KR102116824B1 - Positioning system based on deep learnin and construction method thereof - Google Patents

Abstract

A data collection tool for collecting wireless signal information; A training server for learning the positioning algorithm of the artificial neural network structure and verifying the learned results using the wireless signal information collected by the data collection tool; And a positioning server equipped with a positioning algorithm of an artificial neural network structure learned from the training server, and applying a positioning algorithm of an artificial neural network structure to the inputted wireless signal information to estimate a location where the inputted wireless signal information is obtained. A deep learning based wireless positioning system is disclosed.

Description

Deep learning based positioning system and its construction method {POSITIONING SYSTEM BASED ON DEEP LEARNIN AND CONSTRUCTION METHOD THEREOF}

The present invention relates to a deep learning-based positioning system and a method for constructing the same, and more specifically, deep learning capable of accurately measuring in real time the position of a terminal communicating with an access point in an indoor space where an access point for short-range communication is installed. It relates to a base positioning system and its construction method.

The indoor positioning technique is a technique of locating a user's location inside a building using radio waves, light sources, and the like.

The wireless positioning using radio waves measures the user's location by measuring and processing signals transmitted by devices such as a user's mobile terminal and a base station of a mobile communication network, a Wi-Fi access point, and a Bluetooth access point. At this time, as a method for determining a user's location, a method based on an arrival time of a wireless signal, a measurement method based on an arrival angle using an array antenna, and a method based on signal strength are known.

Typical time-based positioning methods include a Time Of Arrival (TOA) method and a Time Difference Of Arrival (TDOA) method. The TOA method uses the arrival time of the wireless signal and the TDOA uses the difference between the arrival times of the wireless signal. .

In the positioning method based on the arrival angle, two or more base stations measure the direction of a signal coming from the terminal to obtain a direction angle, and use this to measure the position of the terminal. At least two direction angles are required to obtain the location of the terminal.

Cell ID, which is one of the other indoor positioning methods, is a technique for estimating the location of a terminal using information of service coverage. The location of the terminal is located in the cell's service area. The Enhanced Cell ID method, which is an improved Cell ID method, is a method of improving accuracy by adding distance information between a base station and a terminal to the Cell ID method.

Another indoor positioning method, the RF pattern matching method, uses the RF pattern received from the signal source to the receiver's antenna. This RF pattern is a method of estimating the location of the terminal by mapping the location corresponding to the most similar reference RF pattern compared to the reference RF pattern database for the previously established location.

Fingerprint method divides indoor space into virtual grids, measures the strength of each Wi-Fi signal for each grid, sets it into a database in the form of a fingerprint, and then compares the strength of the Wi-Fi signal measured by the database to the location. It is a method that improves the accuracy more than the conventional method as a method of making it available.

As described above, various technical methods for indoor positioning have been developed and applied, but they still do not provide satisfactory performance due to the disadvantages of each technique.

For example, it is true that the time reference method such as TOA or TDOA is difficult to accurately measure the time in a typical wireless device such as Wi-Fi, and is difficult to apply to actual commercialization due to a large error in the measured time.

Of course, it is possible to measure more accurate time compared to Wi-Fi using UWB or RFID wireless signals, but a separate access point and a dedicated receiver (UWB or RFID cannot be used for general devices such as smartphones) are required for commercialization. What is not suitable is reality.

In addition, the conventional fingerprint method has a large number of access points to be collected, and it is difficult to distinguish it from an adjacent grid due to the fluctuation of signal strength, and it is very difficult to implement this with an algorithm. There is a problem in that the RF database must be periodically regenerated due to the change.

Accordingly, the present invention is to provide a deep learning algorithm, in particular, a deep learning based positioning system capable of more accurately measuring a position of a terminal to be measured by applying a deep learning algorithm using a RNN (Recurrent Neural Network) structure, and a method for building the same This is a technical task to be solved.

The present invention as a means for solving the above technical problem,

A data collection tool for collecting wireless signal information;

A training server for learning the positioning algorithm of the artificial neural network structure and verifying the learned results using the wireless signal information collected by the data collection tool; And

A positioning server equipped with a positioning algorithm of an artificial neural network structure learned from the training server, and applying a positioning algorithm of an artificial neural network structure to the inputted wireless signal information to estimate a location where the inputted wireless signal information is obtained;

It provides a deep learning-based wireless positioning system comprising a.

In one embodiment of the present invention, the data collection tool divides a space targeted for positioning into a plurality of unit zones, assigns an identifier to each unit zone, and collects a plurality of radio signal information for each unit zone. have.

In one embodiment of the present invention, the radio signal information, the radio signal information, may include an identifier of an access point capable of receiving a radio signal in each unit zone, and radio signal strength with the access point.

In one embodiment of the present invention, the data collection tool combines the radio signal information with the identifier of the unit zone, and the identifier of the unit zone and the radio signal information associated with it are used to train the training server. It can be divided into data and test data used for verification.

In one embodiment of the present invention, the data collection tool maps a vector value X using the value of the radio signal strength for each access point as an element and a vector value Y corresponding to the identifier of the unit zone, and the vector value X And vector values Y can be mapped one-to-one.

In one embodiment of the present invention, the data collection tool includes a map of an area for collecting the wireless signal information, accumulation information of the map, the unit area displayed on the map, and a wireless signal for each unit area An identification mark indicating whether to collect information and a marker indicating a location to collect the wireless signal information may be displayed.

In one embodiment of the present invention, the data collection tool may delete radio signal information for a mobile access point from among the collected radio signal information.

In one embodiment of the present invention, the training server includes a positioning algorithm of a neural network structure for deep learning, and the weight and bias values of the artificial neural network structure through learning of wireless signal information collected by the data collection tool Can be optimized.

In one embodiment of the present invention, the neural network structure is a Recurrent Neural Network (RNN), and the RNN may be a Long Short Term Memory (LSTM) network.

In one embodiment of the present invention, the training server may set a learning rate for finding a point at which the cost function value becomes minimum.

In an embodiment of the present invention, the training server may be set with a training count indicating a learning amount.

In one embodiment of the present invention, the training server has an input data scale corresponding to the number of access points present in the location target area, has an output data scale of 1, and is a hidden that is equal to or randomly determined as the input data scale. You can have a data scale.

In one embodiment of the present invention, the training server may be set with a sequence length indicating how much radio signal data collected in time to apply to the positioning.

In one embodiment of the present invention, the data collection tool maps a vector value X using the value of the radio signal strength for each access point as an element and a vector value Y corresponding to the identifier of the unit zone, and the vector value X And vector values Y can be mapped one-to-one.

In one embodiment of the present invention, the training server pre-processes the data provided by the data collection tool in a form suitable for application to a neural network structure, and inputs the pre-processed training data into a neural network structure for deep learning, a positioning algorithm. And learning the location algorithm of the learned neural network structure, and inputting the pre-processed test data into the location algorithm of the learned neural network structure to verify the location algorithm of the learned neural network structure.

In one embodiment of the present invention, a measurement target terminal that can be positioned and collect wireless signal information from the access point may be further included.

In one embodiment of the present invention, the measured terminal provides the collected wireless signal information to the positioning server, and the positioning server measures the measured target by applying a neural network structure positioning algorithm to the wireless signal information provided from the measured terminal. The location of the terminal can be derived.

In one embodiment of the present invention, the positioning server estimates the position of the terminal to be measured by pre-processing the wireless signal information provided from the terminal to be measured in a form suitable for application to the neural network structure and then inputting it to the positioning algorithm of the neural network structure. And, the location closest to the location information used for learning of the training server may be provided to the terminal to be measured as the final location information.

In one embodiment of the present invention, the terminal to be measured is equipped with a positioning algorithm of an artificial neural network structure learned from the training server, and can estimate its own position using the collected wireless signal information.

The present invention as another means for solving the above technical problem,

A data collection step of collecting a plurality of radio signal information for each of a plurality of unit zones for dividing a space subject to positioning; And

A training step of learning a positioning algorithm of an artificial neural network structure and verifying the learned results using the wireless signal information collected in the data collection step;

It provides a method for constructing a deep learning-based wireless positioning system including a.

In one embodiment of the present invention, the radio signal information may include radio signal strength for each identifier of an access point capable of receiving radio signal information in each unit zone.

In one embodiment of the present invention, the data collection step includes combining the radio signal information with the identifier of the unit zone, and training the training server using the identifier of the unit zone and the radio signal information associated with it. It can be divided into data and test data used for verification.

In one embodiment of the present invention, the data collection step maps a vector value X using the value of the radio signal strength for each access point as an element and a vector value Y corresponding to the identifier of the unit zone, and the vector value X And vector values Y can be mapped one-to-one.

In one embodiment of the present invention, the training step, by providing a training server including a positioning algorithm of a neural network structure for deep learning, by learning by inputting the wireless signal information collected in the data collection step to the training server The weight and bias values of the artificial neural network structure can be optimized.

In one embodiment of the present invention, the training step pre-processes the data generated in the data collection step in a form suitable for application to a neural network structure, and inputs the pre-processed training data into a neural network structure for deep learning, a positioning algorithm. And learning the location algorithm of the learned neural network structure, and verifying the location algorithm of the learned neural network structure.

According to the deep learning-based positioning system and its construction method, since a positioning algorithm based on deep learning is applied in measuring a location using wireless signal information, the similarity of data and patterns stored in the database is compared with a classical analysis algorithm. Compared to the fingerprint method, more accurate positioning is possible.

In particular, according to the deep learning-based positioning system and its construction method, by implementing a positioning algorithm through deep learning of an RNN structure, data of a previous time affects the current positioning result, thereby excluding discontinuous error features to make it more accurate. Positioning is possible.

1 is a block diagram of a deep learning-based positioning system according to an embodiment of the present invention.
2 is a flowchart illustrating a deep learning-based positioning system construction method according to an embodiment of the present invention.
3 shows an example of a data collection tool display screen of a deep learning based positioning system according to an embodiment of the present invention.
4 and 5 are diagrams illustrating a data structure generated by signal information transformation performed by a data collection tool of a deep learning based positioning system according to an embodiment of the present invention.
6 is a diagram illustrating a neural network structure for implementing a positioning algorithm applied to a deep learning-based positioning system according to an embodiment of the present invention.
7 is a diagram illustrating an example of an input data structure for learning and verifying a positioning algorithm of a deep learning based positioning system according to an embodiment of the present invention.
8 is a diagram illustrating a data processing process performed in a training server of a deep learning-based positioning system according to an embodiment of the present invention.
9 is a diagram illustrating a data processing process performed in a positioning server of a deep learning based positioning system according to an embodiment of the present invention.

Hereinafter, a deep learning based positioning system and a method of constructing it according to various embodiments of the present invention will be described in more detail with reference to the accompanying drawings.

The deep learning-based positioning system according to an embodiment of the present invention includes a data collection tool 10 for collecting wireless signal information required to generate a deep learning-based positioning algorithm, and an artificial neural network structure using the collected wireless signal information. Provides a training server 20 for training the positioning algorithm, a positioning server 30 equipped with a positioning algorithm of an artificial neural network structure learned from the training server 20, and wireless signal information collected by the positioning server 30, It may be configured to include a measurement terminal 40 that receives the location information measured by the positioning server 30 as input to the provided wireless signal information.

In addition, the deep learning-based positioning system according to an embodiment of the present invention may include an application service server 50 that receives location information measured by the location server 30 and provides a location-based service corresponding thereto.

The data collection tool 10 assigns an identifier (unique number) to a unit area in which a space (particularly an indoor space), which is a target of positioning, is divided into a certain size of a grid or a certain section on a movement path, and wireless A plurality of signal information is collected, and the collected information is converted into a format required for learning performed by the training server 20 and provided to the training server 20.

The training server 20 designs and applies an artificial neural network structure applied to a deep learning-based positioning algorithm, and uses the information provided by the data collection tool to find values optimized for the artificial neural network structure to complete the positioning algorithm.

The positioning server 30 is a server that is equipped with a positioning algorithm having an optimized artificial neural network structure learned and verified by the training server 20 and provides a function for estimating the position of the terminal 40 to be measured. The location of the terminal is estimated by using the wireless signal information collected by the terminal to be measured 40 at the current location.

The positioning server 30 sends the estimated location information of the terminal 40 to the terminal 40 to be measured or the application service server 50 (a general user terminal equipped with an application service). Can transmit.

The application service server 50 may provide various location-based application services such as navigation, current location guidance, zone entry / exit services, location-based advertisement services, and the like.

The measured terminal 40 collects wireless signal information at the current location and transmits it to the positioning server 30, and the positioning server 30 applies location information by applying a positioning algorithm applying an artificial neural network structure to the received wireless signal information. Can be estimated. In addition, in an embodiment of the present invention, an artificial neural network positioning algorithm may be directly mounted on a terminal to be measured according to the use of an application service.

2 is a flowchart illustrating a deep learning-based positioning system construction method according to an embodiment of the present invention.

Referring to FIG. 2, a method for constructing a deep learning-based positioning system according to an embodiment of the present invention is performed by a data collection step (S11) performed by the data collection tool 10 and a training server 20 It may be configured to include a learning and verification step (S12) and a step (S13) of building a positioning server (S30). Each step of the method for constructing a deep learning based positioning system according to one embodiment of the present invention will be described in more detail below. Through the description of each step, the action of each component of the deep learning-based positioning system according to an embodiment of the present invention shown in FIG. 1 may be described in more detail.

Data collection step (S11)

The data collection step S11 may be performed by the data collection tool 10.

The data collection tool 10 is a device for collecting wireless signal information in the entire area targeted for positioning, and may be implemented as a mobile communication device such as a smart phone or a laptop computer.

In one embodiment of the present invention, the data collection tool 10 may divide the entire area subject to positioning into a plurality of unit areas (sections set on a grid or a movement path) and collect wireless signal information for each unit area. . The wireless signal information for positioning may be wireless signal strength information of an identifier (eg, Mac address) of a Wi-Fi access point capable of communicating in a corresponding unit zone and each access point.

The information collected by the data collection tool can be divided into training data and test data for learning a deep learning positioning algorithm by combining with collection location information (that is, information for identifying a unit area).

3 shows an example of a data collection tool display screen of a deep learning based positioning system according to an embodiment of the present invention.

The data collection tool 10 may display a screen 11 as shown in FIG. 3 by a pre-installed program and operate as follows.

When the program for data collection is executed in the data collection tool 10, the map 111 of the collection target area (indoor space, etc.) can be loaded. Accumulation information 112 may be displayed at the bottom of the map 111 so as to infer an actual distance on the map.

In addition, on the map 111 displayed on the data collection tool 10, unit areas 113 and 113 'for collecting Wi-Fi signal information may be displayed, and unit areas 113 for collecting Wi-Fi signal information later. The unit area 113 ′ in which the Wi-Fi signal information collection has already been performed may be divided by a method such as color, contrast, or number display.

In addition, a marker 114 capable of marking the location of the unit zone for which data is to be collected may be provided. The marker 114 means a location where Wi-Fi signal information should be collected, and the user may move to a location where the marker is displayed to collect signal information. When the collection is completed, the color or contrast of the unit area can be changed or a number can be assigned to the unit area to indicate that the data collection is completed.

The method of displaying the location of the unit area that needs to be collected may be displayed directly on the map or the current location may be automatically displayed on the map in an environment where GPS or other location is known.

The data collection tool 10 may collect a plurality of signal strengths for a plurality of access points for each unit area, and the number of collections may be set by a user.

The Wi-Fi signal information collected by the data collection tool 10 includes the MAC address of the Wi-Fi access point, the signal strength for each access point, the identifier (identification number) of the unit zone assigned by the user, and the location coordinate (if secured). It can contain.

Meanwhile, the data collection tool 10 not only collects the Wi-Fi signal information as described above, but also assigns the same identification number to the signals included in the unit area, and inputs the input data (X) and output data (Y). Can be converted to This conversion is to convert the collected Wi-Fi signal information into deep learning positioning algorithm input information for use in the training server 20.

Hereinafter, the data conversion performed by the data collection tool 10 will be described in detail.

The data collection tool 10 maps the collected Wi-Fi signal information to a vector value X having a signal strength value for each access point in a unit zone as an element and Y representing a unit zone. X is a value corresponding to the signal strength collected for each access point (each access point is divided into unique values such as MAC addresses) (values obtained by converting measured signal strength values using a constant function) and Y is for each unit area This is the number assigned to the collection location assigned to the user and is the information assigned by the user to be mapped one-to-one with location information.

Here, X and Y are mapped one-to-one, and Y can be finally mapped one-to-one with actual location coordinates. That is, a position coordinate conversion DB corresponding to Y needs to be built. In addition, the collected Wi-Fi signal information for each unit zone can be converted into X and Y vector values.

4 and 5 are diagrams illustrating a data structure generated by signal information transformation performed by a data collection tool of a deep learning based positioning system according to an embodiment of the present invention.

As shown in FIG. 4, in the data collection tool 10, one column becomes a value of Y, which is the number of a unit region in which data is collected, and each row corresponding to the Y value is signal strength for each access point corresponding to the Y value. It can be the value of X that lists the values.

Here, the X value is classified for each Mac address of the access point, and the larger the size, the greater the signal strength. In addition, in the example of FIG. 4, the signal strength can be converted to a range of 0 to 100, and the signal strength of 0 indicates that there is no signal. The change in signal strength may be set differently as needed.

The data collection tool 10 may generate training data for training and test data for verification as shown in FIG. 5 using the data structure shown in FIG. 4.

As shown in FIG. 5, the data collection tool 10 may divide some of the plurality of data collected from the same location number (ie, the same unit zone) into training data and test data. At this time, the same data cannot be used as training data and test data.

On the other hand, the signal strength value for each access point, which becomes the X value, can be converted into a value suitable for calculation when the collected signal strength information is converted into a deep learning positioning algorithm input data structure as shown in FIG. 4 or 5. have. For example, if the actually collected signal strength information is greater than -100 and is less than 0, the actual collected value may be converted into a value within the range of 0 to 100 so that there is no distortion of the signal information.

For optimal positioning performance, the process of collecting Wi-Fi signal information excludes information collected about a mobile access point that may be a factor that increases deep learning errors, and collects data multiple times at different times. It is preferable to use the information collected using the data collection tool 10.

Learning and verification phase (S12)

The learning and verification step (S12) may be performed by the training server 20.

The training server 20 is equipped with a deep learning positioning algorithm, and optimizes and tests the weight (W) and bias (b) values of the hidden layer by using X and Y values of the training data through learning. It is a server that provides a function to verify whether the result of the positioning is predicted by entering data. The training data and test data are data structured with the X and Y values of the signal information collected by the data collection tool 10 as described above.

6 is a diagram illustrating a neural network structure for implementing a positioning algorithm applied to a deep learning-based positioning system according to an embodiment of the present invention. Particularly, FIG. 6 shows an algorithm structure using a Recurrent Neural Network (RNN) among algorithms of various neural network structures applied to deep learning.

As shown in FIG. 6, the output value of the RNN structure for positioning to which the deep learning algorithm according to an embodiment of the present invention is applied is not only the current input value, but also temporally, unlike the convolutional neural network or deep neural network (DNN). It is also characterized by being influenced by the result value of the previous data.

Due to these characteristics, when RNN is applied, the state value at the time t-1 at the previous time also affects the state at the current time t, thereby providing a role to prevent a sudden position error to maintain a more stable positioning result. . On the other hand, since CNN or DNN depends on the input data at time t, the positioning error is more likely to increase discontinuously with the result value at time t-1.

On the other hand, in deep learning networks including RNN, the deeper the layer, the lower the propagation of the learned data. The phenomenon that the learning data characteristic is lost may be a factor that hinders deep learning performance.

To prevent this problem, a Long Short Term Memory (LSTM) network, which is a type of RNN, was developed. The LSTM network is a unit that maintains data without losing data even when the time is prolonged and the layer is deep. In one embodiment of the present invention, the LSTM network is applied to construct a deep learning network and learn to implement a positioning algorithm using data collected by the data collection tool 10.

The learning process to implement the RNN positioning algorithm for optimal performance is a process of finding all weight (W) and bias (b) values between the input radar, hidden layer, and output layer. Learning progress while tuning the values.

-Learning rate-The learning rate can be set externally to find the point where the value of the deep learning algorithm's cost function (also called the cost function or loss function) is the minimum (the point where the derivative value is 0). It is a unit value. If the running rate value is too large, the cost function will diverge without convergence. If it is too small, it will take too much time to converge, so you need to find an appropriate value according to the characteristics of the training data.

-Training Count (Training Count or Learning Count)-The training count is a value indicating the amount of learning and is a value that must be set externally. In general, the higher the training count value, the better the learning (the cost function converges to 0), but when it exceeds a certain value, the learning result does not improve anymore and takes only a lot of time.

-Input data scale (Data dimension)-The data scale is the size of input data input at a specific time t, and the size of the X vector of training data used for RNN input, that is, the target region (number of access points collected from all grid cells) ) Is a value corresponding to the number of all access points used in.

-Hidden data dimension-The hidden data scale is the scale of the data value input to the hidden layer and is generally set to an appropriate value equal to or smaller than the input data scale.

-Output data dimension-The output scale is the scale of the final result value of the RNN algorithm, which is one unit zone determined by the positioning algorithm, so the output scale can be set to 1. The output value at output time t is used, and the output value at the rest of the time is not used.

-Sequence Length-The sequence length represents the size of the RNN networks connected in parallel on the time axis. The sequence length is related to how many data of the previous time will be used, and it is possible to select an appropriate value according to an environment such as a positioning cycle to have optimal positioning performance.

7 is a diagram illustrating an example of an input data structure for learning and verifying a positioning algorithm of a deep learning based positioning system according to an embodiment of the present invention.

The example shown in FIG. 7 is an example in which a location target area is divided into four unit zones (lattice or cell) and a total of six access points are recognized in the location target area. The data collected by the data collection tool as described above is largely divided into training data and test data, and training data used for learning is an X vector (Train X, which is an input value) and a wireless signal strength with each access point. It is divided into Y vector (Train Y, unit area) that is an output value, and test data used for verification is also an X vector (Test X, wireless signal strength with each access point) and Y vector (Test Y) that are output values. , Unit area).

The values of the X vector shown in FIG. 7 may be converted from raw data collected by the sensor of the terminal to be measured into a form that can be used for a positioning algorithm.

8 is a diagram illustrating a data processing process performed in a training server of a deep learning-based positioning system according to an embodiment of the present invention.

As shown in FIG. 8, the training server 20 sequentially performs input data pre-processing step, training process step, and test process step to optimize the RNN positioning algorithm.

In the input data processing step, the training server 20 input data (Training X, Training Y, Test X, Test) to converge the cost function (to have a maximum value of 0) in order to facilitate learning of the RNN positioning algorithm. The element value of Y) is pre-processed by scaling the value between 0 and 1 (MinMaxScaling).

Subsequently, the training server 20 performs a training processing step of inputting the scaled value into an RNN network having a sequence length N to train. As described above, in the RNN network of the present invention, an LSTM unit is used as a cell component of a neural network to prevent a problem of back propagation in which characteristics of learned data are lost.

In the training processing step, the user may set the sequence length, input data scale, hidden data scale, training count, hidden layer size, and the like.

Subsequently, the training server 20 inputs Min-Max scaled test data X into the trained RNN network to derive the positioning result Y ', and how much the derived positioning result matches the scaled output result Y to RMSE (Root). Mean Square Error) can be calculated and verified.

As described above, when learning and verification are completed, an algorithm of RNN network structure for positioning is completed.

Positioning server construction step (S13)

The positioning server 30 may receive and mount the positioning algorithm of the RNN network structure that has been learned and verified by the training server 20 from the training server 20. The positioning server 30 may perform positioning through a process as shown in FIG. 9.

9 is a diagram illustrating a data processing process performed in a positioning server of a deep learning based positioning system according to an embodiment of the present invention.

Specifically, when the measurement terminal 40 collects Wi-Fi signal information and transmits it to the positioning server 30, the positioning server 30 inputs through a pre-processing process (scaling) before applying the positioning algorithm of the RNN network structure. Generate input data Xt.

In the pre-processing of the input data (Xt), scaling (MinMax Scaling) converts the data to a scale applicable to the positioning algorithm, and the scaled input data is input to the RNN network.

The positioning algorithm of the RNN network structure mounted on the positioning server 30 processes the received data and outputs the estimated position Yt 'which is the output value. At this time, the estimated position (Yt ') may be a value that does not exactly match the Yt value used in the actual training by the training server 20, is replaced with the closest unit area of the Yt value used in the actual training The location coordinate of the corresponding unit zone may be the final location information.

Meanwhile, as described above, in addition to the method in which the positioning server 30 receives Wi-Fi wireless signal information (the size of the access point's MAC address and the communication strength therewith) that is an input value from the terminal 40, the terminal to be measured The positioning service may be provided by directly receiving and mounting the positioning algorithm from the training server 20. In this case, the positioning service can be implemented in the same manner as the positioning technique in the above-described positioning server.

In the above description, a deep learning-based positioning system or a construction method using wireless signal information with an access point is mainly described in Wi-Fi communication, but all communication systems (for example, Bluetooth) have elements similar to the access point. , RFID, UWB) in the same way.

In addition, the positioning algorithm used in the present invention can be equally enlarged and applied indoors as well as indoors.

In addition, in order to learn the RNN positioning algorithm used in the present invention, a unique number (Y value) is assigned to classify the collected data by dividing it into regular grids or regular intervals, and a plurality of wireless signals are collected for each grid. The method of dividing the data into a certain ratio to perform learning and verification was suggested, but instead of using the division number assigned to the grid cell as the Y value, the location information (latitude, longitude, or latitude longitude) of the collected data can be calculated directly. Information) can be applied to construct by training and verifying using Y value.

 Also, in the above description, although an RNN is mainly applied as an artificial neural network structure used for deep learning, a positioning system may be constructed by applying CNN or DNN instead of RNN within the scope of the present invention.

On the other hand, when a deep learning-based positioning system and a method for constructing it according to various embodiments of the present invention are applied to a very wide range, the wide range is divided into a plurality of deep learning learning areas, and the above-described ones are described for each area. The process of establishing a deep learning-based positioning system according to various embodiments of the present invention may be performed, and other information (eg, MSC information, base station information, access point information) that can be collected by the terminal to be measured by the area selection method ) Can be constructed by first selecting the area corresponding to the estimated position calculated and applying the deep learning positioning algorithm learned in the area.

In addition, if the environment for collecting data among the unit areas corresponding to the lattice or the moving path segmented target area is not created and the wireless signal information cannot be collected, the area uses a deep learning-based positioning algorithm. Learning to generate is not done. Terminals belonging to the non-learning area should be pre-filtered so that deep learning-based positioning algorithms according to various embodiments of the present invention are not performed, and positioning algorithms of other methods can be applied.

As a method of determining a terminal in a non-learning area, which can be applied to various embodiments of the present invention, when the input data X of a deep learning-based positioning algorithm is generated using signal information collected by the terminal, the non-learning area includes Since there is no learned access point signal information, all signal information values of X are very likely to be 0, and thus can be performed by filtering a terminal having such signal information. In other words, when comparing the identification information (Mac address) of the access point collected by the terminal to be measured and the identification information of the access point used in the positioning algorithm, the same identification information may not be present or there may be only a very small percentage. The terminal can be filtered.

In the above, although shown and described in relation to a specific embodiment of the present invention, it is understood that the present invention can be variously improved and changed within the limits that do not depart from the technical spirit of the present invention provided by the following claims. It will be apparent to those skilled in the art.

10: data collection tool 20: training server
30: positioning server 40: terminal to be measured
50: application service server

Claims (26)

A data collection tool for collecting wireless signal information;
A training server for learning a positioning algorithm of an artificial neural network structure and verifying the learned result using the wireless signal information collected by the data collection tool; And
The positioning server is equipped with a positioning algorithm of an artificial neural network structure learned from the training server, and a positioning server for estimating a location where the input wireless signal information is obtained by applying a positioning algorithm of the artificial neural network structure to the input wireless signal information,
The training server includes a positioning algorithm of a neural network structure for deep learning, and optimizes the weight and bias values of the artificial neural network structure through learning of wireless signal information collected by the data collection tool. Running-based wireless positioning system. The method according to claim 1,
The data collection tool divides a space targeted for positioning into a plurality of unit zones, assigns an identifier to each unit zone, and collects a plurality of radio signal information for each unit zone. system. The method according to claim 1 or claim 2,
The radio signal information, a deep learning-based radio positioning system, characterized in that it comprises an identifier of an access point capable of receiving a radio signal in each unit zone and the radio signal strength with the access point. The method according to claim 3,
The data collection tool combines the radio signal information with the identifier of the unit zone, and the training data used for training of the training server and the test data used for verification of the identifier of the unit zone and the radio signal information coupled thereto. Deep learning based wireless positioning system, characterized by classifying. The method according to claim 4,
The data collection tool maps the vector value X having the value of the radio signal strength for each access point as an element and the vector value Y corresponding to the identifier of the unit region, and the vector value X and the vector value Y are mapped one-to-one. Deep learning-based wireless positioning system, characterized in that characterized in that. The method according to claim 2,
The data collection tool, an identification mark indicating a map of an area for collecting the wireless signal information, the accumulated information of the map, the unit area displayed on the map, and whether to collect wireless signal information for each unit area And a marker indicating a location to collect the wireless signal information. The method according to claim 1,
The data collection tool is a deep learning-based wireless positioning system, characterized in that the wireless signal information for the mobile access point among the collected wireless signal information. delete The method according to claim 1,
The neural network structure is a deep learning-based wireless positioning system, characterized in that the RNN (Recurrent Neural Network). The method according to claim 9,
The RNN is a deep learning based wireless positioning system, characterized in that the LSTM (Long Short Term Memory) network. The method according to claim 1,
The training server, a deep learning-based wireless positioning system, characterized in that the learning rate (learning rate) is set to find the point where the cost function value is the minimum. The method according to claim 1,
The training server, a deep learning-based wireless positioning system, characterized in that a training count (Training Count) indicating the learning amount is set. The method according to claim 1,
The training server has an input data scale corresponding to the number of access points present in a location target region, has an output data scale of 1, and has a hidden data scale equal to or randomly determined from the input data scale. Deep learning based wireless positioning system. The method according to claim 10,
The training server, a deep learning-based radio positioning system characterized in that the sequence length is set to indicate how much to apply the radio signal data previously collected in time. The method according to claim 3,
The data collection tool maps a vector value X having the value of the radio signal strength for each access point as an element and a vector value Y corresponding to the identifier of the unit region, and the vector value X and the vector value Y are mapped one-to-one. Deep learning-based wireless positioning system, characterized in that characterized in that. The method according to claim 5,
The training server pre-processes the data provided by the data collection tool in a form suitable for application to a neural network structure, inputs the pre-processed training data into a neural network structure for deep learning, learns a positioning algorithm, and learns the learned neural network structure Deep learning-based wireless positioning system, characterized in that by verifying the positioning algorithm of the learned neural network structure by inputting the test data into the positioning algorithm of. The method according to claim 3,
A deep learning-based wireless positioning system further comprising a terminal to be measured and capable of collecting wireless signal information from the access point. The method according to claim 17,
The measured terminal provides the collected collected wireless signal information to the positioning server, and the positioning server derives the position of the measured terminal by applying a neural network structure positioning algorithm to the wireless signal information provided by the measured terminal. Deep learning based wireless positioning system, characterized in that. The method according to claim 18,
The positioning server pre-processes the wireless signal information provided by the device to be measured in a form suitable for application to a neural network structure, inputs it to a positioning algorithm of the neural network structure, estimates the position of the device to be measured, and trains the training server. A deep learning based wireless positioning system, characterized in that a location closest to the used location information is provided to the terminal to be measured as final location information. The method according to claim 17,
The measurement terminal is equipped with a positioning algorithm of an artificial neural network structure learned from the training server, and a deep learning-based wireless positioning system characterized in that it estimates its own location using the collected wireless signal information. A data collection step of collecting a plurality of radio signal information for each of a plurality of unit zones for dividing a space subject to positioning; And
A training step of learning the positioning algorithm of the artificial neural network structure and verifying the learned results using the wireless signal information collected in the data collection step,
In the training step, a training server including a positioning algorithm of a neural network structure for deep learning is provided, and the wireless signal information collected in the data collection step is input to the training server to learn the weight and bias of the artificial neural network structure. Deep learning based wireless positioning system construction method characterized by optimizing the value. The method of claim 21,
The wireless signal information, deep learning based wireless positioning system construction method comprising the wireless signal strength for each identifier of an access point capable of receiving wireless signal information in each unit zone. The method according to claim 22,
In the data collection step, the wireless signal information is combined with the identifier of the unit zone, and the identifier of the unit zone and the wireless signal information coupled thereto are training data used for training of the training server and test data used for verification. Deep learning based wireless positioning system construction method characterized in that separated by. The method according to claim 23,
In the data collection step, a vector value X having the value of the radio signal strength for each access point as an element and a vector value Y corresponding to the identifier of the unit region are mapped, and the vector value X and the vector value Y are mapped one-to-one. Deep learning based wireless positioning system construction method, characterized in that characterized in that. delete The method according to claim 24,
In the training step, the data generated in the data collection step is pre-processed in a form suitable for application to a neural network structure, and the pre-processed training data is input to a neural network structure for deep learning to learn positioning algorithms and the learned neural network structure. A deep learning-based wireless positioning system construction method characterized by verifying a positioning algorithm of a learned neural network structure by inputting pre-processed test data into a positioning algorithm of.

Images