The second data increase unit,
The apparatus for generating learning data, characterized in that the predetermined data is added to an inner division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected disability type. A method for generating learning data for diagnosing network failures in a learning data generating device, the method comprising:
performing a classification and labeling operation for each type of failure on the alert data collected from network devices and vectorizing them into a predetermined dimension;
selecting insufficient data that meets a predetermined criterion among the vectorized data and increasing the insufficient data by interpolating the selected insufficient data;
Learning to receive vectorized data with increased insufficient data as input data, extract features of the input data and cluster them by type of failure to generate compressed data, and minimize the difference between the output data from which the compressed data is restored and the input data performing the steps; and
and adding predetermined data to a cluster area corresponding to a type of disability selected by a user among the areas of the compressed data. 8. The method of claim 7,
The step of increasing the insufficient data comprises:
Method according to claim 1, characterized in that by interpolating the selected insufficient data, insufficient data is increased in each of an inner region and an outer region of the insufficient data. 9. The method of claim 8,
The step of increasing the insufficient data comprises:
In increasing the insufficient data in the inner region, data is added to the inner division point of the line segment connecting the two insufficient data,
In increasing the insufficient data in the outer region, data is added to an outer division point of a line segment connecting two pieces of insufficient data. 9. The method of claim 8,
The step of increasing the insufficient data comprises:
and adjusting the inner region and the outer region by applying a weight. 8. The method of claim 7,
The step of performing the learning is
Expressing the occurrence probability of each element element of the vectorized data in which the insufficient data is increased as the sum of the probability density function for each type of disorder to estimate the average and variance of all data, and to generate compressed data using the estimated mean and variance A method characterized in that. 8. The method of claim 7,
The step of adding the data is:
Method according to claim 1, characterized in that predetermined data is added to an inner division point of a line segment connecting any two pieces of data selected in the inner region of the cluster region corresponding to the selected disability type. A computer program stored in a computer-readable recording medium as a computer program for executing the method according to any one of claims 7 to 12 through a computer system.

Images