KR102444841B1 - System for deducting reason of quality of mobile communication network based on artificial intelligence - Google Patents

Abstract

The present invention, the information extraction unit for extracting quality-related information from the communication network quality measurement data; a pre-processing unit performing pre-processing on the quality-related information extracted by the information extracting unit; a model data generator for generating model data to be input into a neural network model from the preprocessed quality-related information; a neural network model for learning to classify a mobile communication network quality class based on the model data generated by the model data generator; an inference unit for generating mapping information indicating a correlation between quality-related information constituting the inspection data and a mobile communication network quality class with respect to inspection data, which is an object to determine a cause of a poor quality; and a report module for providing result information indicating the cause of the quality defect with respect to the inspection data based on the mapping information generated by the inference unit.

Description

SYSTEM FOR DEDUCTING REASON OF QUALITY OF MOBILE COMMUNICATION NETWORK BASED ON ARTIFICIAL INTELLIGENCE

The present invention relates to an artificial intelligence-based mobile communication network quality defect reason inference system, and more particularly, to a system capable of efficiently, quickly and accurately analyzing and inferring the cause of a mobile communication network quality defect based on artificial intelligence. .

A mobile communication network quality measuring device is a device that evaluates the quality of a communication network based on communication-related information for the optimization of a communication network provided by a mobile communication company. Perform tasks such as verification and call quality analysis.

In particular, although 5G mobile communication is rapidly spreading in recent years, the communication network is not yet stable, so the need to accurately measure and analyze the quality of the communication network and take appropriate measures is growing.

As a conventional method of measuring the quality of a communication network, for example, a mobile communication operator measures the field using measuring equipment in roads, public facilities, major buildings, railways, subways, ships, floating distances, and densely populated areas nationwide on a quarterly basis. and optimization engineers have been manually analyzing data from poor quality areas.

Manual analysis process by optimization engineer,

1) Among the vast log data measured periodically, log files of poor quality calls are collected for each region in charge,

2) Infer the cause of the defect using the data chart, table, log message viewer, etc. of the quality analysis tool,

3) When a certain amount of analysis contents in the same area is accumulated, it is a process of taking appropriate measures to improve the communication quality of the area according to the analysis contents.

For example, if the signal of the 5G network is weak and the distance from the adjacent base station is far, take measures to expand the base station, and if the 5G network signal is good and there are enough base stations in the adjacent area, take measures to optimize the base station .

Here, the quarterly measurement data is 5G speed measurement data on a call basis. For example, in the case of KT, one of the domestic 5G mobile communication operators, about 2 million or more measurement data are generated every quarter. Therefore, there is a limit to the task itself of an optimization engineer to manually analyze the location and cause of deterioration in service quality. In addition, there is a problem that quality analysis is not consistent depending on the ability of the analysis engineer.

Republic of Korea Patent Publication No. 10-1518287 (2015.05.07. Announcement)

An object of the present invention is to solve the above problems, and it is an object of the present invention to provide a system capable of efficiently, quickly and accurately analyzing and inferring causes of quality defects in a mobile communication network based on artificial intelligence.

Another object of the present invention is to provide a system capable of providing consistent analysis results by utilizing a vast amount of data related to the quality of a mobile communication network as learning data based on artificial intelligence.

Another object of the present invention is to provide a system that can efficiently utilize data related to the quality of a mobile communication network as learning data.

In order to solve the above problems, the present invention provides an artificial intelligence-based mobile communication network quality defect reason inference system, comprising: an information extraction unit for extracting quality-related information from communication network quality measurement data; a pre-processing unit performing pre-processing on the quality-related information extracted by the information extracting unit; a model data generator for generating model data to be input into a neural network model from the preprocessed quality-related information; a neural network model for learning to classify a mobile communication network quality class based on the model data generated by the model data generator; an inference unit for generating mapping information indicating a correlation between quality-related information constituting the inspection data and a mobile communication network quality class with respect to inspection data, which is an object to determine a cause of a poor quality; and a report module for providing result information indicating the cause of the quality defect with respect to the inspection data based on the mapping information generated by the inference unit.

Here, the information extracting unit may extract quality-related information from the communication network quality measurement data through sampling at a preset period for each measurement call, and add unique identifier information to the quality-related information.

In addition, the model data generator, the preprocessed quality-related information for each unique identifier, consists of two-dimensional matrix data in the form of X × Y, and converts it into three-dimensional matrix data in the form of X × Y × X to model data can be generated (where X is the number of items constituting the quality-related information, and Y is a preset time interval).

In addition, the neural network model includes a feature extraction unit for generating a feature map for the model data, and a classification unit for outputting a classification result for the model data based on the feature map generated by the feature extraction unit, the feature extraction The unit performs a convolution operation at least once by at least one filter on the model data, which is three-dimensional matrix data in the form of X × Y × X, to obtain a feature map that is three-dimensional matrix data in the form of X × Y × N including at least one or more convolutional layers to be extracted, wherein N is the number of channels of a filter used to generate the final feature map, and the classifier is configured to convert the feature map finally generated by the feature extractor to Y quality. A mapping layer for mapping each item of related information may be included, and the classification unit may classify the mobile communication network quality class by weights for N nodes constituting the mapping layer.

In addition, the convolutional layers of the feature extraction unit may be all composed of 3D matrix data in the form of X×Y×N, where N is the number of channels of a filter used to generate each convolutional layer.

In addition, the mapping layer of the classification unit may be obtained by adding all the elements of each 2D matrix to each of the N X×Y type 2D matrix data finally generated by the feature extraction unit and calculating an average value thereof. have.

In addition, the inference unit may generate mapping information based on a final convolution layer that is a feature map finally output from the feature extractor of the neural network model with respect to the inspection data and a weight between the mapping layer and the mobile communication network quality class.

Also, the inference unit may generate mapping information by performing a weighted sum operation on the last convolutional layer transmitted from the feature extraction unit and the weights learned from the classifier of the neural network model.

Advantageous Effects of Invention According to the present invention, it is possible to provide a system capable of efficiently, quickly and accurately analyzing and inferring a cause of a quality defect in a mobile communication network based on artificial intelligence.

In addition, the present invention can provide consistent analysis results by utilizing a vast amount of data related to the quality of the mobile communication network as learning data based on artificial intelligence, and efficiently convert the data related to the quality of the mobile communication network into learning data. A system that can be used can be provided.

In addition, according to the present invention, there is an advantage in that it is possible to analyze a large amount of quality-related data that was not possible when analyzed by a conventional engineer. Since there is a human/time limit in human/time limitations in human/time limitations in analyzing a bad call individually, analysis has been carried out on only some calls, but according to the present invention, there is an effect that a vast amount of quality-related data can be used for learning and analyzed.

In addition, if only the format of input data is matched in various areas such as field measurement using 5G measuring equipment, measurement result upload, data collection, analysis using analysis tools, and web analysis system, the quality of the artificial intelligence-based mobile communication network according to the present invention is poor Since it can be applied to the cause inference system, the system according to the present invention can be freely extended.

In addition, according to the present invention, it is possible to provide a method for detecting, reporting, and responding to real-time events that were not possible when analyzed by conventional engineers. Conventionally, it was difficult to respond in real time because the time of analyzing an event occurring in a specific area was delayed due to the procedures of quality measurement/measurement data upload/data collection/engineer confirmation. A system and method capable of detection may be provided. For example, it will be possible to build a system that automatically tracks log files uploaded from the upload server of measurement data, infers the cause of failure by using the system according to the present invention, and collects and provides the inferred results by region.

1 is a diagram showing the configuration of an artificial intelligence-based mobile communication network quality defect reason inference system 100 according to the present invention.
2 is a diagram illustrating an operation of the information extraction unit 10 .
3 and 4 are diagrams for explaining the configuration and operation of the preprocessor 20 .
5 is a diagram for explaining the operation of the data generator 30 .
6 is a diagram for explaining the configuration of the neural network model 40 of the present invention.
7 is a diagram showing classification accuracy in the neural network model 40 according to the present invention.
8 and 9 are diagrams illustrating a process of generating mapping information in the inference unit 50 .
10 and 11 exemplarily show result information provided by the report module 60 .

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram illustrating the configuration of an artificial intelligence-based mobile communication network quality defect cause inference system 100 (hereinafter, simply referred to as “system 100”) according to the present invention.

Referring to FIG. 1 , the system 100 includes an information extraction unit 10 , a preprocessor 20 , a model data generation unit 30 , a neural network model 40 , an inference unit 50 , and a report module 60 . includes

The information extraction unit 10 performs a function of extracting quality-related information from the communication network quality measurement data.

The pre-processing unit 20 performs pre-processing on the quality-related information extracted by the information extracting unit 10 .

The model data generator 30 performs a function of generating model data to be input to the neural network model 40 from the quality-related information preprocessed in the preprocessor 20 .

The neural network model 40 is responsible for performing learning for classifying a mobile communication network quality class based on the model data generated by the model data generator 30 .

The reasoning unit 50 performs a function of generating mapping information indicating a correlation between quality-related information constituting the inspection data and a mobile communication network quality class for inspection data, which is an object of understanding a cause of a quality defect.

The report module 60 is responsible for providing result information indicating the cause of the mobile communication network quality defect with respect to the inspection data based on the mapping information generated by the inference unit 50 .

The system 100 performs deep learning-based learning to classify the mobile communication network quality class by a plurality of quality-related information by these components, and for inspection data that is a target for identifying the cause of the quality defect By generating mapping information indicating a correlation with the quality-related information, it operates to provide information inferred as a cause of poor quality in the mobile communication network from among the quality-related information.

Hereinafter, these components and overall operations of the system 100 will be described in more detail with reference to FIG. 2 or less.

2 is a diagram illustrating an operation of the information extraction unit 10 .

The information extraction unit 10 performs a function of extracting quality-related information from the communication network quality measurement data, as described above.

Here, the communication network quality measurement data, as log data measured by the mobile communication network quality measurement equipment (refer to (a) of FIG. 2), for example, may be collected from a data server (not shown) provided by the mobile communication network operator.

The information extraction unit 10 extracts quality-related information related to the quality of the mobile communication network from the collected log data, and extracts the main information required from the log data in the specifications provided for each mobile chip vendor of the mobile communication terminal. Accordingly, quality-related information can be extracted as shown in FIGS.

As the quality-related information, as an embodiment, in the present invention, 5G mobile communication network quality measurement data is used, and may be, for example, the following data.

1) RSRP (Reference Signal Received Power)

- Power of the reference signal received by the terminal

- Average value of power measurement of RS (Reference Signal) and loaded RE (Resource Element)

- Used as a reference value for H/O and Cell Reselection

2) RSRQ (Reference Signal Received Quality)

- Ratio of Reference Signal Power to Power received by the terminal

- Since interference is included in the measurement along with signal strength, H/O decision index along with RSRP

3) SINR (Signal to Interference Noise Ratio)

- It is not defined in 3GPP Spec and implemented by each terminal manufacturer

- In general, the quality of RF is judged by the signal power of the serving cell versus the noise based on the RS and PDSCH power of the serving cell.

4) DL/UL MCS (Modulation Coding Scheme)

- It refers to the level that determines how many bits of information to transmit RF on the physical symbol according to the channel report status from the system to the terminal and how to do channel coding to overcome the channel and modulation.

5) Layer

- A value related to RI, which is a value that reports to the system how much independence of paths between Multi Ants is maintained to operate MIMO, means the number of spatially separated layers

6) PDSCH/PUSCH BLER (Block Error Rate)

- Physical channel block error rate

7) PUSCH Power (Physical Uplink Shared Channel Power)

- PUSCH transmission power (size of power transmitted from the terminal to the base station)

8) Rx/Tx Num of RB (Resource Block)

- The minimum number of resources that contain actual data

9) PCI (Physical Cell Identity)

- Unique ID of the cell the terminal is receiving service from

10) SSB (Synchronization Signal Block) Index

- Detect changes in the serviced beam in relation to beamforming

11) Network

- The type of network serviced by the terminal (WCDMA, LTE, 5G, ...)

12) LTE PDSCH/PUSCH Throughput

- Throughput of LTE physical channel

13) 5G PDSCH/PUSCH Throughput

- Throughput of 5G physical channel

14) Data measurement speed

- Throughput of Application Layer

- Used to generate class information of bad (0)/good (1) in the training data preparation stage

The information extraction unit 10 extracts the quality-related information as described above by performing sampling at a preset period (eg, 1 second) from the communication network quality measurement data for each measurement call.

In addition, the information extraction unit 10 adds unique identifier (ID) information for each measurement call to classify each measurement call to the extracted quality-related information. The unique identifier information may be composed of, for example, measurement date and time + terminal phone number + sequence index.

3 and 4 are diagrams for explaining the configuration and operation of the preprocessor 20 .

The pre-processing unit 20, as described above, is a means for performing pre-processing on the quality-related information extracted by the information extraction unit 10, and is input to the neural network model 40 from the model data generation unit 30 to be described later. A function of performing pre-processing, such as processing quality-related information in a form suitable for generating model data to be performed, and generating information on the result of classification of a mobile communication network quality class that acts as a label when learning is performed in the neural network model 40 carry out

To this end, the preprocessor 20 includes, as shown in FIG. 3 , a data processing unit 21 , a data scaling unit 22 , an additional correction unit 23 , and a missing value processing unit 24 as an embodiment. can do.

The data processing unit 21 may, for example, perform the following processing.

1) Using the PCI value, it is processed as a value indicating whether or not the PCI has been changed within the measurement call (pci_cng) - In addition, based on the 5G PDSCH throughput for 3 seconds before and after the PCI change, up to the weight of 3 seconds after the occurrence of pci_cng Calculate and process

2) Using the SSB Index value, it is processed into a value indicating whether or not the SSB Index value has been changed in the measurement call (ssb_cng)

3) Using the network value, it is processed into a value indicating whether the network is serviced in 5G (is_nr5g)

4) Processed as a ratio between LTE PDSCH Throughput and 5G PDSCH Throughput (lte_vs_nr_rate) - LTE PDSCH Throughput / (LTE PDSCH Throughput + 5G PDSCH Throughput)

The data scaling unit 22 may perform robust scaling, which is less sensitive to outliers and outliers, on at least some of the quality-related information extracted by the information extraction unit 10 .

The additional correction unit 23 multiplies the following quality-related information among the data passed through the data scaling unit 22 by -1, so that a large numerical value indicates a large influence on the quality defect.

rsrp, rsrq, sinr, rx_rb, tx_rb, layer, dl_mcs, ul_mcs

Also, since the additional correction unit 23 applies robust scaling, a median may have negative and positive values based on 0, and a negative region indicating good dual network signal may be set to 0. This is to enable the reasoning unit 50 to more clearly infer the cause of the failure later. Some input information that does not significantly affect the quality of 5G experience has some meaningless weight in the learning process, and although its size is small in the inference unit 50, it has an effect. For example, in the case of DL measurement, the influence of UL-side input information is small, and in the case of UL measurement, the influence of DL-side input information is small.

The missing value processing unit 24 is a means for filling items with no values with zeros.

Through this process, the pre-processing unit 20 performs pre-processing of quality-related information so that the classification unit 40 obtains an optimal classification result for the quality-related information extracted by the information extraction unit 10 . can

For example, as described above, when the quality-related information is configured by 5G mobile communication network quality measurement data, the information output from the preprocessor 20 may be configured as follows.

Call ID, RSRP, RSRQ, SINR, DL MCS, UL MCS, Layer, PDSCH BLER, PUSCH BLER, PUSCH Power, Tx Num of RB, Rx Num of RB, PCI change, SSB change, 5G service, LTE and 5G Throughput Rate

4 exemplarily shows the operation of the preprocessor 20 for quality-related information by 5G mobile communication network quality measurement data. As described above, processing, scaling, additional correction, and missing value processing for quality-related information It can be seen that etc. have been performed.

Next, the model data generation unit 30 will be described.

The model data generator 30 performs a function of generating model data to be input to the neural network model 40 from the quality-related information preprocessed in the preprocessor 20 as described above.

Here, the model data input to the neural network model 40 may be training data and verification data used when learning of the neural network model 40 is performed, and inspection data that is a target for determining the cause of the quality defect.

To this end, the model data generating unit 30, as described above, configures the preprocessed quality-related information into two-dimensional matrix data in the form of X × Y for each unique identifier, and sets it to 3 in the form of X × Y × X Transform into dimensional matrix data to create model data.

Here, X is the number of items constituting the quality-related information, and Y is a preset time interval.

5 is a diagram for explaining the operation of the data generator 30 .

Figure 5 (a) shows the quality-related information pre-processed by the pre-processing unit 20 as described above. Here, the left field is unique identifier (ID) information, and information in a red box constitutes quality-related information. Each quality-related information consists of 15 items, which include information sampled every second for 10 seconds for unique identifier information.

This is composed of two-dimensional matrix data in the form of X×Y, as shown in FIG. 5( b ). Here, since quality-related information consists of 15 items, X is 15. As shown in FIG. Also, since the time interval is set to 10 seconds, Y is 10.

Accordingly, two-dimensional matrix data in the form of 15×10 can be obtained.

Here, since the number of times sampled per second may be different for each unique identifier information, in order to make this constant, the following method may be used.

In general, quality-related information is measured in units of 5 or 10 seconds, so set Y to 10, discard quality-related information that is less than 5 seconds, and fill in the remainder with 0 if it is between 5 and 10 seconds. make it data In addition, 10 seconds or more and less than 15 seconds use the 10 seconds after the data, and data for 15 seconds or more are generated by dividing the data into n pieces of 10-second data according to the method described above.

In this way, as shown in (b) of FIG. 5 , the quality-related information pre-processed by the pre-processing unit 20 can be configured as 15×10 2D matrix data.

Next, as shown in (c) and (d) of FIG. 5 , the model data generation unit 30 expands the two-dimensional matrix data in the form of X×Y to X pieces to form 3 pieces of the form of X×Y×X. Create model data by constructing dimensional matrix data.

Here, the expansion of the two-dimensional matrix data in the form of X×Y to X is, as shown in FIG. 5(c), for each of the X items constituting the quality-related information, Y columns belonging to each item. A method of leaving only data and filling all column data belonging to the remaining items with 0 can be used. If this process is performed for each of the X items constituting the quality-related information, as shown in FIG. As shown in d), model data can be generated by composing 3D matrix data in the form of X×Y×X.

This is to enable a separate filter value for each item because the same filter value used in the neural network model 40 is used for each input information when using the two-dimensional matrix data in the form of X×Y. it is for

The three-dimensional matrix data in the form of X×Y×X is input to the neural network model 40 .

In addition, the data generation unit 30 transmits, to the neural network model 40, the mobile network quality class classification result information acting as a label generated by the preprocessor 20 as described above along with this process. It can be used for learning to classify mobile network quality classes.

Next, the neural network model 40 will be described.

As described above, the neural network model 40 is responsible for performing learning to classify the mobile communication network quality class based on the model data generated by the model data generator 30 .

As described above, since the model data generated by the model data generator 30 is 3D matrix data in the form of X×Y×X, it may be viewed in the form of a multi-channel image. Therefore, the neural network model 40 in the present invention preferably uses a deep learning neural network model such as a Convolutional Neural Network (CNN) model suitable for learning and classifying data in the form of multi-channel images.

As is known in the prior art, the CNN model extracts a feature map through convolution from input data in the form of an image, and learns weights to classify the input data based on the feature map. , image classification, and object detection.

Such a CNN model generally includes a feature extractor for extracting a feature map, and a classifier for classifying input data based on the extracted feature map.

Here, the feature extraction unit is configured by disposing at least one convolution layer performing a convolution operation and a pooling layer performing pooling, and the classification unit transforms the feature map into one dimension. It operates to output a classification result for input data by a flatten layer and at least one fully connected layer, and the CNN model uses the feature extractor and the parameters in the classifier. Learning is performed through the process of adjusting so that the error of and weights is minimized.

The CNN model itself is not a direct object of the present invention, and is well known in the prior art, and thus detailed description thereof will be omitted.

In the present invention, a process of extracting a feature map such as a convolutional layer and a pooling layer known in such a CNN model is used, but the following configuration is used.

6 is a diagram for explaining the configuration of the neural network model 40 of the present invention.

Referring to FIG. 6 , the neural network model 40 includes a feature extraction unit 41 and a classification unit 42 .

The feature extraction unit 41 performs a function of generating a feature map for the model data. To this end, the feature extraction unit 41 performs a convolution operation at least once by using at least one filter on the model data, which is three-dimensional matrix data in the form of X×Y×X, to form 3 of the X×Y×N shape. At least one convolutional layer (C1, C2, C3, C4, 411) for extracting a feature map that is dimensional matrix data is included.

Here, N is an integer of 2 or more, and means the number of channels of the filter in the convolution operation. Since the number of channels of the filter may be different for each filter used in the convolution operation, N of each of the convolution layers C1, C2, C3, C4, 411 may have a different value.

Here, each of the convolutional layers C1, C2, C3, C4, and 411 is 3D matrix data in the form of X×Y×N. As described above, N is a variable value because it may be different for each filter, but X and Y are the same values as in the input model data. Accordingly, the kernel size of the filters between the convolutional layers C1, C2, C3, C4, and 411 is 1, and the pooling layer used in the conventional general CNN model is not used.

This is to enable the inference unit 50 to accurately identify a mapping relationship between quality-related information constituting inspection data and a mobile communication network quality class for inspection data, which is a target for determining the cause of quality defects, as will be described later. In this configuration, although classification accuracy may be somewhat lowered, there is an advantage in that it is possible to obtain an accurate arrangement of input quality-related information and weights of each node of the classification unit 42 to be described later.

On the other hand, in the present invention, the methods of extracting the feature map by the convolution layers (C1, C2, C3, C4, 411) and the filter in the feature extraction unit 41 are not a direct object of the present invention, and methods known by the prior art can be used, so a detailed description will be omitted.

Meanwhile, as described above, the classification unit 42 performs a function of outputting a classification result for the model data based on the feature map generated by the feature extraction unit 42 .

To this end, the classification unit 42 maps the feature map finally generated by the feature extraction unit 41 (ie, the last convolutional layer C4) to each item of X pieces of quality-related information. ), and classifies the mobile network quality class by weights (w ₁ , w ₂ , ..., w ₁₅ ) for X nodes constituting the mapping layer 421 .

Here, the last convolutional layer (C4) is composed of two-dimensional matrix data in the form of N X×Y (where N is the number of channels of the filter used when the last convolutional layer is created), and a mapping layer (421) can be obtained by performing global average pooling (GAP) on each of the N pieces of X×Y type 2D matrix data, for example, by adding all elements of the matrix and obtaining an average value thereof.

Alternatively, the mapping layer 421 may be obtained by performing Global Maximum Pooling (GMP) for obtaining the largest value among each element of the matrix on each of the N pieces of X×Y type 2D matrix data.

Accordingly, the mapping layer 421 is composed of a total of N nodes, which can be viewed as 1×N matrix data.

Each node constituting the mapping layer 421 finally has N×2 weights connected to the mobile network quality class, that is, bad and good nodes, respectively. While repeating the process of adjusting these weights, a classification result for the input model data is output, and learning of the neural network model 40 is performed through this process.

The classification unit 42 of this configuration uses a mapping layer 421 without using a flatten layer and a pre-coupling layer that converts a feature map into one dimension when compared with a conventional general CNN model. There is a difference in that the classification result is output.

On the other hand, in the present invention, the method of adjusting the weight by the mobile communication network quality class classification result information acting as a label in the present invention is not a direct object of the present invention, and since a method known by the prior art can be used, this is also Detailed description will be omitted.

7 is a diagram showing classification accuracy in the neural network model 40 according to the present invention.

As shown in (a) and (b) of Figure 7, it can be seen that as the epoch is repeated, the classification accuracy (model accuracy) is gradually increasing, and finally it can be seen that the accuracy exceeds 95%. have.

Through this process, a trained model file may be generated as shown in FIG. 7C , and parameters and weights used for learning are stored therein, which may be used in the inference unit 50 .

Next, the reasoning unit 50 will be described.

As described above, the inference unit 50 performs a function of generating mapping information indicating a correlation between quality-related information constituting the inspection data and a mobile communication network quality class with respect to inspection data, which is a target for determining the cause of the quality defect, as described above. .

The inference unit 50 is a feature map that is finally output from the feature extraction unit 41 of the neural network model 40 learned through the process as described above with respect to the inspection data, that is, the last convolution layer C4 and the previous Based on the weight between the mapping layer 421 and the mobile communication network quality class learned through the process as described above, the mapping information indicating the correlation between the quality-related information constituting the inspection data and the mobile communication network quality class is generated.

Here, the inspection data refers to data that is a target for determining the cause of the quality defect in a state in which learning is completed in the neural network model 40 as described above. When the communication network quality measurement data is input to the information extraction unit 10 in order to identify the cause of the quality defect, as described above, through the information extraction unit 10, the pre-processing unit 20, and the model data generation unit 30, Model data for the inspection data is generated. The model data generated for the test data is input to the neural network model 40 .

The feature extraction unit 41 of the neural network model 40 generates a feature map through the process as described above for the inspection data, and transfers the finally generated feature map, that is, the last convolutional layer, to the inference unit 50 . do. In this case, the feature map for the inspection data is not necessarily transmitted to the classification unit 42 .

8 and 9 are diagrams illustrating a process of generating mapping information in the inference unit 50 .

As shown in FIG. 8 , the inference unit 50 includes the N last convolutional layers C4 transmitted from the feature extraction unit 41 and the weights (w ₁ , w ₂ , ... w ₁₅ ) as described above. Mapping information can be obtained by performing a weighted sum operation. Here, the weights w ₁ , w ₂ , ... w ₁₅ are fixed values determined when learning in the neural network model 40 is completed. That is, the mapping information can be obtained by a weighted sum operation that multiplies each of the N last convolution layers corresponding to each weight, and adds them, with a total of N weights, when the inspection data is classified as defective, As a result, the mapping information becomes matrix data of the same dimension as the input quality-related information, ie, two-dimensional matrix data in the form of X×Y.

When the weighted sum operation is performed in this way, the reasoning unit 50 can accurately map 1:1 which item of the quality-related information constituting the inspection data has an influence on the cause of the quality defect with respect to the input inspection data. Mapping information can be obtained.

The reasoning unit 50 may obtain the mapping information as shown in FIG. 9(b) by performing the weighted sum operation as described above on the test data input as shown in FIG. 9(a).

In the red box in FIG. 9 (b), the weight (w ₁ ) is reflected for the first column "rsrp" among the quality-related information constituting the inspection data, which is the value of the red box in FIG. 9 (a). This indicates that the value has changed.

In this way, when the weighted sum operation of all weights w ₁ , w ₂ , ... w ₁₅ and the last convolution layer C4 transmitted from the feature extraction unit 41 is performed, quality-related information constituting the inspection data It is possible to obtain mapping information indicating a relationship between the and the mobile communication network quality class.

This mapping information represents a 1:1 mapping relationship between the quality-related information constituting the input inspection data and the last convolutional layer C4 transmitted from the feature extraction unit 41. means that

FIG. 9(c) shows a heat map displayed by visualizing such mapping information.

In (c) of FIG. 9 , a darker color is indicated to have a higher value, and therefore it can be intuitively understood that a darker color is a major cause of quality defects.

8 and 9, since it is assumed that the inspection data is poor among the mobile communication network quality classes, the mapping information has been described as indicating the cause of the poor quality, but even when the inspection data is good among the mobile communication network quality classes, the present invention is maintained can be applied. In this case, the mapping information indicates the main cause of good quality.

Next, the report module 60 will be described.

As described above, the report module 60 is responsible for providing result information indicating the cause of the mobile communication network quality defect with respect to the inspection data based on the mapping information generated by the inference unit 50 .

As described above, the mapping information generated by the inference unit 50 is the same size as the quality-related information, that is, two-dimensional matrix data in the form of X×Y.

In the preceding example, since X is 15 and Y is 10, this data indicates which items in the items constituting 15 quality-related information measured for 10 seconds have affected the mobile network quality class.

The report module 60 calculates, for example, five items and probabilities, and five items and a score, for the reasons of failure of the entire call (all 10 seconds) in order for this data, in order for every second, the result information can be created and provided.

10 and 11 exemplarily show result information provided by the report module 60 .

As shown in FIG. 10 , the result information provided by the report module 60 may include, for example, the following.

1) Call_id (unique ID of measurement call)

2) classification prediction accuracy ([probability of bad, probability of good])

- Each probability that the neural network model predicted bad/good

- For example, if [0.9, 0.1], the neural network model predicts that the corresponding input has a 90% probability of being bad and a 10% probability of being good.

3) Defect reason inference information per second

- Top 5 items and scores (0~1) for defective causes every second

- Since each row of mapping information means the second of the measured data, based on this, a defect cause value is generated per second based on the size of the mapping information.

4) Inference information on the cause of all failures in the call

- Top 5 items and probability of failure (0-100%)

- By using the size of the mapping information for each column of mapping information, the total defect cause value of the call is generated.

- In addition, statistical data obtained from the learning data for each item (quartile statistical data, subject: all, bad calls (per call), bad calls (per second))

5) Mapping information

6) Heat map if needed

In addition, the report module 60 may provide such result information in the form of a report as shown in FIG. 11 .

In the above, the present invention has been described with reference to the preferred embodiment according to the present invention, but the present invention is not limited to the above embodiment, and other various modifications and variations are possible.

100...Artificial intelligence-based mobile communication network quality defect reason inference system
10...information extraction unit
20...preprocessor
30...Model data generator
40...Neural Network Model
50...the reasoning part
60...Report module

Claims (8)

As an artificial intelligence-based mobile communication network quality defect reason inference system,
an information extraction unit for extracting quality-related information from communication network quality measurement data;
a pre-processing unit performing pre-processing on the quality-related information extracted by the information extracting unit;
a model data generator for generating model data to be input into a neural network model from the preprocessed quality-related information;
a neural network model for learning to classify a mobile communication network quality class based on the model data generated by the model data generator;
an inference unit for generating mapping information indicating a correlation between quality-related information constituting the inspection data and a mobile communication network quality class with respect to inspection data, which is an object to determine a cause of a poor quality; and
A report module for providing result information indicating a cause of poor quality with respect to the inspection data based on the mapping information generated by the inference unit
including,
The information extraction unit extracts quality-related information from the communication network quality measurement data through sampling at a preset period for each measurement call, and adds unique identifier information to the quality-related information,
The model data generation unit generates model data by constructing the preprocessed quality-related information into two-dimensional matrix data in the form of X × Y for each unique identifier, and converting it into three-dimensional matrix data in the form of X × Y × X (where X is the number of items constituting the quality-related information, and Y is a preset time interval).
delete delete The method according to claim 1,
The neural network model includes a feature extraction unit generating a feature map for model data, and a classification unit outputting a classification result for the model data based on the feature map generated by the feature extraction unit,
The feature extracting unit performs a convolution operation at least once by at least one filter on the model data, which is the three-dimensional matrix data of the X × Y × X form, to obtain X × Y × N three-dimensional matrix data. at least one or more convolutional layers for extracting the feature map, wherein N is the number of channels of the filter used to generate the final feature map;
The classification unit includes a mapping layer that maps the feature map finally generated by the feature extraction unit to each item of X pieces of quality-related information,
The classification unit classifies the mobile network quality class according to the weights for the N nodes constituting the mapping layer.
5. The method according to claim 4,
The convolutional layers of the feature extraction unit are all composed of three-dimensional matrix data in the form of X×Y×N, where N is the number of channels of the filter used to generate each convolutional layer. Inference system for the cause of poor mobile network quality.
5. The method according to claim 4,
The mapping layer of the classification unit is obtained by adding all the elements of each 2D matrix to each of the N X X Y type 2D matrix data finally generated by the feature extraction unit and obtaining an average value thereof An artificial intelligence-based mobile communication network quality inference reasoning system.
5. The method according to claim 4,
The inference unit generates mapping information based on a weight between the last convolution layer, which is a feature map finally output from the feature extracting unit of the neural network model, and the weight between the mapping layer and the mobile communication network quality class with respect to the inspection data An intelligence-based mobile communication network quality defect reason inference system.
8. The method of claim 7,
The inference unit generates mapping information by performing a weighted sum operation on the last convolutional layer transmitted from the feature extraction unit and the weights learned from the classification unit of the neural network model. Defective quality reason inference system.

Images