KR20200017612A - Method for positioning learning by using Deep learning - Google Patents

Abstract

The present invention relates to a location positioning learning method using deep learning. Specifically, provided are a method for estimating location positioning using a machine-learned artificial intelligence neural network model for signal strength of a radio-based positioning resource and an apparatus performing the same. According to the present invention, the location positioning learning method comprises the steps of: (a) designating a convolutional neural network (CNN) including a plurality of convolutional layers; and (b) learning using input data and a label.

Description

Method for positioning learning by using deep learning

The present invention relates to a location learning method using deep learning, and to a method for estimating location using an artificial neural network model machine learning on the signal strength of the radio wave-based positioning resources and apparatus for performing the same will be.

Location information estimation technology is a technology that can track and track the location of assets or people in real time, which can be usefully used in various fields. It is widely used in various fields, from generalized services such as car navigation technology using GPS satellite information, to indoor and outdoor positioning technology for moving objects using a special terminal such as active RFID technology.

As indoor positioning is developed, various members are being used. The most popular is WIFI-based WPS, which uses the existing infrastructure as it is, and commercial facilities or office spaces are sufficient APs for WIFI-based positioning (WPS). In the indoor positioning field, there is a need for a deep learning method that can establish a learning strategy and model necessary for positioning. However, the popular input and output method of deep learning is One Hot Encoding (OHE), which classifies data into orthogonal matrix tables according to types and processes them as input or output. This method can express only one attribute at a time, it can not express the relationship between them and takes up a lot of memory, and is currently used as an output expression method for classification problems. In addition, Deep Learning expresses data in different ways to process actual data as input for each domain. In the well-known image recognition field, each pixel value of image data RGB is used as a CNN (Convolutional Neural Network) input. Also, in the field of natural language processing, it is impossible to input directly like an image, so that the word bundle is made into a vector and used as input. Generally, in the case of natural language, the number of words is 100,000 or more. Therefore, in case of expressing OHE, input size and memory capacity are needed, and since the meaning of words cannot be expressed, the entire word set is usually limited to a certain number (eg 10,000), and this is expressed in OHE, and the proximity of words in sentences It is learned as a vector form and used as the input of deep learning (CNN, RNN, etc.).

However, general positioning resources in indoor positioning include the use of radio waves such as WIFI AP or BLE tag, and natural resources generated from building structures such as geomagnetism. In the case of geomagnetism, the horizontal component (Mh) and vertical component (Mv) of the X, Y, Z or geomagnetism of the sensor can be used as input for deep learning. However, in the case of WIFI and BLE, the data is represented by an identifier (ID) and a signal strength indicator (RSSI) such as a MAC address.In the case of a radio-based positioning resource, the identifier (ID) and the signal strength (RSSI) are dependent. Since it must be able to perform 1: 1 mapping because it has a relationship with each other, it is impossible to use techniques such as OHE and Word2Vec, which are mainly used in the deep learning described above. It has variable size every hour, and the range of identifier of WIFI AP or BLE tag (MAC address, etc.) is very large, and only a few of the APs appear in a building, which uses very large memory when using general OHE. As the environment changes, there is a problem in that the AP or BLE tag additionally deleted should be flexibly coped with.

1020180058191 A

The present invention was devised to solve such a problem, and location estimation can be performed using a neural network model machine-learned for signal strength by converting a radio-based positioning resource into a representation applicable to deep learning. Its purpose is to.

Positioning learning using deep learning according to the present invention to achieve the above object comprises the steps of: (a) designating a convolutional neural network (CNN) including a plurality of convolutional layers; And (b) training using input data and labels.

The input data is either a WIFI AP or a BLE tag which is a radio wave-based positioning resource.

The input data is signal strength and is input in one of One Hot Encoding (OHE) form, RSSI value form, and RSSI vector form including beer place.

According to the present invention, since an identifier, which is a radio wave-based positioning resource, and a signal strength are matched in a 1: 1 relationship, and converted into a representation applicable to deep learning, indoor positioning is effective.

1 is a location learning structure using deep learning according to the present invention, the first embodiment of the input location resource representation.
2 is a second embodiment of an input location resource representation according to the present invention;
3 is a third embodiment of an input location resource representation according to the present invention;
4 is a flow chart as a first embodiment of a positioning method using deep learning according to the present invention.
5 is a flowchart as a second embodiment of a positioning method using deep learning according to the present invention.
6 is a view showing the configuration of a positioning device using a deep learning according to the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms or words used in the present specification and claims should not be construed as being limited to the ordinary or dictionary meanings, and the inventors should properly explain the concept of terms in order to best explain their own inventions. It should be interpreted as meaning and concept corresponding to the technical idea of the present invention based on the principle that it can be defined. Therefore, the embodiments described in the specification and the drawings shown in the drawings are only the most preferred embodiment of the present invention and do not represent all of the technical idea of the present invention, various modifications that can be replaced at the time of the present application It should be understood that there may be equivalents and variations.

1 is a position location learning structure using deep learning according to the present invention, and is a view showing a first embodiment of an input location resource representation.

The CNN includes an input layer, a hidden layer, and an output layer. The input layer receives propagation based positioning resources. The hidden layer is composed of a plurality of layers, each layer including a convolutional layer and a pooling layer. The convolution layer performs a convolution operation on the propagation-based positioning resource input to each layer using a convolution filter and generates a feature. In addition, the size of the feature generated through pooling is reduced, and the convolutional layer and the pooling layer may be represented as feature extraction units. The output layer classifies classes by combining various features expressed by the feature extractor. In this case, the output layer may be configured as a fully connected layer and may be represented by a classifier. In general, the structure of a CNN (e.g., the number of hidden layers, the number and size of filters in each layer, etc.) is predetermined and the weight matrix of the filters (especially convolution filters) in each layer is already known. The answer to which class belongs is estimated by using the known data. Such known data are called 'learning data', and the process of determining the weighting matrix of the filter is called 'learning'.

The learning model using the CNN of the present invention is largely composed of a training phase and a prediction phase. The learning phase consists of data collection, data feature filtering (or feature), hypothesis, Cost function definition, and learning sequence. The prediction phase is a step of estimating the learning result using real data. In the present invention, the filter for processing the indoor positioning propagation environment can be defined and used in various ways, but in the present invention, the filter is defined by the following [intuition] and [strategy].

[intuition]

 (1) Similar APs / tags and RSSIs tend to be detected at similar locations.

  The RSSI of the AP / tag may vary due to the shape or object of space.

  The AP / tag detected may vary according to the scanning timing.

  The same AP / tag and RSSI can be detected at non-adjacent locations

 ② RSSI tends to be proportional to distance.

  RSSI is proportional to distance as it is closer, not proportional to distance.

  -It is often not proportional due to the shape of space and the attenuation of radio wave intensity by objects.

 ③ Larger RSSI tends to have higher reliability.

  -Larger RSSI tends to have less measurement deviation.

  -The smaller the RSSI, the larger the measurement deviation.

[strategy]

 ① The learning is carried out so that similar positions are estimated for similar APs / tags and RSSIs.

  -Learning to reflect the signal attenuation effect due to the shape or object of space.

  -Learning to reflect missing AP / tag.

  When the same AP / tag and RSSI exist in different locations, learning is performed based on probability.

 ② Measure between AP / Tag RSSI and location distance.

  -According to the space, the learner learns a section in which the RSSI for each AP / tag and the distance are proportional to each other.

 ③ Study the reliability of RSSI strength and deviation.

By the above [intuition] and [strategy], the most basic filter of the present invention defines a compound product to filter a feature that a plurality of APs appear at the same time. For example, for m virtual input nodes, when the number of scanned APs / tags is p and the corresponding APs / tags are APi in order of signal strength, the number of dimensions of the input vector is two. In this case, the scanned AP list may be expressed as a vector having a px2 size.

And when AP scan list is called {AP1, AP2, ... APP}, AP appears simultaneously APi, ... APi + k Let's say A filter K for filtering a feature in which a plurality of APs appear at the same time becomes a vector having a dimension number k x 2 ( k can be regarded as a window size for performing CNN composite products, and various values of 2 or more can be used) This filter is applied to h APs according to the number of scanned APs. The function Ci using the defined feature is defined as shown in [Equation 1].

[Equation 1]

은 다음의 [수식 2]와 같이 정의한다.Here, b is a bias, F is an activation function, and various nonlinear functions such as sigmoid function or ReLU, which are frequently used in CNN, can be used. The filter K uses the same weight to generate a feature map (feature map), and can define various filters according to the size of k to extract various characteristics of the propagation environment. In a similar manner to the above, various intuitions and strategies described in [Strategy] were defined as several filters. In this case, the filtering result is generated as many as the number of filters, and may be subjected to a process of pooling again the results of each filter as necessary. Pooling can also be handled in various forms, in the present invention uses MaxPooling as an example. The result of the filter of the input scan list is represented by p + h-1 vectors, so the MaxPooling result

Is defined as in the following [Formula 2].

[Formula 2]

At this time, the vector size of the final pooling result should match the number of position labels.

On the other hand, when defining the hypothesis function and the cost function, the hypothesis function can be variously defined as a function used when estimating the position learning and position, the present invention uses a softmax regression function as an example. Using q vectors output as the result of the CNN ( Y _i ) is defined as shown in Equation 3 to output q positions and probabilities using softmax regression analysis.

[Equation 3]

Cost function is a function that calculates the accuracy of the learning result in the position learning. The smaller the error distance between the estimated position and the actual position, the smaller the value and the larger the error distance, the larger the value. In the present invention, a cross-entropy function is used as an example and can be derived from the hypothesis function.

라고 정의하고 해당 라벨의 확률은 1로 정의한다. 상기 가설 함수의 추정 위치 라벨에 대한 확률을

라 할 때 Cost함수는 다음의 [수식 4]와 같다.When learning a location, label the actual location

And the probability of the label is defined as 1. The probability for the estimated position label of the hypothesis function

The cost function is as shown in [Equation 4].

[Equation 4]

And the radio wave environment learning for indoor positioning separates the actual dataset for learning and verification, and learns only with the learning data. Learning is aimed at minimizing the Cost function, and is performed by automatically propagating the weight inside the CNN by backpropagating the weight of the Cost function. Thereafter, the accuracy of the learning may be determined by evaluating whether the actual learning is properly performed using the verification data.

Meanwhile, a method of converting propagation-based positioning resources into an expression applicable to CNN is described as follows.

First, the identifier (MAC address) of the radio wave-based positioning resource is processed as an input node of an artificial neural network used in deep learning by using a WIFI AP or a BLE tag as a fixed input node, and as an input, the first embodiment of FIG. As described above, the signal strength (RSSI) of the one-hot encoding type is input or the signal strength (RSSI) value is input as it is. At this time, the input of the unscanned AP causes all values to be processed to 0 value, and the WIFI AP or BLE tag and the radio wave strength (RSSI) are mapped 1: 1. At this time, by processing the identifier of each AP as a virtual input node, the identifier of the AP / tag and the signal strength (RSSI) can be mapped 1: 1 to be used for learning, and the positioning environment changes (additionally deleted by the AP over time). If necessary, location environment learning can be handled flexibly by adding and removing input nodes.

Second, the method using the vector input and the virtual input node is the same as the third embodiment of FIG. When using the same method as the first and second embodiments described above, an additional storage (disk or memory) must be used because a node that is not used as input data must always exist due to a fixed input node. That is, in the general case, the number of WIFI APs or BLE tags detected in one scan is several to at most several hundred. If there are 4000 APs detected in the building, more than 3000 input nodes are unnecessary at every input. Thus, in the same manner as in the third embodiment of FIG. 3, the number of input nodes may be defined as the maximum number of APs recognized in scanning (m) and used as an input in the form of a [MAC address, RSSI] vector. have. When the number of scanned APs of the actual training data is m or less, the vector is inputted [0,0]. In order to improve the learning efficiency, scanned data is sorted and inputted by RSSI, and this method has an advantage of saving storage (memory / disk).

4 is a flow chart as a first embodiment of the positioning method using the signal strength according to the present invention.

The terminal is located in an arbitrary space and executes a positioning application, thereby scanning a plurality of radio wave-based positioning resources (S100). The radio wave-based positioning resource may be either a WIFI AP or a BLE tag. In this case, the terminal transmits an information request message for positioning determination together with the wave-based positioning resource information scanned to the server.

Thereafter, the terminal receives information including a convolutional layer corresponding to the plurality of radio wave-based positioning resources scanned in step S110, a convolution filter for neural networks, and an artificial neural network weight from the server (S110).

Then, the terminal extracts the feature vector through the convolutional layer by using the ID and signal strength of the scanned radio wave-based positioning resources as input (S120). At this time, in the present invention, as described above in the embodiments of FIGS. 1 to 3, the signal strength used as an input is in the form of One Hot Encoding (OHE), RSSI value form, or RSSI vector form including beer. You will use either.

When the feature vector is extracted in step S120, the feature vector is received as an input, and a location probability value is estimated for each propagation-based location resource scanned in step S100 in a neural network of one or more layers (S130).

Then, the positional positioning of the terminal is determined from the probability value estimated in step S130 (S140), and the terminal displays the determined positional positioning.

5 is a flow chart as a second embodiment of the positioning method using the signal strength according to the present invention.

Since the terminal is located in an arbitrary space and executes a location positioning application, the terminal scans a plurality of radio wave-based positioning resources. At this time, the server receives a plurality of radio wave-based positioning resources scanned from the terminal (S200). The radio wave-based positioning resource may be either a WIFI AP or a BLE tag. In this case, the terminal transmits an information request message for positioning determination together with the wave-based positioning resource information scanned to the server.

Thereafter, the server extracts the feature vector through the convolutional layer using the ID and signal strength of the radio wave-based positioning resource received in step S200 as input (S210). At this time, the server already knows the convolutional layer for the propagation-based location resource, the convolutional filter and the neural network weight for the neural network through learning, and this information is stored. In the present invention, as described above in the embodiments of FIGS. 1 to 3, the signal strength used as an input is in the form of One Hot Encoding (OHE), RSSI value form, or RSSI vector form including beer. You will use either.

When the feature vector is extracted in step S210, the feature vector is received as an input, and a location probability value is estimated for each propagation-based location resource received in step S200 in a neural network of one or more layers (S220).

And the positional positioning of the terminal is determined from the probability value estimated in step S220 (S230), the server transmits the determined positional positioning to the terminal (S240).

FIG. 6 is a diagram illustrating a configuration of a computer device equipped with a positioning application using signal strength according to the present invention, which may be a terminal or a server. In the first embodiment of FIG. 4, the computer device is a terminal. A second embodiment of the is done at the server.

The positioning device 100 using the signal strength includes a processor 110, a nonvolatile storage unit 120 storing programs and data, a volatile memory 130 storing programs being executed, and other devices for communicating with other devices. The communication unit 140 and a bus which is an internal communication path between these devices. Running programs can include device drivers, operating systems, and various applications. Positioning using the signal strength of the present invention is operated by the positioning application 210, the storage unit 120, the convolution filter and the neural network weight is stored, the information including the convolution filter and the neural network weight from the server The terminal is downloaded from the server and used. Although not shown, the electronic device includes a power supply unit such as a battery.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the scope of equivalents of the claims to be described.

100: positioning device using deep learning
110: processor
120: storage unit
130: memory
140: communication unit
210: positioning application

Claims (3)

(a) designating a convolutional neural network (CNN) comprising a plurality of convolutional layers; And
(b) training using input data and labels
Positioning learning method using deep learning comprising a. The method according to claim 1,
The input data is any one of a WIFI AP or a BLE tag which is a radio wave-based positioning resource.
Positioning learning method using deep learning, characterized in that. The method according to claim 1,
The input data is signal strength and is input in one of One Hot Encoding (OHE) form, RSSI value form, and RSSI vector form including beer place.
Positioning learning method using deep learning, characterized in that.

Images