WO2022065541A1

WO2022065541A1 - Machine-learning-based blocked call detection system and control method thereof

Info

Publication number: WO2022065541A1
Application number: PCT/KR2020/012867
Authority: WO
Inventors: 김종주
Original assignee: 주식회사 지니테크
Priority date: 2020-09-23
Filing date: 2020-09-23
Publication date: 2022-03-31

Abstract

Disclosed are a machine-learning-based blocked call detection system and a control method thereof. A control method of a blocked call detection system according to the present invention comprises the steps of: collecting basic call processing data, including call quality per unit time, about individual calls; generating a distribution plot of call quality on the basis of the call quality per unit time included in the basic call processing data; determining each parameter of an artificial intelligence system by performing machine-learning using, as input values, the generated distribution plot of the call quality and at least one piece of information included in the basic call processing data; and applying at least one among real-time call connection data or CDR data to the artificial intelligent system to determine blocked calls.

Description

Failure call detection system based on machine learning and its control method

The present invention relates to a system for detecting a faulty call and a control method therefor, and more particularly, to a system for detecting a faulty call based on machine learning and a method for controlling the same.

With the recent development of communication technology, wired, internet, and wireless telephones are widely used. However, due to the rapid change in the communication environment, proper equipment expansion is not made or due to various other reasons, communication service users (customers) have reduced call quality and There are many complaints about this.

However, in the prior art, when the call quality is poor, that is, in detecting a faulty call, the operator of the communication service provider relies on manual operation.

Although identifying and resolving problems quickly when they are found contributes to attracting more customers and increasing customer loyalty, there is a problem in that manual operation by the operator is neither quick nor accurate.

In other words, countless changes in the communication environment occur in the middle of a call being connected and the call quality changes at every moment. Manual analysis alone is inevitably inaccurate in determining whether a call is faulty.

(Prior Patent Literature) Korean Patent Publication No. 10-2009-0008948

The present invention has been devised to solve the above-mentioned problems of the prior art, and its object is to calculate appropriate statistical data based on collected call processing-related data, and to provide a system for detecting a faulty call through machine learning and a control method thereof will be.

In order to achieve the above object, there is provided a system for detecting a faulty call according to the present invention, comprising: a data collecting unit for collecting basic call processing data including call quality per unit time for an individual call; a statistical data generator for generating a distribution map of call quality based on call quality per unit time included in the basic call processing data; Each parameter of the artificial intelligence system is determined by performing machine learning using at least one piece of information included in the basic call processing data collected from the basic call processing data and the distribution of call quality generated by the statistical data generator as input values. a machine learning processing unit; and a determination unit that applies at least one of real-time call connection data and CDR (Call Detail Record) data to the artificial intelligence system to determine a faulty call.

In addition, in order to achieve the above object, there is provided a control method of a faulty call detection system according to the present invention, comprising the steps of: collecting basic call processing data including call quality per unit time for an individual call; generating a distribution map of call quality based on call quality per unit time included in the basic call processing data; determining each parameter of an artificial intelligence system by performing machine learning using at least one piece of information included in the basic call processing data and a distribution of the generated call quality as input values; and applying at least one of real-time call connection data and call detail record (CDR) data to the artificial intelligence system to determine a faulty call.

1 is a schematic configuration diagram of an entire system including a faulty call detection system according to an embodiment of the present invention;

Figure 2 is a functional block diagram of the faulty call detection system of Figure 1;

3 is a diagram showing an example of data that can be acquired in real time for an Internet call;

4 is a view showing an example of data in which statistical data generated by the faulty call detection system according to an embodiment of the present invention is additionally reflected in the data of FIG. 3;

5 is a view showing an example of a call quality distribution in a graph form;

6 is a diagram showing the structure of CNN processing;

7 is a view showing a process in which data that can be acquired in real time for an Internet call is accumulated and stored and machine learning is performed based on the accumulated and stored data;

8 is a diagram illustrating a process in which necessary data is accumulated in CDR data stored after a PSTN call is made, and machine learning is performed based on the accumulated data.

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

For convenience, some claims of the claims include alphabets such as '(a)', but these alphabets do not define the order of each step.

A schematic configuration of the entire system including the faulty call detection system 100 according to an embodiment of the present invention is shown in FIG. 1 .

In the drawing, the calling terminal 300 is a terminal making a call to the other party, and the called terminal 400 is a terminal receiving a call from the other party.

As such, the call processing system 200 exists in the middle on the communication path between the calling terminal 300 and the called terminal 400, and the function of the call processing system 200 is to receive a call connection request from the calling terminal 300 and respond accordingly. It performs verification processing and call connection processing to the called terminal 400, and further performs a function of managing call connection and call-related information between the calling terminal 300 and the called terminal 400, that is, call processing data. do.

As such, the process of storing outgoing, incoming, call connection and call processing data between the calling terminal 300, the called terminal 400, and the call processing system 200 corresponds to a known technique, and thus a more detailed description thereof will be omitted.

On the other hand, the faulty call detection system 100 communicates with the above-described call processing system 200 and performs a function of determining whether each call is a faulty call after a preset time elapses after a call connection time or a call connection is made. do.

In particular, for real-time faulty call detection, a device for packet mirroring or port mirroring between the call processing system 200 and the faulty call detection system 100 may be further provided. .

In particular, the faulty call detection system 100 performs a function of detecting whether a faulty call exists after performing machine learning.

FIG. 2 shows an example of a specific functional block of the faulty call detection system 100 .

As shown in the figure, the failure call detection system 100 includes a data collection unit 110 , a data normalization processing unit 130 , a statistical data generation unit 120 , a machine learning processing unit 140 , and a determination unit 150 . can be configured.

First, the data collection unit 110 collects basic call processing data related to each individual call.

Here, the call processing data may be real-time data at the time when a call connection occurs, or may be a kind of call detail record (CDR) data stored after the call is terminated.

In particular, the basic call processing data collected by the data collection unit 110 may include call quality per unit time, and further include a caller identification number, called party identification number, call start time, call end time, or call time. .

In addition, the basic call processing data collected by the data collection unit 110 may include customer claim information corresponding to the call, which is information indicating whether there was a customer complaint or dissatisfaction related to the call quality of the call. .

Here, the call quality per unit time means the call quality measured at a preset time interval. For example, if a call is made for 100 seconds and the preset time interval is 20 seconds, a total of five call quality is measured.

Here, the call quality may correspond to at least one of LQ (Listening Quality), which is a value that does not consider delay occurring during a call, and CQ (Conversation Quality), which is a value that considers delay occurring during a call, such call quality measurement value or Since the measurement method itself corresponds to a known technique, a detailed description thereof will be omitted.

However, in the present embodiment, call quality may be divided into uplink and downlink, respectively.

Of course, the call quality may be separately generated in another configuration according to the present embodiment, but in the present embodiment, it is assumed that the call quality is included in the basic call processing data as an example.

In addition, the caller identification number and the called party identification number may include identification numbers for devices located on a communication path as well as identification numbers for each call target terminal.

That is, the caller identification number may include an identification number of the calling terminal 300 (eg, a phone number of the calling terminal 300) as well as identification numbers of devices corresponding to the calling party, and the called party identification number includes the called terminal ( 400) (for example, the phone number of the called terminal 400), as well as identification numbers of devices corresponding to the called party may be included.

Accordingly, even with the same call quality, when each communication path is different, machine learning can be performed including a case where the customer feels different.

The statistical data generator 120 performs a function of generating a distribution map of call quality based on call quality per unit time included in the basic call processing data.

Specifically, the statistical data generating unit 120 selects a preset call quality section for each call quality per unit time included in the basic call processing data, and uses the ratio of each call quality section for the entire call to the distribution of call quality can create

For example, based on call quality 100, the call quality section is divided into 5 levels: less than 30, between 30 and less than 50, between 50 and less than 70, between 70 and less than 90, and above 90. , the statistical data generating unit 120 determines to which section of these five steps the measured call quality per unit time (for example, a time preset to 100 ms, 1 s, 10 seconds, etc.) belongs, and measures over the entire call A distribution map of call quality can be generated by determining the ratio belonging to each section for the call quality.

3 shows an example of basic call processing data, and FIG. 4 shows a state in which data related to the distribution of call quality is added using the basic call processing data. That is, in FIG. 4 , the data added in FIG. 3 is shown in shading.

3 and 4 illustrate a logical structure of data, it does not necessarily correspond to a table included in a specific database, and at least some of the data may be stored or generated in a separate form in a memory or the like.

In addition, in FIG. 5, the call quality distribution for each of the uplink and the downlink is shown in graphic form.

5, in the case of uplink and downlink, the case of less than 30 among the total call quality per unit time collected is 11.11%, the case of 30 or more to less than 50 is 11.11%, and the case of 50 or more to less than 70 is 0%, It can be seen that 44.44% of the cases were 70 or more and less than 90, and 33.33% of the cases were 90 or more.

In FIG. 5, both the uplink and the downlink show the same distribution as an example, but the call quality distribution of the uplink and the downlink may be different depending on the communication environment.

The data normalization processing unit 130 sets at least one piece of information included in the basic call processing data collected by the data collection unit 110 and the distribution of call quality generated by the statistical data generation unit 120 to the same length according to a preset algorithm. It performs the function of performing normalization processing so that it becomes .

For example, the data normalization processing unit 130 extracts or receives pre-stored call processing data and the generated call quality distribution map, and performs a function of performing normalization processing to have the same length according to a preset algorithm for each field.

Here, the data normalization processing unit 130 may extract the call processing data when it is stored by itself, or may receive it from the external server 200 .

In particular, the data normalization processing unit 130 normalizes the call processing data and the call quality distribution for each field. For example, the normalization size for the outgoing phone field is 15 digits, and the outgoing phone number included in the call processing data is In the case of '010-5555-1111', an additional '0' is inserted to align 15 digits to this, and it can be made in the form of '000001055551111'.

In particular, as described above, the data normalization processing unit 130 does not normalize only the extracted/received basic call processing data, but also normalizes the distribution of call quality generated by the statistical data generation unit 120. .

The machine learning processing unit 140 performs machine learning with at least one piece of information included in the basic call processing data collected from the basic call processing data and the distribution of call quality generated by the statistical data generating unit 120 as an input value. It performs the function of determining each parameter of the artificial intelligence system.

In particular, when the normalization processing for each field is performed by the normalization processing unit as described above, the machine learning processing unit 140 performs machine learning using the normalized basic call processing data and the normalized call quality distribution as input values to perform artificial intelligence. Each parameter of the system can be determined.

That is, in the case of an artificial intelligence system, particularly, deep learning, which is a type of machine learning, the result may vary depending on the parameter value of each layer constituting the neural network configuration. It determines the value and performs the function to be reflected in the corresponding artificial intelligence system.

Since the process of machine learning corresponds to the process of calculating parameter values (eg, matrix values) in each layer of the artificial intelligence system, itself corresponds to a well-known technology, and thus a more detailed description will be omitted.

The machine learning processing unit 140 performs a unique function in processing machine learning, that is, after the data normalized by the data normalization processing unit 130 is formed into one-dimensional image data, the one-dimensional image data By performing machine learning by CNN (Convolutional Neural Network), each parameter of the CNN can be determined and reflected.

For the process of forming a one-dimensional image, if some of the data of FIG. 4 is used as an example, the data normalization processing unit 130 collects the outgoing IP (CALLER_IP), the incoming IP (CALLEE_IP), in response to the uplink. If it is less than 30, from 30 to less than 50, from 50 to less than 70, from 70 to less than 90, from 90 or more among the total call quality per unit time, it is less than 30 out of the total call quality per unit time collected in response to the downlink, 30 or more and less than 50, 50 or more and less than 70, 70 or more and less than 90, and 90 or more, respectively '121.111.0.1', '212.0.0.112', '11.11', '11.11', '0', respectively ', '44.44', '33.33', '11.11', '11.11', '0', '44.44', and '33.33' After creating '1111', '0000', '4444', '3333', '1111', '1111', '0000', '4444', '3333' Data '1211110000012120000001121111111100004444333311111111000044443333' is created and this is converted into one-dimensional image data.

Here, the one-dimensional image refers to an image in which pixels are connected to each other only in one direction (eg, a horizontal direction) and pixels are not connected to each other in the other direction (eg, a vertical direction).

The machine learning processing unit 140 that has performed the one-dimensional image processing in this way determines and reflects each parameter of the CNN by performing machine learning on the one-dimensional image by the CNN.

A process of processing a one-dimensional image with n pixels by applying the CNN algorithm is shown in FIG. 6 .

Referring to FIG. 6 , the configuration of the neural network, which is an AI (artificial intelligence) model, includes an input, a layer, a prediction result, a target, a loss function, and an optimizer. ), where the input means that the above-described real-time data or CDR data and data encoded in a normalized manner with respect to the call quality distribution are input.

In addition, a layer is a layer constituting a neural network, and can be modeled to be optimized for the Layer 1 Dimension CNN algorithm to be suitable for real-time data processing or CDR data and call quality distribution processing.

The loss function is an important component that defines the feedback signal to be used for learning. According to the deep learning guidelines, binary crossentropy for two class classification and categorical crossentropy for multiple class classification ), mean square error in the case of regression, and CTC (connection temporlclassification) in the case of a sequence, and since multiple classes exist, categorical crossentropy may be applied.

The optimizer is a component that determines the learning progress method, and determines the weight update of the neural network based on the loss function, and may apply stochastic gradient descent (SGD).

As described above, the data that has been normalized is composed of a one-dimensional image array to enable CNN modeling, and an optimal value for the layer parameter can be derived by machine learning through iterative convolution operation on the image.

However, in performing machine learning, information (AI_LABEL in FIG. 4) on whether or not a faulty call is determined is required for a kind of labeling, and the machine learning processing unit 140 determines a faulty call based on customer claim information included in the basic call processing data. can determine whether

That is, the machine learning according to the present embodiment performs various layer parameters for determining a faulty call, and an expected result (Result Y' in FIG. 6 ) and a faulty call determination value (Target Y in FIG. 6 ) for data input during machine learning ), learning takes place in a direction in which the prediction difference gradually decreases. In other words, if there is a customer claim, it is determined as a faulty call.

As another example, the machine learning processing unit 140 may be based on a ratio (hereinafter, referred to as a 'low quality ratio') that is less than or equal to a preset reference quality among call quality per unit time for a failure call determination value necessary for performing machine learning.

For example, if the call quality value is stored at 20 second intervals for a call made for 10 minutes, and the percentage of call quality that is less than or equal to a preset value (eg 50 points) based on 100 points is 15% or more A call can be judged as a disability call.

Furthermore, the machine learning processing unit 140 may determine the failure call determination by combining the above-described customer claim information and the low quality ratio.

For example, when the customer's claim information is input as high, medium, or low by the manager, the failure call determination can be determined by combining the low quality ratio and the scores assigned to each of the high, middle, and bottom at a preset ratio.

When the configuration shown in FIG. 6 is made, the process itself in which machine learning is performed according to the CNN algorithm corresponds to a known technology, and thus a more detailed description will be omitted.

Meanwhile, the determination unit 150 performs a function of determining a faulty call by applying at least one of real-time call connection data and call detail record (CDR) data to the artificial intelligence system.

That is, as described above, after each parameter of the artificial intelligence system is determined by machine learning, the determination unit 150 transmits real-time call connection data or CDR data as an input value to the corresponding artificial intelligence system to determine whether a faulty call exists. will be.

In particular, the determination unit 150 may classify and process the processing according to the call connection method. If the call connection is a connection through the Internet, after extracting real-time call connection data, the corresponding call connection data is applied to the artificial intelligence system to apply the faulty call , and if the call connection is through a Public Switched Telephone Network (PSTN), the error call can be determined by applying the CDR information stored after the call connection is terminated to the AI system.

As such, the overall processing method according to the call connection method is illustrated in FIGS. 7 and 8 .

7 shows a processing process when a SIP call is generated through the Internet.

Referring to the same figure, when a packet transmitted and received to and received from the call processing system 200 is received through packet mirroring, the failure call detection system 100 extracts real-time call-related information from the real-time extraction module of the real-time data extraction block to detect AI It is transmitted to the block, and the AI detection block detects a fault signal by applying it to a pre-established AI model (that is, corresponds to an artificial intelligence system in which the above-described parameters are determined and reflected), as well as machine learning processing for the corresponding real-time data. do additional

Further progress of such machine learning processing means updating the parameters of the artificial intelligence system, and accordingly, it is possible to continuously track and manage even when the faulty call pattern is changed.

8 illustrates a processing process when a call is generated through the PSTN.

Referring to the same figure, the CDR collection block of the failure call detection system 100 periodically collects the CDR data accumulated and stored after the PSTN call is generated, and transmits it to the AI detection block. In the AI detection block, the CDR data is previously built It is applied to the AI model that has been used to detect faulty calls and additionally performs machine learning processing on the CDR data.

Meanwhile, it goes without saying that the process of performing each of the above-described embodiments may be performed by a program or application stored in a predetermined recording medium (eg, computer-readable). Here, the recording medium includes an electronic recording medium such as a random access memory (RAM), a magnetic recording medium such as a hard disk, and an optical recording medium such as a compact disk (CD).

In this case, the program stored in the recording medium may be executed on hardware such as a computer or smart phone to perform each of the above-described embodiments. In particular, at least one of the functional blocks of the fault call detection system 100 according to the present invention described above may be implemented by such a program or application.

In addition, the present invention is not limited to the specific embodiments described above, but can be practiced with various modifications and modifications within the scope without departing from the gist of the present invention.

In particular, although the above-described embodiment mainly describes the case in which a call is connected, it goes without saying that a case in which the call is not properly connected and fails may be included. In this case, the call quality information per unit time may be filled with blanks.

According to the present invention, it is possible not only to increase the accuracy of detecting a faulty call, but also to make it possible to detect a faulty call even when the faulty call pattern is changed by automation through machine learning for the faulty call pattern.

In particular, in performing machine learning, since the call time of each individual call is quite variable, if the call quality per unit time for each call is used as it is, the accuracy of the artificial intelligence module generated as a result of machine learning is considerably lowered. When the distribution of call quality is first generated based on the call quality per hour, the normalization process can be consistently performed, and the accuracy of the artificial intelligence module generated as a result of machine learning is increased.

Furthermore, the accuracy of the artificial intelligence module can be improved by performing machine learning based on the failure call determination in consideration of the customer claim information and the low quality ratio at the same time.

Claims

(a) collecting basic call processing data including call quality per unit time for individual calls;

(b) generating a distribution map of call quality based on call quality per unit time included in the basic call processing data;

(c) each parameter of the artificial intelligence system by performing machine learning using at least one piece of information included in the basic call processing data of step (a) and the distribution map of call quality generated in step (b) as input values determining;

(d) applying at least one of real-time call connection data and CDR (Call Detail Record) data to the artificial intelligence system to determine a faulty call.
According to claim 1,

Before step (c),

(e) performing normalization processing so that at least one piece of information included in the basic call processing data of step (a) and the distribution of call quality generated in step (b) have the same length according to a preset algorithm; including more,

In the step (c), the control method of a fault call detection system, characterized in that each parameter of the artificial intelligence system is determined by performing machine learning using the normalized data as an input value.
3. The method of claim 2,

The step (b) is,

(b1) selecting a preset call quality section for each call quality per unit time included in the basic call processing data;

(b2) a control method of a faulty call detection system comprising the step of generating a distribution map of call quality using the ratio for each call quality section for the entire call.
3. The method of claim 2,

In step (c), after forming the data normalized in step (e) into one-dimensional image data, machine learning is performed on the one-dimensional image data by a Convolutional Neural Network (CNN) to each of the CNNs. A control method of a faulty call detection system, characterized in that the parameter is determined.
3. The method of claim 2,

The basic call processing data of step (a) includes customer claim information corresponding to the call,

The method for controlling a faulty call detection system, characterized in that the determination of a faulty call necessary for performing the machine learning in step (c) is based on customer claim information included in the basic call processing data.
3. The method of claim 2,

The method for controlling a faulty call detection system, characterized in that the determination of a faulty call necessary when performing the machine learning of step (c) is based on a ratio of the call quality per unit time that is less than or equal to a preset reference quality.
3. The method of claim 2,

In step (a), the basic call processing data includes a caller identification number, called party identification number, call start time, call end time, and call quality per hour.
a data collection unit for collecting basic call processing data including call quality per unit time for individual calls;

a statistical data generator for generating a distribution map of call quality based on call quality per unit time included in the basic call processing data;

Each parameter of the artificial intelligence system is determined by performing machine learning using at least one piece of information included in the basic call processing data collected from the basic call processing data and the distribution of call quality generated by the statistical data generator as input values. a machine learning processing unit;

and a determination unit for determining a faulty call by applying at least one of real-time call connection data and CDR data to the artificial intelligence system.
9. The method of claim 8,

A normalization processing unit that performs normalization processing so that at least one piece of information included in the basic call processing data collected by the data collection unit and the distribution of call quality generated by the statistical data generation unit have the same length according to a preset algorithm, further including,

The machine learning processing unit performs machine learning using the normalized data as an input value to determine each parameter of the artificial intelligence system.
10. The method of claim 9,

The statistical data generating unit selects a preset call quality section for each call quality per unit time included in the basic call processing data, and generates a distribution map of call quality using the ratio for each call quality section for the entire call. A fault call detection system, characterized in that.