KR20220164158A - Method for alarming danger of disease infection employing sensing of same location based in cloud computing and deep learning - Google Patents

Abstract

An embodiment relates to a disease infection risk notification method using co-location detection based on cloud computing and deep learning. Specifically, the method relates to a co-location detection system using mobile magnetometer data, wherein a co-location detection task is executed on magnetometer reading values by cloud computing and deep learning-based approaches. Therefore, one of most decisive ways to prevent or mitigate epidemics and an outbreak of the epidemics is provided as a co-located detection system. In addition, a magnetometer is one particularly useful solution by achieving high accuracy similar to GPS and also consuming significantly less battery.

Description

Method for alerting danger of disease infection employing sensing of same location based in cloud computing and deep learning}

The content disclosed herein relates to a technology for alarming the risk of disease infection, and more particularly, to a technology that allows people in the vicinity to be notified of the risk of disease infection through a mobile terminal when a specific user is infected. it's about

Unless otherwise indicated herein, material described in this section is not prior art to the claims in this application, and inclusion in this section is not an admission that it is prior art.

Respiratory-related infectious viruses such as swine flu, MERS, and COVID-19 are highly contagious and cause many victims.

Recently, as the number of confirmed cases of COVID-19 is increasing, transmission to various places is becoming a reality. If the number spreads further, it may be difficult to accommodate all confirmed cases, and in this case, first of all, not only confirmed cases but also contacts to prevent the spread of the Corona 19 virus Suspected persons must also be quarantined.

And, in this situation, it is safe to avoid gatherings in advance, but in exceptional cases, it is also important to keep the people around you away from each other and make them feel safe.

Thus, the disclosure herein provides one of the most decisive ways to prevent or mitigate the occurrence of these epidemics and epidemics.

Prior art in this background is to the extent that it is shown in the following patent documents.

(Patent Document 1) KR102193757 Y1

In addition, the above prior art confirms the situation, such as mild or suspected degree of disease among confirmed patients for a disease invented by a virus, etc., through a mobile phone and is helpful.

The disclosed content provides a disease infection risk notification method using cloud computing and deep learning-based co-location detection, which can provide one of the most decisive ways to prevent or mitigate the outbreak of epidemics and epidemics as a co-location detection system. want to do

And, in this case, mobile magnetometers can achieve high accuracy similar to GPS, as well as significantly lower battery consumption, providing one of the particularly useful solutions.

The disease infection risk notification method using cloud computing and deep learning-based co-location detection according to the embodiment,

It relates to a co-location detection system using mobile magnetometer data, characterized in this case by performing co-location tasks on magnetometer readings from cloud computing and deep learning-based approaches.

According to embodiments, co-location detection systems provide one of the most critical ways to prevent or mitigate epidemics and outbreaks. And, at this time, magnetometers are one of the particularly useful solutions because they can achieve high accuracy similar to GPS as well as consume significantly less battery.

1 is a diagram for schematically explaining a disease infection risk notification method using cloud computing and deep learning-based co-location detection in one embodiment.
2 is a view showing the entire system to which a disease infection risk notification method using cloud computing and deep learning-based same-location detection according to an embodiment is applied as a whole.
3 is a block diagram showing the configuration of a management information processing device to which a disease infection risk notification method using cloud computing and deep learning-based same-location detection according to an embodiment is applied.
4 is a flow chart sequentially illustrating the operation of a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment.
5 is a view for explaining a window sliding operation applied to a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment.

1 is a diagram for schematically explaining a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment.

As shown in FIG. 1 , a disease infection risk notification method using co-location detection based on cloud computing and deep learning according to an embodiment relates to a co-location detection system using mobile magnetometer data.

These co-location detection systems are one of the most crucial ways to prevent or mitigate epidemics and outbreaks. While many sensors such as GPS can be deployed to accomplish this task, magnetometers are one particularly useful solution because they can achieve high accuracy similar to GPS, while consuming significantly less battery.

Specifically, it uses cloud computing and deep learning-based approaches to perform co-location sensing tasks on magnetometer readings.

Additionally, Figure 1a shows the relationship between co-location and other entities as the proposed system.

Specifically, in the method for notifying risk of disease infection according to an embodiment, the user's mobile phone first collects three-axis magnetometer data in the background, and manages the data whenever it is connected to the Internet, the information processing device 100, that is, Send to RESTful API server.

This server controls and secures all incoming requests from clients, and data received by the server is stored inside the database for later use.

Now, assuming that an arbitrary user is infected with an infectious disease, this user first notifies the server of the infection (see the right side of FIG. 1).

The server then fetches the user's magnetometer data and other user's data from the database and passes them to the co-location detector.

Next, after performing the same location detection algorithm, if someone is at high risk of getting the disease, the result is sent to the server.

And, if there is a user with a high risk of getting the disease, the server notifies the user of the result.

Next, in this disease infection risk notification method, when data is received from the server, as shown in b of FIG. 1, a deep learning model, that is, an autoencoder, is learned using the magnetic field data of the infected user. For reference, b of FIG. 1 shows the inside of the co-location detector.

When the learning process is complete, we send data from other users to the autoencoder to check whether they are in the same location.

This can be done by analyzing the mean square error (MSE) values between the original magnetometer data and the reconstructed magnetometer data of different users.

The co-location detector notifies the server when a user who has been in the same location as an infected user or a user who has a high risk of being infected with a disease is found.

Meanwhile, additionally, these autoencoders will be described in more detail.

First, if such an autoencoder can reconstruct magnetometer data of another user with a low MSE value after reconstructing magnetometer data from an infected user, it can infer that the two magnetometer data have the same or similar shape. So, this means that the two magnetometer data are likely to be in the same location. Two advantages of this approach are:

1. Unlike traditional approaches, the infected user's data can be used only once to train the model and used to perform colloquial detection on as many other users' data as desired, so that the infected user's data can be used again with the data of other users. No need to compare.

2. Unlike Euclidean distance or other similar approaches, co-location detection accuracy is increased because the two or more magnetometer data are not aligned or affected by time differences. The proposed system with this advantage can indicate that two people were at the same place at the same time, but cannot confirm that they were at the same place even though they were actually at the same place. To solve this problem, cut the infected user data into data with a window size of 5 minutes, train the model, feed the data of other users for the same time ㅁ1 minute, and analyze the MSE to complete the investigation of all data. Repeat the process until

Fig. 1c shows an autoencoder model of the same position detector, and a typical deep autoencoder infrastructure (input layer, hidden layer, latent space representation and output layer) is used.

We used 480 neurons for the input/output layer, which corresponds to the size of a sliding window representing 480 data points or 2 minutes of data.

In addition, a hidden layer with 60 neurons was implemented, and 10 neurons were also used for latent space representation.

2 is a diagram showing a system as a whole to which a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment is applied.

As shown in FIG. 2, the system according to an embodiment includes a user mobile terminal located in a plurality of different areas, a nearby mobile terminal 110, and a management information processing device connected to the mobile terminal 110 through a private network ( 100).

Additionally, it includes a hospital management information processing device and a police station management information processing device that are connected to the management information processing device 100 and provide additional medical services and safety services.

The mobile terminal 110 , that is, whenever the mobile terminal 110 of the user and a neighbor connects to the Internet, collects 3-axis magnetometer data on its own and transmits it to the management information processing device 100 .

The management information processing device 100 collects three-axis magnetometer data from a mobile terminal 110 such as a user to generate and register a deep learning model according to an embodiment (the deep learning model will be described later). In this state, when the user is infected, the management information processing device 100 is notified of the fact from the user's mobile terminal. Therefore, when the notification is received, the magnetic field data of the infected user is learned and reconstructed from the deep learning model. Therefore, it is checked whether the reconstructed user's magnetic field data and the magnetic field data of other nearby users are at the same location according to a preset same location detection format. As a result of the check, if the reconstructed user's magnetometer data and other user's magnetometer data are in the same location, the corresponding user's mobile terminal is notified of the result, and if they are not in the same location, the result is not notified.

3 is a block diagram showing the configuration of a management information processing device to which a disease infection risk notification method using cloud computing and deep learning-based same-location detection according to an embodiment is applied.

As shown in FIG. 3, the management information processing device 100 according to an embodiment includes an I/F unit 101, a RESTful API server 102, the same location detector 103, and a database 104. include

The I/F 101 communicates with an external mobile terminal, that is, a client, collects 3-axis magnetometer data, and notifies the fact of risk of infection.

The RESTful API server 102 controls all received requests from clients and maintains security, and stores the collected data in the database 104 for later use. Then, when the user is infected with an infectious disease, the RESTful API server 102 is notified of the infection through the user's mobile terminal. Then, the RESTful API server 102 retrieves the user's magnetometer data and other user's data from the database 104 and delivers them to the same location detector. At this time, before transmitting the magnetometer data, the data is pre-processed through the above-described deep learning model to be reconstructed so that it can be applied to the same position detector model.

The co-location detector 103 determines whether or not they are in the same location by performing a preset co-location detection algorithm on the magnetic field data of the user and the person around them transmitted from the RESTful API server 102 . So, if you are in the same location or if someone is at high risk of getting a disease, the resulting value is sent to the RESTful API server 102. And, through this, the RESTful API server 102 notifies the result to the user when there is a user with a high risk of getting a disease.

The database (!04) accumulates 3-axis magnetometer data collected through an external mobile terminal or stores related user registration information.

4 is a process flow chart showing operations of a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment in order.

As shown in FIG. 4, in the method for notifying risk of disease infection according to an embodiment, whenever a mobile terminal 110-1 or 110-2 such as a user is connected to the Internet, the mobile terminal 110-1 or 110 -2) collects the 3-axis magnetometer data and creates and registers the deep learning model below.

Next, when the user is infected in this state, the fact is notified from the user mobile terminal 110-1.

Then, when this notification is received, the magnetic data of the infected user is learned and reconstructed from the deep learning model.

Therefore, it is checked whether the reconstructed user's magnetic field data and the magnetic field data of other nearby users are at the same location according to a preset same location detection format.

As a result of the check, if the reconstructed user's magnetometer data and other user's magnetometer data are in the same location, the corresponding user's mobile terminal is notified of the result, and if they are not in the same location, the result is not notified.

Additionally, the deep learning model is implemented as follows.

a) First, three-axis magnetometer data of the mobile terminals 110-1 and 110-2 of the user are collected, and the horizontal strength (H) and the total strength (F) of the magnetic field are calculated according to [Equation 1] below. It is calculated using the Pythagorean theorem.

At this time, for reference, in order to calculate the total intensity (F), the horizontal intensity (H) must first be found. Finding F and H can be done using the Pythagorean theorem shown in [Equation 1].

[Equation 1]

Here, mag ² _I,j = (mag_x _I,j , mag_y _I,j ), mag_f _I,j = (H _I,j , mag_z _I,j ). i represents i ^(th) device and j represents j ^(th) data point. This step is important because the three-axis magnetometer data cannot be used directly for the same position detection.

Additionally, the magnetometer data of this user and the magnetometer data of other users are made as follows.

First, the magnetic field data of the infected user is configured as follows.

이다.in other words,

to be.

는 특정 시간 간격 내의 m 윈도우 이동 중에 상기 이동시간의 윈도우 행렬을 나타내고,

는 타임스탬프 t 동안의 3축 자기계이다.here

Represents a window matrix of the movement time during m window movement within a specific time interval,

is the three-axis magnetometer during timestamp t.

In addition, the magnetic field data of other nearby users is configured as follows.

이다.in other words,

to be.

는 확인할 데이터가 있는 사용자를 나타내고,

는 감염된 사용자의 데이터와 동일한 형식을 나타낸다.here

represents a user with data to check,

indicates the same format as the infected user's data.

b) Then, downsampling is performed by averaging the timestamp values of the total strength of the magnetometer thus calculated.

For example, specifically averaging the values of the timestamps down the magnetometer total strength to 250 ms bins. The purpose of this step is to use the same frequency for all data and to reduce noise.

c) Next, the downsampled total strength of the magneto field is normalized to be within the range of 0 and 1 defined according to [Equation 2] below.

That is, in order to increase the model accuracy, the total intensity (F) is normalized so that all values are within the range of 0 and 1 formally defined as [Equation 2].

[Equation 2]

where F' is the normalized result.

d) Therefore, training data is generated by converting the normalized total strength of the magnetometer into mㅧn matrix training data in the window sliding format below.

That is, before training the autoencoder model, we must first generate the mxn matrix training data using a window sliding technique. This window sliding technique will be described in more detail with reference to FIG. 5 below.

As described above, one embodiment relates to a co-location detection system using mobile magnetometer data, in which case co-location detection is performed on magnetometer readings from a cloud computing and deep learning based approach.

Thus, this co-location detection system thus provides one of the most decisive ways to prevent or mitigate epidemics and epidemic outbreaks. And, at this time, magnetometers are one of the particularly useful solutions because they can achieve high accuracy similar to GPS as well as consume significantly less battery.

5 is a diagram for explaining a window sliding operation applied to a disease infection risk notification method using cloud computing and deep learning-based co-location detection according to an embodiment.

As shown in FIG. 5 , the window sliding operation according to an embodiment first generates mxn matrix training data before training the magnetometer data reconstructed through the deep learning model as an autoencoder model.

For reference, here is the window sliding technique used to generate training data (720x480).

In this case, the execution step (format) of the window sliding technique is performed as follows.

1) First, data for a preset initial time is copied to the first row of the matrix.

Specifically, here we copy the first two minutes of data into the first row of the matrix (480 data points).

2) Then, the copied data is divided into one data point.

3) Next, copying to the second row of the matrix during the preset movement time and repeating the process until the data ends with the window of the preset standard size.

That is, copy the next 2-minute data (after shifting by one data point) into the second row of the matrix, repeating the process until the 5-minute window data is over.

4) Thus, matrix training data is generated by repeating the repetition operation a preset number of times.

That is, by repeating steps 1) to 3) above 10 times, a 7210x480 matrix for training the autoencoder model is created.

With this approach, increasing the amount of training data has the advantage of improving model accuracy.

On the other hand, in this case, an autoencoder model according to an embodiment is configured as follows, for example.

Specifically, after generating training data using the magnetometer data of an infected person, this data is input to an autoencoder model and trained. At this time, MSE is used as the objective function and is defined as [Equation 3].

[Equation 3]

은 입력,

은 재구성된 데이터이다. MSE를 목적 함수로 사용하기로 결정한 이유는 MSE가 실제로 예상치의 분산과 아래의 [식 4]로 정의되는 예상치의 편향 제곱의 합이기 때문이다.where n represents the total number of training data (rows of the training matrix),

is the input,

is the reconstructed data. The reason we decided to use the MSE as the objective function is that the MSE is actually the sum of the variance of the expectations and the squared biases of the expectations, defined as [Equation 4] below.

[Equation 4]

Here, by minimizing MSE, the proposed model aims to prevent overfitting and underfitting in the process of regression work by minimizing both the bias and variance of the estimates. Because autoencoders reconstruct input data by minimizing the difference between input and output, it can be classified as a regression task, so MSE can be an ideal objective function for autoencoders.

So, when the autoencoder model training is completed, data of other users is provided.

를 테스트 데이터로 모델에 공급하고 테스트 MSE를 교육 MSE와 비교한다. 테스트 MSE가 교육 MSE와 유사할 경우, 동일 위치 검출기는 다른 사용자가 동일한 위치에서 감염되었을 수 있는 지표를 서버에 반환한다.That is, data from other users at the end of the training process.

as test data to the model and compare the test MSE to the training MSE. If the test MSE is similar to the training MSE, the colocation detector returns an indicator to the server that other users may have been infected from the same location.

100: management information processing device 110: mobile terminal
101: I/F unit 102: RESTful AIP server
103: co-location detector 104: database

Claims (3)

A first step of generating and registering the following deep learning model by collecting 3-axis magnetometer data from the user mobile terminal whenever the user mobile terminal is connected to the Internet;
a second step of receiving a notification from a user mobile terminal when the user is infected;
a third step of learning and reconstructing magnetic field data of an infected user from the deep learning model when the notification is received;
a fourth step of checking whether or not the reconstructed user's magnetic field data and the magnetic field data of other nearby users are at the same location according to a preset same location detection format; and
As a result of the check, if the reconstructed user's magnetic field data and the magnetic field data of other nearby users are in the same location, the result is notified to the corresponding user's mobile terminal, and if they are not in the same location, the result is not notified Fifth step; including,

The deep learning model,
a) The three-axis magnetometer data of the user mobile terminal is collected, and the magnetic field horizontal intensity (H) and total intensity (F) are calculated using the Pythagorean theorem according to [Equation 1] below,
[Equation 1]

Here, mag ² _I,j = (mag_x _I,j , mag_y _I,j ), mag_f _I,j = (H _I,j , mag_z _I,j ). i represents i ^(th) device and j represents j ^(th) data point.
b) downsampling the calculated total intensity of the magnetometer by averaging the timestamp values;
c) normalize the downsampled total strength of the magneto field to be within the range of 0 and 1 defined according to [Equation 2] below,
[Equation 2]

Here, F' represents the normalized result.
d) generating training data by transforming the normalized total strength of the magnetometer into training data of m ㅧ n matrix in the following window sliding format;
The window sliding format,
1) copy the data for a preset initial time to the first row of the matrix;
2) Dividing the copied data into one data point,
3) Copy to the second row of the matrix during the next preset movement time and repeat the process until the end of the data with the window of the preset reference size;
4) generating matrix training data by repeating the repetition operation a preset number of times; Disease infection risk notification method using co-location detection based on cloud computing and deep learning, characterized by. The method of claim 1,
In the fourth step,
When it is confirmed whether the reconstructed user's magnetometer data and the magnetic field data of other users in the vicinity are in the same position, the following autoencoder model educates and confirms whether or not they are in the same position,
The autoencoder model,
a) Define using MSE as the objective function according to [Equation 3] below,
[Equation 3]

where n represents the total number of training data (rows of the training matrix),

is the input,

represents the reconstructed data.
b) the MSE becomes an objective function made according to [Equation 4] below; Disease infection risk notification method using co-location detection based on cloud computing and deep learning, characterized by.
[Equation 4]

The method of claim 2,
The magnetic field data of the infected user,
Consists of the following [Equation 5],
[Equation 5]

here

Represents a window matrix of the movement time during m window movement within a specific time interval,

is the triaxial magnetometer during timestamp t.
The magnetic field data of other users in the vicinity,
Consisting of [Formula 6] below; Disease infection risk notification method using co-location detection based on cloud computing and deep learning, characterized by.
[Equation 6]

here

represents a user with data to check,

indicates the same format as the infected user's data

Images