JP7710142B2 - Position estimation device, position estimation method, and position estimation program - Google Patents

Description

The present invention relates to a position estimation device, a position estimation method, and a position estimation program.

The Internet of Things (IOT), in which various devices are connected to the Internet, is becoming a reality. Various devices, such as automobiles, drones, and construction machinery vehicles, are becoming wirelessly connected. Regarding wireless communication standards, the wireless LAN (Local Area Network) defined by the standard IEEE 802.11, Bluetooth (registered trademark), cellular communication using LTE and 5G, LPWA (Low Power Wide Area) communication for IOT, ETC (Electronic Toll Collection System), VICS (Vehicle Information and Communication System), ARIB-STD-T109, and other wireless communication standards that support them are also developing and are expected to become more widespread in the future.

In order to ensure high throughput and reliability, wireless communication devices have adopted MIMO (Multiple input multiple output) communication technology using multiple antennas. MIMO communication technology can improve throughput and reliability by utilizing channel information that indicates how radio waves are propagating between the transmitting and receiving sides. For example, wireless communication devices on the transmitting side support a function for transmitting a feedback signal that conveys channel information to wireless communication devices on the receiving side (see Non-Patent Document 1).

There is also known technology that uses channel information related to radio wave propagation to estimate the position of a wireless communication device (see Non-Patent Documents 2 and 3). For example, the position of a wireless communication device is identified based on the arrival time and level of wireless signals wirelessly communicated with multiple base stations.

“Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications”, IEEE Computer Society, IEEE Std 802.11, 2016, p.2396-p.2400 Y. Tao and others, “A Novel System for WiFi Radio Map Automatic Adaptation and Indoor Positioning”, in IEEE Transactions on Vehicular Technology, vol. 67, no. 11, November 2018, p.10683-p.10692 H. CAO, 5 others, “Indoor Positioning Method Using WiFi RTT Based on LOS Identification and Range Calibration”, in Proc., ISPRS International Journal of Geo-Information, October 2020

However, conventional position estimation technologies had the problem of requiring wireless communication devices to communicate wirelessly with multiple base stations simultaneously and with high time resolution capabilities, resulting in high costs for estimating the position of an object.

The present invention has been made in consideration of the above circumstances, and an object of the present invention is to provide a technology that can estimate the position of an object at low cost and with high accuracy.

A position estimation device of one embodiment of the present invention comprises a wireless communication unit that receives a wireless signal transmitted from a wireless communication unit of a wireless terminal and acquires channel information related to radio wave propagation from the wireless signal, an input feature generation unit that converts the channel information into input features that can be input to a position estimation model, stores the input features, and outputs multiple input features corresponding to multiple times to the position estimation model, and a position estimation model utilization unit that inputs the multiple input features into a position estimation model that models the relationship between the channel information related to radio wave propagation and position information of the wireless terminal using machine learning, thereby estimating and calculating the position of the wireless terminal.

A position estimation method according to one aspect of the present invention is a position estimation method performed by a position estimation device, which includes the steps of receiving a radio signal transmitted from a wireless communication unit of a wireless terminal and acquiring channel information related to radio wave propagation from the radio signal, converting the channel information into input features that can be input to a position estimation model, storing the input features, and outputting a plurality of input features corresponding to a plurality of times to the position estimation model, and inputting the plurality of input features into a position estimation model that models the relationship between the channel information related to radio wave propagation and the position information of the wireless terminal using machine learning, thereby estimating and calculating the position of the wireless terminal.

A position estimation program according to one aspect of the present invention causes a computer to function as the position estimation device described above.

The present invention provides technology that can estimate the position of an object at low cost.

FIG. 1 is a diagram showing the overall configuration of a wireless communication system. FIG. 2 is a diagram showing a position estimation flow (first example) for estimating the position of a wireless terminal. FIG. 3 is a diagram showing a training flow (first example) for training a location estimation model. FIG. 4 is a diagram showing a training flow (second example) for training a location estimation model. FIG. 5 is a diagram showing an expansion flow for expanding training data. FIG. 6 is a diagram showing an example (first example) of expanding the training data. FIG. 7 is a diagram showing an example (second example) of expanding the training data. FIG. 8 is a diagram showing an example (third example) of expanding the training data. FIG. 9 is a diagram showing a position estimation flow (second example) for estimating the position of a wireless terminal. FIG. 10 is a diagram showing the experimental environment area. FIG. 11 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 12 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 13 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 14 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 15 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 16 is a diagram showing a cumulative probability distribution of the positional deviation amount error. FIG. 17 is a diagram illustrating a hardware configuration of the position estimation device.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the description of the drawings, the same parts are given the same reference numerals and the description will be omitted.

Summary of the Invention
The present invention estimates the position of a specific object by using commonly used general-purpose wireless communication information, that is, by using channel information related to radio wave propagation included in a wireless signal of a wireless communication terminal and a position estimation model. Specifically, a position estimation model is generated in advance by modeling the relationship between the channel information included in the wireless signal and the position information of a specific object by machine learning, and channel information acquired from the wireless signal (input feature converted so that the channel information can be input to the position estimation model) is input to the position estimation model, thereby estimating the position of the specific object in the real world at the time of position estimation.

In this way, the present invention estimates the position of a specific object using channel information regarding radio wave propagation contained in the wireless signal of a wireless communication terminal and a position estimation model, making it possible to estimate the position of a specific object using general-purpose wireless communication, and providing a technology that can estimate the position of a specific object at low cost.

Channel information is information about how radio waves propagate between a transmitting wireless communication terminal and a receiving wireless communication terminal, and about the communication quality of wireless communication. For example, it is information about the received power and radio wave propagation coefficient in wireless communication, the channel matrix that represents the state of radio wave propagation between multiple antennas equipped on a transmitting wireless communication terminal and multiple antennas equipped on a receiving wireless communication terminal in MIMO (Multiple Input Multiple Output) communication technology, and the signal-to-noise-interference power.

The input features are the features of channel information that have been converted so that they can be input to the position estimation model. For example, the input features are the unconverted channel information itself, or numerical values obtained by performing various calculations on the channel information.

A specific object is a mobile wireless terminal located in the same environment as the wireless communication terminal. The position of a specific object is, for example, a position on the path along which the specific object is moving, a position in two-dimensional space (such as a map), or a position in three-dimensional space. In addition to this position information, more detailed physical conditions such as orientation and speed may also be estimated.

[Overall configuration of wireless communication system]
FIG. 1 is a diagram showing the overall configuration of a wireless communication system according to this embodiment.

The wireless communication system includes a position estimation device 1 and wireless terminals 2-1 to 2-Q, where Q is an integer greater than or equal to 1.

The position estimation device 1 estimates the positions of the wireless terminals by collecting channel information of the wireless signals transmitted from the wireless terminals 2-1 to 2-Q.

The wireless terminals 2-1 to 2-Q are equipped with wireless communication units 2-1-1 to 2-Q-1. The wireless communication units 2-1-1 to 2-Q-1 transmit wireless signals including pilot signals that are known during transmission and reception, or channel information between the wireless communication units. The wireless communication units are the wireless communication units 1-1 to 1-R provided in the position estimation device 1, or other wireless communication units.

The position estimation device 1 receives the above-mentioned wireless signals via the wireless communication units 1-1 to 1-R. From the received wireless signals, the position estimation device 1 obtains channel information between the wireless communication units 2-1-1 to 2-Q-1 and any of its own wireless communication units 1-1 to 1-R.

Then, the location estimation device 1 inputs the acquired channel information to the input feature generation unit 1-2. The input feature generation unit 1-2 converts the channel information into input features suitable for input to the location estimation model, and inputs the converted input features to the location estimation model utilization unit 1-3.

Then, the position estimation model utilization unit 1-3 inputs the input features of the channel information collected from the wireless terminals 2-1 to 2-Q into a position estimation model that models the relationship between the wireless terminal position information and the input features of the channel information using machine learning, thereby estimating the position of the moving or stopped wireless terminal 2-i (1≦i≦Q).

The location estimation device 1 is, for example, a device installed in a base station installed at a key location, in the cloud, etc. The location estimation device 1 may have any configuration including a wireless communication unit capable of communicating with the wireless terminal 2-i, or a wireless communication unit capable of decoding wireless signals transmitted from the wireless terminal 2-i. The location estimation device 1 does not necessarily need to have the function of transmitting wireless signals, and may be a device that only receives them.

As shown in FIG. 1, the location estimation device 1 includes, for example, wireless communication units 1-1 to 1-R, an input feature generation unit 1-2, a location estimation model utilization unit 1-3, and a location estimation model training unit 1-4.

The wireless communication units 1-1 to 1-R are communication units that perform wireless communication or receive wireless signals. The wireless communication units 1-1 to 1-R may be compatible with multiple frequencies, multiple frequency bands, or multiple wireless communication systems. In FIG. 1, the wireless communication units 1-1 to 1-R are shown configured to receive wireless signals transmitted from the wireless terminal 2-i.

Any of the wireless communication units 1-1 to 1-R may be a base station that communicates with the wireless communication unit 2-i-1 of the wireless terminal 2-i. Alternatively, the wireless terminal 2-i may communicate with any base station not shown in FIG. 1 and receive a wireless signal that it transmits to that base station. In addition, as described below, the wireless terminal 2-i can handle cases where the wireless terminal 2-i moves while connected to multiple base stations by switching the position estimation model based on the ID of the base station with which the wireless terminal 2-i is connected. The wireless communication units 1-1 to 1-R on the receiving side may collect channel information of multiple wireless communication units 2-i-1 by transmitting wireless signals on channels of the same frequency.

The wireless communication units 1-1 to 1-R each have a function of receiving a wireless signal and acquiring, from the received wireless signal, channel information regarding radio wave propagation between the wireless terminal 2-i and the position estimation device 1, or channel information regarding radio wave propagation between a wireless communication device other than the wireless terminal 2-i and the position estimation device 1. When the wireless communication units 1-1 to 1-R acquire channel information corresponding to the wireless terminal 2-i, they input the acquired channel information to the input feature generation unit 1-2.

The input feature generator 1-2 has a function of converting the input channel information into input features that can be input to a position estimation model and storing them. The input feature generator 1-2 inputs multiple input features corresponding to multiple times, which are the converted channel information, to the position estimation model utilization unit 1-3.

The position estimation model utilization unit 1-3 has a function of estimating and calculating the position of at least one wireless terminal 2-i among the wireless terminals 2-1 to 2-Q that are moving or stopped by inputting multiple input feature amounts corresponding to multiple times into a position estimation model, and outputting the estimated and calculated position information. The position estimation model utilization unit 1-3 may use a position estimation model that has been generated in advance, or may use a position estimation model that is generated and updated by the position estimation model training unit 1-4.

The location estimation model is a model generated by training the relationship between the location information of wireless terminal 2-i and channel information (≒ input features) related to radio wave propagation acquired from wireless terminal 2-i using machine learning, and is a model that outputs location information using the input features. Alternatively, the location estimation model may be generated by generating a space equivalent to the real space in a simulation space using digital twin technology, etc., and using the relationship between the virtually generated wireless terminal and the channel information calculated by simulation. The location estimation model may be a location estimation model created from the relationship between the channel information measured by another location estimation unit and the wireless terminal.

The location estimation model training unit 1-4 has a function for generating a location estimation model and a function for updating the location estimation model.

For example, the position estimation model training unit 1-4 separately acquires data regarding the position information of the wireless terminal 2-i, and generates or updates the position estimation model by training a position estimation model capable of estimating the position of the wireless terminal 2-i based on the relationship between the acquired position information and channel information.

Alternatively, the location estimation model training unit 1-4 may generate or update a location estimation model from input features generated from channel information obtained when the wireless terminal 2-i is located at a predetermined, known location and performs a predetermined, known operation, rather than separately acquiring data regarding location information.

The location information of the wireless terminal 2-i that is acquired separately can be realized by any means of acquiring location measurement data, for example, sensors, cameras, radio positioning, SLAM (Simultaneous Localization And Mapping), GPS (Global Positioning System), and sensors, cameras, etc. installed in the environment, mounted on the wireless terminal 2-i.

The position measurement data is, for example, position information and time information of the wireless terminal 2-i, and is stored in the wireless terminal 2-i, and then input to or stored in the position estimation model training unit 1-4 for use. The position estimation model is updated or generated in the position estimation model training unit 1-4 using input features corresponding to the same time as the position measurement data, which are generated in the input feature generation unit 1-2, and training data consisting of the position measurement data.

The location estimation model training unit 1-4 can train the location estimation model by learning the relationship between the stored location and time information of the wireless terminal 2-i and the input feature and time information also stored in the storage unit using training data organized on the same time axis. The performance of the location estimation model can be improved by forming various combinations of training data in a short time. The method of generating the training data will be described later.

In other words, the location estimation model training unit 1-4 has the function of generating a training data set (a training data set including multiple training data) in which multiple location information of the wireless terminal 2-i corresponding to multiple times obtained by a method other than the method of estimation calculation by the location estimation model utilization unit 1-3 (the method of the present invention) and multiple input features corresponding to multiple times are aligned at the same time, and generating or updating a location estimation model using the training data set so as to improve the accuracy of the relationship between the input features and the location information, and updating the weights, biases, etc. of the location estimation model.

In addition, the location estimation model training unit 1-4 has a function of generating or updating a training data set using input features related to channel information acquired from the wireless terminal 2-i at the time when it is detected that the wireless terminal 2-i is located at a predetermined known location or when it is detected that the wireless terminal 2-i is performing a predetermined known operation, and the predetermined known location information of the wireless terminal 2-i.

In addition, the location estimation model training unit 1-4 has a function of generating or updating a location estimation model using multiple training data sets in which training data is extracted at different time intervals from multiple training data included in the training data set, and updating the weights, biases, etc. of the location estimation model.

[Operation of wireless communication system]
[Wireless terminal position estimation method (first example)]
FIG. 2 is a diagram showing a position estimation flow (first example) for estimating the position of a wireless terminal.

First, the wireless communication unit 2-i-1 of the wireless terminal 2-i performs wireless communication with a predetermined wireless communication device and starts transmitting a wireless signal including a pilot signal or channel information (step S1-1).

Next, the wireless communication unit 1-1 of the position estimation device 1 receives the wireless signal transmitted from the wireless communication unit 2-i-1 of the wireless terminal 2-i, and acquires channel information regarding radio wave propagation from the received wireless signal (step S1-2).

Next, the input feature generation unit 1-2 converts the acquired channel information regarding radio wave propagation into multiple input features suitable for input to a position estimation model and corresponding to multiple times, and stores them (step S1-3).

The input feature is, for example, channel information, which is a coefficient of radio wave propagation between antennas, or a numerical value obtained by performing various calculations on the channel information. It is a coefficient related to a channel matrix whose elements are received power, signal power, signal-to-noise power ratio, signal-to-noise interference power ratio, radio wave propagation coefficient between multiple transmitting and receiving antennas, a channel vector which is a column or row vector of a channel matrix, or a value obtained by dividing them by a constant value. It is a standardization to make constant the absolute value or Frobenium norm of a correlation matrix, a channel matrix, or a channel vector obtained by multiplying a channel matrix by a Hermitian matrix of a channel matrix. It is a conversion of a correlation matrix, a channel matrix, or a channel vector into angle information. It is a division of an absolute value or Frobenium norm of a correlation matrix, a channel matrix, or a channel vector by a constant value. It is a conversion of a correlation matrix, a channel matrix, or a channel vector into decibel units and subtraction by a constant value. It is also a value obtained by using a moving average to obtain the average value of the values obtained from power information, power ratio information, or channel information in a recent fixed time period, and normalizing it by that value. It is also possible to use time series information of at least one of these pieces of information. A specific method for calculating the input feature amount will be described later.

The position estimation device 1 may also generate auxiliary information other than the channel information to be input to the position estimation model, and add the generated auxiliary information to the input features (step S1-4). Examples of auxiliary information include specific information for identifying the wireless terminal 2-i, information on the equipment and owner of the wireless terminal 2-1, temperature and position information on the camera, sensor, etc., status information on the position, speed, settings, etc. of the estimation target device and its surrounding reflective structures, temperature, humidity, congestion status, and setting information of the wireless communication system.

Finally, the location estimation model utilization unit 1-3 inputs multiple input features corresponding to multiple times, which are the converted channel information, into the location estimation model, thereby estimating, calculating, and outputting the location information of the wireless terminal 2-i (step S1-5).

[Training method for localization model (first example)]
FIG. 3 is a diagram showing a training flow (first example) for training a location estimation model.

First, the wireless terminal 2-i transmits a wireless signal from which channel information can be obtained while being located at a known position or performing a known operating pattern at a specific time (step S2-1).

Next, the wireless communication unit 1-1 of the position estimation device 1 detects that the wireless terminal 2-i is located at a known position or that the wireless terminal 2-i is performing a known operation, and acquires channel information from the wireless signal received from the wireless terminal 2-i at the time of detection (step S2-2). Here, the wireless terminal 2-i may be any of the wireless terminals 2-1 to 2-Q that are to be subject to position estimation, or it may be one or more wireless terminals selected from among them.

Next, the input feature generation unit 1-2 generates input features from the acquired channel information (step S2-3).

At this time, the position estimation device 1 may generate auxiliary information other than the channel information to be input into the position estimation model, and add the generated auxiliary information to the input features (step S2-4).

Finally, the location estimation model training unit 1-4 generates a training data set using the generated input features and the position or movement of the wireless terminal 2-i at a time corresponding to the time of the input features, and generates or updates a location estimation model using the training data set (step S2-5).

[Training method for localization model (second example)]
FIG. 4 is a diagram showing a training flow (second example) for training a location estimation model.

First, the location estimation model training unit 1-4 generates or acquires location information from another location estimation technology or measurement results other than the present invention, and generates and acquires training data that associates the location information corresponding to the same time with input features (step S3-1).

For example, in a wireless terminal 2-i of an autonomous vehicle, robot, etc., a log related to location information is stored, notified to the location estimation device 1 at a certain frequency, and input features corresponding to the time of the log are stored in the location estimation device 1. By organizing both by time, training data can be obtained.

At this time, the position estimation device 1 may generate auxiliary information other than the channel information to be input into the position estimation model, and add the generated auxiliary information to the input features to generate training data (step S3-2).

Finally, the location estimation model training unit 1-4 generates or updates a location estimation model using the obtained training data (step S3-3). For example, the location estimation model training unit 1-4 updates the weights or biases of the location estimation model using the training data set to improve the accuracy of the relationship between the input features and the location information.

[Method of expanding training data]
5 is a diagram showing an extension flow for extending training data. This extension method is effective when the location estimation model uses time-series data including multiple pieces of time information as input features.

First, the location estimation model training unit 1-4 acquires training data for updating or generating a location estimation model (step S4-1). The training data includes location information and input features over time.

Next, the location estimation model training unit 1-4 increases the number of valid data by changing the time series arrangement of the acquired training data or by arbitrarily setting multiple training data to be used (step S4-2). For example, the location estimation model training unit 1-4 generates multiple training data sets in which training data is extracted from multiple training data with different extraction frequencies in the time axis direction. In this way, the number of valid data can be increased, and the accuracy of location estimation can be further improved.

At this time, the position estimation device 1 may generate auxiliary information other than the channel information to be input into the position estimation model, and add the generated auxiliary information to the input features to generate training data (step S4-3).

Finally, the location estimation model training unit 1-4 generates or updates a location estimation model using the generated training data (step S4-4). For example, the location estimation model training unit 1-4 updates the weights or biases of the location estimation model.

In addition, the location estimation model training unit 1-4 may determine the order in which to use multiple training data sets depending on the acquisition frequency of the input features to be input to the location estimation model or the speed of the target location detection object, and generate or update the location estimation model using the multiple training data sets in the determined order.

Generally, the entire time-series training data used for training is divided into batches of data, and the process of updating the coefficients and weights of the entire model based on the output error is performed on all data as one epoch.

From the multiple training data sets generated by this embodiment, those with the greatest deviation from the frequency of the input information obtained from the location estimation target to be estimated are selected, and model updating is performed using the multiple training data sets in the selected order. Finally, one epoch of model updating is completed using the time series training data with the smallest deviation from the frequency at the time the location prediction is performed, and this process is repeated multiple times to improve the location estimation accuracy of the location estimation model.

In this way, by increasing the amount of valid data per epoch and updating the model again under conditions that are closest to the time series conditions at the time of use, the location estimation performance of the location estimation model obtained for each epoch can be improved.

In particular, if the speed conditions of the target for localization are the same during training and inference, the training data of the time series with the time interval closest to the acquisition interval of the input features during inference is used last. If the expected speed of the target for localization when performing localization is V times the speed when the training data is generated, and if the time interval for acquiring the input features during inference is Tp and the time interval of the time series acquired during training is Tt, the training data of the time series with the time interval at which the target for localization is expected to move the distance closest to the distance moved during the time interval Tp during inference may be used last. Here, if time series training data of three time intervals, Tt, 2Tt, and 3Tt, are generated from the training data generated with the interval Tt, training for each epoch can be performed so that the training data of the time series with a value close to VTt is used last.

Figure 6 shows an example of training data expansion (first example). This training data consists of input features In-1 to In-L and the location information (X, Y) of the wireless terminal to be estimated arranged over time. The acquisition times do not need to match exactly. In this example, the data is organized in a time span of 100 ms. It is also possible to supplement data over time, or to copy missing values from values earlier in time. Other supplementation techniques used in machine learning can also be used.

In the example in Figure 6, training data originally obtained at 100 ms intervals is re-extracted at 200 ms intervals. In the case of a 200 ms interval, it can be seen that two patterns of training data sets are obtained depending on the starting point of the re-extraction. If the total number of time steps of the training data contained in the original training dataset is D, then when the interval is 200 ms, the number of time steps is D/2, and two training datasets are created. If both the 100 ms and 200 ms intervals are used as training data, it becomes D+(D/2)×2, and it can be seen that the training data is doubled.

By reconstructing the training data for different time widths, such as 300 ms or 400 ms periods, it is possible to increase the number of training data by the number of different time widths. Resetting the time step width to a longer value simulates an increase in the moving speed of the target for location estimation. If the channel information is not significantly affected by the moving speed of the wireless terminal, then this increases the amount of pseudo data simulating various moving speeds of the wireless terminal from the same training data set.

Figure 7 shows an example of training data expansion (second example). Training data with a pattern in which training data is randomly removed is generated. Unlike when the time width is simply a fixed multiple of the minimum time width, this simulates more complex wireless terminal movements. In Figure 6, the speed of the wireless terminal is equivalently increased, but with this method, a large amount of training data can be generated without drastically increasing the speed.

Figure 8 is a diagram showing an example (third example) of expanding the training data. This is a method of reversing the chronological order of the training data. Training data is generated by rearranging the data from future to past. In this way, pseudo data equivalent to a wireless terminal moving in the opposite direction to its actual movement can be generated. In addition, by using the training data set generated in Figure 8 together with the training data sets shown in Figures 6 and 7, it is possible to further increase the amount of training data.

[Wireless terminal position estimation method (second example)]
Next, a method of estimating the position when the wireless terminal 2-i moves a long distance, such as when the wireless terminal 2-i moves through the communication areas of a plurality of base stations, will be described.

When wireless communication units 1-1 to 1-R are arranged in a wide area and input features are generated from channel information estimated using pilot signals, multiple position estimation models are prepared corresponding to wireless communication units 1-1 to 1-R, or multiple position estimation models are prepared for combinations of wireless communication units 1-1 to 1-R, and the position of wireless terminal 2-i is estimated by inputting the data into the appropriate position estimation model.

When generating input features by feedback of channel information contained in a wireless signal from wireless terminal 2-i, the channel information does not change regardless of which wireless communication unit receives the signal, so it is necessary to determine to which base station wireless terminal 2-i has fed back the wireless signal. The flow in this case is shown in Figure 9.

First, the wireless terminal 2-i communicates with an appropriate base station from among the multiple installed base stations and transmits a wireless signal including channel information (step S5-1).

Next, the wireless communication unit 1-1 of the position estimation device 1 receives the wireless signal and acquires channel information and the base station ID of the destination base station (step S5-2).

Next, the location estimation model utilization unit 1-3 selects a location estimation model corresponding to the base station ID from among the multiple location estimation models (step S5-3).

Next, the input feature generation unit 1-2 converts the channel information into input features (step S5-4).

At this time, the position estimation device 1 may generate auxiliary information other than the channel information to be input into the position estimation model, and add the generated auxiliary information to the input features (step S5-5).

Finally, the position estimation model utilization unit 1-3 outputs the position of the wireless terminal 2-i by inputting the input features into the selected position estimation model (step S5-6).

[Method of collecting channel information and calculating input features]
An example of a method for collecting channel information and a method for calculating input feature quantities will be described below.

[How to collect channel information and how to calculate input features (first example)]
In the first method, the wireless terminal 2-i transmits a pilot signal that is known in transmission and reception. By transmitting a known pattern in advance, the wireless communication unit 1-r (1≦r≦R) of the position estimation device 1 can obtain a channel matrix between the antenna (number of receiving antennas: Mr) of its own wireless communication unit 1-r and the antenna (number of transmitting antennas: Ni) of the wireless communication unit 2-i-1 that transmitted the pilot signal. If OFDM (orthogonal frequency division multiplexing) is used in various wireless communication systems, it is possible to obtain a channel matrix of subcarriers corresponding to multiple frequencies.

From the channel matrix H of "the number of transmitting antennas Ni × the number of receiving antennas Mr" obtained in this manner, an input feature to be input to the position estimation model utilization unit 1-3 is generated. For example, when a channel matrix is obtained for a plurality of subcarriers by OFDM, the channel matrix of the η-th subcarrier is defined as H _η . As a method of converting into an input feature, first, as shown in formula (1), the channel matrix H _η is separated into a normalized channel matrix G _η normalized by a predetermined norm, and amplitude information γ _η or power information γ _η ² .

G _η can be set so that, for example, ||G _η || _F =1. ||·|| _F represents the Frobenium norm. γ _η generally has a large fluctuation range, and the value can change by 10 to the power of 5 or more. For this reason, it is possible to convert γ _η and γ _η ² into dB, specify their maximum and minimum values, and use values converted to express the range as a normalized value within a range of 0 to 1. It is also possible to use a value γ _all obtained by selecting or averaging multiple values for different frequency conditions and antenna conditions.

Alternatively, as shown in equation (2), the amplitude information γ _η may be separated for each antenna, and the norm value may be normalized to a certain value to obtain each column vector g _{1, η} to g _{Mr, η} and its amplitude values γ _{1, η} to γ _{Mr, η} .

g _{a, η} can be set as a prescribed vector so that, for example, ||g _{a, η} || _F = 1. It is also possible to use values obtained by converting _{γ a, η} or γ _{a, η} ² into dB, prescribing their maximum and minimum values, and expressing the range as a normalized value within a range of 0 to 1. It is also possible to use the value γ _{a, all} corresponding to the a-th column vector obtained by selecting or averaging multiple values for η.

Alternatively, as shown in equation (3), the amplitude information γ _η may be separated for each antenna, and row vectors g' _{1, η} to g' _{Ni, η} with their norm values normalized to a certain value, and their amplitude values γ' _{1, η} to γ' _{Ni, η} may be obtained.

g' _{b, η} can be set as a prescribed vector so that, for example, ||g' _{b, η} || _F = 1. It is also possible to use values obtained by converting γ' _{b, η} or γ' _{b, η} ² into dB, prescribing their maximum and minimum values, and expressing the range as a normalized value within a range of 0 to 1. It is also possible to use the value γ' _{b, all} corresponding to the b-th column vector obtained by selecting or averaging multiple of these for η.

The channel matrix H _η , normalized channel matrix G _η , normalized vectors g _{a, η} , and normalized vectors g' _{b, η} can have the real and imaginary parts of each element as input features, or can be used as input information as imaginary numbers, or can be converted into another format such as angle information, or quantized.

Also, as described above, the power of the received signal may be obtained as γ _η ² at the stage of simply receiving something, without obtaining a channel matrix such as H from the received signal. Such information can be obtained as RSSI (Received Signal Strength Index) in many systems. In addition, any information related to power, such as RSRQ (Reference Signals Received Power) and RSRP (Reference Signal Received Quality) in LTE and 5G systems, may be used.

In addition, since power information fluctuates over time depending on the humidity and temperature of the radio wave propagation environment, the temperatures of the terminal and base station, etc., it is also effective to use moving averages or comparative information for the amplitude value information. In other words, if the value of amplitude information, power information, RSSI, etc. acquired at a certain time is γ[t], the calculated value as shown in equation (4) may be used as the power ratio information ω[t].

The value obtained may be used as is or may be expressed in dB.

Also, correlation matrices H _η H _η ^H , H _η ^{H H} _η generated using a channel matrix H _η can be used. Correlation matrices G _η G _η ^H , G _η ^H G _η generated using a normalized channel matrix G _η can be used. Matrices ΣH η , ΣG _η , ΣH _η H η ^H , ΣH _η H _{H η} , _{ΣG η} ^G _η ^H , ΣG _η ^H G _η obtained by summing or averaging the channel matrix H η , normalized channel matrix G η , and their correlation matrices H _{η H η H} , H _{η H} ^H _η , ^{G η} _{G η H} , and G _η ^H G _η for a plurality of frequencies can be used. Eigenvalues, diagonal matrices, and unitary matrices obtained by performing QR decomposition, SVD (Singular value decomposition), eigenvector decomposition ^, and the like of these matrices can be used.

Furthermore, by using the above matrices ΣH _η , ΣG _η , ΣH _η H _η ^H , ΣH _η ^H H _η , ΣG _η G _η ^H , and ΣG _η ^H G _η , the power characteristics for the direction of arrival for the communication device, obtained by the direction of arrival estimation technology, may be used as the input feature. For example, the values obtained by multiplying the vector components by (1, exp(jdθ), exp(j2dθ), ..., exp(jNdθ)) can be calculated for θ. It is also possible to generate θ at any angle interval from 0 to 2π, and use the outputs for multiple θ as the input feature. d is a predetermined constant. N is the number of elements in the vector.

That is, the input feature generating unit 1-2 receives, as input features, the received power of a wireless signal, signal power, power ratio information obtained from a moving average of the received power and signal power, a channel matrix consisting of radio wave propagation coefficients between multiple antennas, a correlation matrix or pseudo-correlation matrix of a channel matrix, an operational matrix obtained by signal processing of a channel matrix, an operational matrix obtained by signal processing of a correlation matrix or pseudo-correlation matrix, an operational matrix obtained by signal processing of a channel matrix, correlation matrix, or pseudo-correlation matrix corresponding to multiple frequencies, a unitary matrix obtained by linearly operating a channel matrix, a correlation matrix or pseudo-correlation matrix, Among features relating to a unitary matrix obtained by linearly operating columns, a unitary matrix obtained by linearly operating an operation matrix, a diagonal matrix obtained by linearly operating a channel matrix, a diagonal matrix obtained by linearly operating a correlation matrix or a pseudo-correlation matrix, a diagonal matrix obtained by linearly operating an operation matrix, a triangular matrix obtained by linearly operating a channel matrix, a triangular matrix obtained by linearly operating a correlation matrix or a pseudo-correlation matrix, and a triangular matrix obtained by linearly operating an operation matrix, the input feature generator 1-2 generates values of phase, amplitude, real component, imaginary component, and standardized values of the range of coefficients of one or more of the features. The input feature generator 1-2 stores such input feature as time-series data, and outputs input feature corresponding to a plurality of times from the past to the present to the position estimation model.

In a wireless communication system that uses equalization technology, it is possible to estimate the arrival time, power, and phase conditions of the path of an incoming electrical signal. Even channel information obtained in this way as a time series can be used as input features for a position estimation model by normalizing the power, converting it to frequency components, and using features and angle information extracted by existing direction-of-arrival technology.

[Method of collecting channel information and calculation of input features (second example)]
In the second method, the wireless terminal 2-i communicates with a specific wireless base station, estimates channel information from a pilot signal transmitted from and received by the specific wireless base station, generates feedback information by quantizing the signal in some way, and transmits a wireless signal including the generated feedback information. The wireless communication unit 1-r of the position estimation device 1 receives the wireless signal and obtains channel information between the specific wireless base station and the wireless terminal 2-i, which is included in the received wireless signal.

First, a pilot signal that is known for transmission and reception is transmitted from a specific known wireless base station. By transmitting a known pattern in advance, the wireless communication unit 2-i-1 of the wireless terminal 2-i can obtain the channel matrix between its own receiving antenna (number of receiving antennas: Ni) and the antenna (number of transmitting antennas: Mt) of the specific wireless base station that transmitted the pilot signal. In the case of OFDM, which is used in various wireless communication systems, it is possible to obtain the channel matrix of subcarriers corresponding to multiple frequencies.

From the channel matrix _Hα of "the number of transmitting antennas Mt × the number of receiving antennas Ni" obtained in this manner, an input feature to be input to the position estimation model utilization unit 1-3 is generated. For example, when a channel matrix is obtained for a plurality of subcarriers by OFDM, the channel matrix of the η-th subcarrier is defined as _Hα,η . Here, when a specific wireless base station is the wireless communication unit 1-r of the position estimation device 1, the number of transmitting antennas Mt is equal to the number of receiving antennas Mr defined for the wireless communication unit 1-r in the first method. In this case, the channel matrix Hα _,η corresponds to the transposed matrix of the channel matrix _Hη .

Therefore, as a method of converting into input features, similar to the first method, the channel matrix H _α,η is separated into a normalized channel matrix G _α,η normalized by a predetermined norm and amplitude information γ _α,η or power information γ _η2 as shown in equation ( ⁵ ).

G _{α, η} can be set so that, for example, ||G _{α, η} || _F = 1. ||·|| _F represents the Frobenium norm. _{γ α, η} and γ _{α, η} ² may be converted to dB, their maximum and minimum values may be specified, and the converted values may be expressed as normalized values within a range of 0 to 1. A value γ _{α, all} obtained by selecting or averaging multiple of these may also be used. When γ _{α, all} is obtained by averaging, it may be averaged using true values, or may be averaged after being converted to dB, or the true values may be averaged in dB units.

Alternatively, as shown in equation (6), the amplitude information γα _,η may be separated for each antenna, and the norm value may be normalized to a certain value to obtain column vectors gα _,1,η to gα _,Ni,η and their amplitude values γα _,1,η to γα _,Ni,η .

g _α,a,η can be set as a prescribed vector so that, for example, ||g _α,a,η || _F = 1. It is also possible to use values obtained by converting _{γ α,a,η} or γ _α,a,η ² into dB, prescribing their maximum and minimum values, and expressing the range as a normalized value within a range of 0 to 1. It is also possible to use the value γ _α,a,all corresponding to the a-th column vector obtained by selecting or averaging multiple of these for η.

Alternatively, as shown in equation (7), the amplitude information γα _,η may be separated for each antenna, and row vectors g'α _,1,η ∼ g'α _,Mt,η with their norm values normalized to a certain value, and their amplitude values γ'α _,1,η ∼ γ'α _,Mt,η may be obtained.

g' _α,b,η can be set as a prescribed vector so that, for example, ||g' _α,b,η || _F = 1. It is also possible to use values obtained by converting γ' _α,b ,η or γ' _α,b,η ² into dB, prescribing their maximum and minimum values, and expressing the range as a normalized value within a range of 0 to 1. It is also possible to use the value γ' _α,b,all corresponding to the b-th column vector obtained by selecting or averaging multiple of these for η.

The channel matrix H _{α, η} , normalized channel matrix G _{α, η} , normalized vector g _{α, a, η} , and normalized vector g' _{α, b, η} can have the real and imaginary parts of each element as input features, or can be used as input information as imaginary numbers, or can be converted into another format such as angle information, or quantized.

Also, correlation matrices Hα _{,ηHα ,ηH and Hα} , _ηHHα,η generated using channel matrices _Hα,η can be used. Correlation matrices Gα, ^ηGα ,ηH _{and Gα} ^, ηHGα _,η generated using normalized channel matrices Gα,η can be used. Matrices ΣHα _{, η} ,ΣGα _,η , ^ΣHα ,η,ΣHα _, ^ηHα ,η,ΣHα, _ηHHα,η , ^ΣHα _{,ηHHα , η} ,ΣGα _{,ηGα ,ηH} and ^ΣGα _,ηHGα ^, _η obtained by summing or averaging the channel matrices _Hα ,η, normalized channel matrices ^Gα , _η and their correlation matrices _Hα,ηHα , _{ηH, Hα ,ηHHα,} ^{η ,} _{Gα , ηGα ,ηH and} ^Gα _, ^ηHGα _,η for a plurality of frequencies can be used. Eigenvalues, diagonal matrices, and unitary matrices obtained by performing QR decomposition, SVD, eigenvector decomposition, etc. of these matrices can be used.

As an example, we will show a case where the feedback of channel information used in the wireless LAN standard IEEE 802.11n/ac/ax is used. In this case, the right singular matrix obtained by SVD of the channel matrix H _α,η in the above example is compressed into angle information, which is quantized and generated corresponding to multiple frequencies, and then fed back.

Here, U _η is a left singular matrix of Ni×Ni, Σ _η is a diagonal matrix of Ni× _Ni with eigenvalues λ _1,η to λ Ni,η as diagonal elements, 0 is a matrix with zero elements of Ni×(Mt-Ni), (V _S,η V _N,η ) ^H is a right singular matrix of Mt×Mt, V _S,η is a vector corresponding to the eigenvalue, and V _N,η corresponds to the 0 matrix. Ni is the number of receiving antennas. Mt is the number of transmitting antennas. In the feedback, vectors and eigenvalues are generated for the number of streams to be spatially multiplexed. In the following, it is assumed that feedback is performed for the number of receiving antennas Ni, but any number less than Ni may be used. The vectors V _η of the right singular matrix Ni are fed back, and the SNR value obtained from the eigenvalues. The SNR is not calculated for each subcarrier, but is averaged. For example, λ _1,η to λ _Ni,η are averaged over the produced subcarriers, and λ _1, to λ _Ni are divided by the thermal noise level to provide feedback as S ₁ to S _Ni . V _η can be converted to angle information (φ, ψ) and fed back as shown in equation (9).

This expression is written with a focus on a certain frequency, and the V matrix in equation (9) exists for the specified number of subcarriers, and angle information is generated for each subcarrier. It is then quantized with a specified number of quantization bits, stored in a wireless signal, and transmitted. The wireless communication unit 1-r of the position estimation device 1 can obtain this angle information and SNR information, and can also obtain the RSSI information of the wireless signal.

The angle information may be used as the input feature as it is. The sine and cosine components calculated from the angle information may be used as the input feature. A matrix obtained by converting the angle information back to a right singular matrix using formula (9) may be used. After converting the angle information back to a right singular matrix, an averaging matrix obtained by averaging the right singular matrix or its correlation matrix in the frequency direction may be used. A matrix obtained by further performing signal processing such as QR decomposition on the averaging matrix may be used. For example, when converting back to a correlation matrix, V _η corresponding to the η-th subcarrier is obtained, but only a representative value of the SNR as a whole is obtained. For this reason, a pseudo-correlation matrix obtained as an approximation is calculated, for example, as shown in formula (10).

Here, Θ represents the group of subcarriers from which V _η is obtained, and N _Θ is the number of elements of Θ. Since the lower triangular part and the upper triangular part of the off-diagonal elements of the correlation matrix obtained here are complex conjugates of each other, when using them as input features in machine learning, the diagonal elements and the lower left half or upper right half of the off-diagonal elements of the pseudo-correlation matrix R obtained here can be used. Since the off-diagonal elements are imaginary numbers, general machine learning algorithms can be applied by using the real and imaginary parts as real numbers and using them as input features.

In addition, in equation (10), even if the number of columns of V _η fed back or the number of SNRs (eigenvalues) change over time, in order to prevent the correlation matrix of equation (10) from being affected, the correlation matrix of equation (10) may be calculated using only the first vector and the first SNR of V _η , or the number of columns and SNRs of vector number N _Θ ', which is smaller than the number of columns currently obtained (N _Θ in the case of equation (10)), may be selected to calculate the correlation matrix and obtain it as the input feature.

By calculating as above, the elements of the calculated R can be in the range of 0 to 1 or -1 to 1, and the standardization process can be omitted when inputting it into deep learning.

Furthermore, as mentioned above, when calculating the correlation matrix by selecting the first vector, the first diagonal element is 1, so simply

Here, V _η,1 represents the first column vector of V _η . As described above, in order to standardize the range of R, the denominators of equations (10) to (12) may be omitted and divided by a standardization coefficient.

In addition, when the right singular matrix V is restored from angle information, the imaginary part of the last element of each column vector is always 0, so if the right singular matrix is obtained as an M×Ni matrix, the value 2×Mt×Ni-Ni becomes meaningful information from the numerical values of the real and imaginary parts of each element. For example, if the right singular matrix is a 4×1 matrix, a total of 7 elements with a real part of 4 and an imaginary part of 3 are meaningful information, and if a 4×2 matrix is obtained, a total of 14 elements with a real part of 8 and an imaginary part of 6 are meaningful information. Since the imaginary part of the last element of each column is 0, the last element of each column does not need to be used.

[Experimental Results]
The position estimation method and its effects according to this embodiment will be described with reference to specific examples and indoor experimental results.

Two wireless communication units 1-1 and 1-2 were installed in the indoor experimental environment area shown in Figure 10, and the position of wireless terminal 2-1 was estimated. Wireless terminal 2-1 was traveling in a figure-8 pattern in the center of the experimental environment area, and was transmitting a wireless signal including channel information as feedback to the base station (AP: Access Point).

The wireless terminal 2-1 is requested by the base station AP to report channel information every 100 ms, and transmits feedback of angle information of the channel information using the channel information feedback method defined in the wireless LAN standard IEEE 802.11ac. The base station AP has four antennas. The wireless terminal 2-1 and the wireless communication units 1-1 and 1-2 each have two antennas.

The two wireless communication units 1-1 and 1-2 can obtain angle information and SNR (λ ₁ , λ ₂ ) generated from the right singular matrix of the channel matrix between the wireless terminal 2-1 and the base station AP. The wireless communication unit 1-1 can also obtain an RSSI value (γ ₁₁ , γ ₁₂ ), and the wireless communication unit 1-2 can also obtain an RSSI value (γ ₂₁ , γ ₂₂ ). There are two RSSI values each because the wireless communication units 1-1 and 1-2 each have two receiving antennas, and the two values correspond to the respective antennas.

The input feature generator 1-2 calculates a pseudo correlation matrix from the angle information according to the above formula, averages the calculated pseudo correlation matrix by frequency, calculates the pseudo correlation matrix R in formula (10), and divides R by the sum of the diagonal terms of the obtained R for normalization. For the pseudo correlation matrix R, a total of 16 input features are generated, including four diagonal terms and 12 off-diagonal terms, which are the real and imaginary parts of the six elements in the lower left half. This is the pseudo correlation matrix feature (R ₁ , …, R ₁₆ ), denoted as R. The dB values of the two SNRs (λ ₁ , λ ₂ ) included in the feedback signal are divided by a value that exceeds the value that can be observed in the experimental environment, and the normalized coefficient is denoted as SNR. Moreover, the dB value of RSSI (γ ₁₁ , γ ₁₂ ) at the receiving antenna of wireless communication unit 1-1 is normalized to be distributed in the range of 0 to 1 and expressed as RSSI, and the result obtained using wireless communication unit 1-1 is used as RSSI.

Here we explain how to generate a location estimation model.

To generate a position estimation model according to this embodiment, an autonomous robot, wireless terminal 2-1, was driven for four days in an indoor experimental environment area. The lines remaining in the figure-eight shape shown in Figure 10 are the actual route of the autonomous robot, and to complicate the movement, the robot heads toward a goal defined as a 60 cm square, but within this square, the point it actually aims for was set randomly, with a 50% probability of skipping.

At this time, the position estimation model training unit 1-4 generates training data by organizing the highly accurate autonomous robot position information obtained from the LIDAR (light detection and ranging) and tire control information provided in the wireless terminal 2-1, and the above-mentioned input feature amounts related to propagation, in the same time series at 200 ms intervals. As mentioned above, the input feature amount generation period is 100 ms, but the training data is generated by selecting and using the latest information obtained in a time span divided by 200 ms intervals.

Then, using the generated training data for 8 hours, a localization model was trained using a deep neural network that uses GRU (Gated Recurent Unit) and direct connections. The learning rate was 0.0002, and the optimization algorithm was ADAM. The GRU had 1 hidden layer and a dimension of 400, and the weights and biases of the localization model were updated by backpropagation to output information on X and Y coordinates within the experimental environment area, using two direct connection layers with input 400 and output 400, and one direct connection layer with input 400 and output 2. The updated localization model that minimized the MSE using 30 minutes of data used for validation was used.

Figure 11 is a graph plotting the amount of position deviation as an error using a cumulative distribution function (CDF) when the position of wireless terminal 2-1 is estimated using the position estimation model generated by this embodiment.

The "SNR/RSSI/R" graph shows the location estimation error results using a location estimation model generated by inputting 20 samples of each of the aforementioned pseudo-correlation matrix R, SNR, and RSSI (a total of 20 (=16 + 2 + 2)) as time-series data into an RNN (Recurrent Neural Network), training, and generating the model.

The graph "SNR/RSSI/R (immediately after training)" shows the error results obtained by using the location estimation model on test data measured at different times from the training data after running for four days. A lower location estimation error can be seen than in "SNR/RSSI/R".

The other error results (graphs for "R", "SNR/RSSI", and "single-time SNR/RSSI") are the cumulative probability distribution of the error when the location estimation model generated using four days of training data is used to measure the same features 30 days later and estimate the location.

First, the results using a total of 20 input features of "SNR/RSSI/R" show that the location estimation error has increased compared to the characteristics used immediately after training, increasing from a median error of 7.7 cm to 12.7 cm, but still achieves a highly accurate location estimation effect with a location estimation error of about 10 cm. This is significantly higher than the estimation result of "SNR/RSSI", which was trained using a total of 20 samples of the four input features of SNR and RSSI in a time series and applied to the training data 30 days later. "Single-time SNR/RSSI", which does not use time series data and uses only a single sample in the time direction, has an even larger location estimation error. This confirms the high location estimation accuracy achieved by the pseudo correlation matrix R.

Figure 12 shows the cumulative probability distribution of the estimation errors for "SNR/RSSI/R", "R", and "SNR/RSSI" for test data from different times, data 34 days later, and data 6 days later, trained on four days of data as in the previous case, with the dimension of the GRU increased to 2000, the first two fully connected layers having an input of 2000 and an output of 2000, and the last layer having an input of 2000 and an output of 2, and the model is then trained on four days of data as in the previous case.

These results show that the results of the pseudo-correlation matrix R have further improved, approaching the estimation error of "SNR/RSSI/R". In addition, it can be seen that although the performance of "SNR/RSSI" is greatly improved when used immediately after training, the location estimation accuracy deteriorates after one or two months.

Figures 13 and 14 show graphs verifying the above-mentioned training method. Here, we consider training a localization model for the following cases using training data from feedback signals obtained at 100 ms intervals, with the GRU dimension set to 500, the first two fully connected layers with input 500 and output 500, and the last layer with input 500 and output 2.

Case 1 is a case in which 40 samples of input features every 100 ms are input to a deep learning position estimation model in time series using a feedback signal obtained every 100 ms, and a position estimation model is trained to estimate the position of the robot.

Case 2 is a case in which 40 samples of input features every 200 ms are input to a deep learning position estimation model in time series using a feedback signal obtained every 200 ms to train a position estimation model that estimates the robot's position.

In Figure 13, the case where only the original training data obtained every 100 ms was used when training the position estimation model is shown as "100 ms interval data only."

Using the data expansion method shown in Figure 6, a data set with 100 ms intervals is reconstructed into two data series with 200 ms intervals, and during training, the data of the two data series is divided into batch data, and the output error for all batch data is evaluated as one epoch, and the coefficients and weights of the entire location estimation model are updated. In addition, the model is updated from the output of each batch data from the training data with 200 ms intervals, and after considering all the data, batch data is generated from the training data with 100 ms intervals, and all the data in this is considered as one epoch, and the location estimation model is updated. The estimation error in this case is described as "200, 100 ms interval data".

Furthermore, the "400, 300, 200, 100 ms interval data" is the result of repeating one epoch of training using four data series spaced at 400 ms intervals, three data series spaced at 300 ms intervals, two data series spaced at 200 ms intervals, and data series spaced at 100 ms intervals, in that order. When applied to the 200 ms training data generated by the data expansion of the present invention, the estimation error was reduced by 7%.

In other words, in Fig. 13, the location estimation model training unit 1-4 updates the location estimation model using multiple training data sets, such as "data at intervals of 400, 300, and 200 ms," and then finally updates the location estimation model using the original training data set, which is "data at intervals of 100 ms," that was previously generated. This makes it possible to further reduce the location estimation error, as described above.

In Figure 14, the effect of the data augmentation method in Figure 6 is also verified for a location estimation model that estimates location using time-series input features every 200 ms. First, as with the data augmentation method shown in Figure 6, training data for "only data at 200 ms intervals" is used for learning by extracting two training data sets at 200 ms intervals from training data that existed at 100 ms intervals. Comparing Figure 14 with Figure 13, it can be seen that the location estimation performance is almost the same as that of Figure 13, even though the frequency of use of the data used is halved.

In addition, three pieces of data spaced 300 ms apart were generated from the original 100 ms training data, and the location estimation model was trained using the training data spaced 300 ms apart, followed by training the location estimation model using training data spaced 200 ms apart. This is shown as "300, 200 ms interval data." As with the case of "100 ms interval data only" shown in Figure 13, it can be confirmed that the location estimation accuracy is improved and errors are reduced.

Furthermore, similarly high results were obtained for the results of "400, 300, 200 ms interval data" using a location estimation model trained using data at intervals of 400, 300, and 200 ms in that order. In this way, data augmentation can be used to expand the limited training data, thereby improving the location estimation accuracy of the location estimation model.

In other words, in Fig. 14, the location estimation model training unit 1-4 updates the location estimation model by using later training data sets, among multiple training data sets such as "400, 300, and 200 ms interval data," the training data sets whose time intervals are closer to the original training data set, which is the "100 ms interval data" generated in advance. This makes it possible to further reduce the location estimation error, as described above.

To confirm the effect of the pseudo correlation matrix, the V matrix restored from feedback information is averaged by the number of subcarriers instead of the 16 features of the pseudo correlation matrix R, and the difference in location estimation accuracy using the information V of the real and imaginary parts of the resulting 4 × 2 matrix is shown in Figure 15. From Figure 15, it can be confirmed that the location estimation performance is higher when the pseudo correlation matrix R, SNR, and RSSI are used than when the V matrix is averaged.

Fig. 16 is a diagram comparing the estimation accuracy in the experimental environment area shown in Fig. 10 using the RSSI values ( _γ11 , _γ12 ) in wireless communication unit 1-1 and two RSSI values ( _γ21 , _γ22 ) in wireless communication unit 1-2, setting the GRU dimension to 400, the first two fully connected layers having input 400 and output 400, and the last layer having input 400 and output 2, and training with four days of data as in the previous case, and using only the RSSI values ( _γ11 , _γ12 ) in wireless communication unit 1-1. Although the effect of adding wireless communication unit 1-2 is not large, the median error of 6.48 cm is improved to 5.97 cm.

[Effects of this embodiment]
According to this embodiment, the position estimation device 1 includes wireless communication units 1-1 to 1-R that receive wireless signals transmitted from wireless communication units of wireless terminals and acquire channel information related to radio wave propagation from the wireless signals, an input feature amount generation unit 1-2 that converts the channel information into input feature amounts that can be input to a position estimation model, stores the input feature amounts, and outputs a plurality of input feature amounts corresponding to a plurality of times to the position estimation model, and a position estimation model utilization unit 1-3 that inputs the plurality of input feature amounts into a position estimation model that models the relationship between the channel information related to radio wave propagation and the position information of the wireless terminal by machine learning, thereby estimating and calculating the position of the wireless terminal. In other words, the position of the wireless terminal is estimated using the channel information related to radio wave propagation included in the wireless signal of the wireless terminal and the position estimation model, so that the position of the wireless terminal can be estimated by general-purpose wireless communication, and a technology that can estimate the position of a wireless terminal at low cost and with high accuracy can be provided.

[others]
The present invention is not limited to the above-described embodiment, and various modifications are possible within the scope of the present invention.

The position estimation device 1 of the present embodiment described above can be realized, for example, as shown in FIG. 16, using a general-purpose computer system including a CPU 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906. The memory 902 and the storage 903 are storage devices. In the computer system, the CPU 901 executes a predetermined program loaded onto the memory 902, thereby realizing each function of the position estimation device 1.

The position estimation device 1 may be implemented in one computer. The position estimation device 1 may be implemented in multiple computers. The position estimation device 1 may be a virtual machine implemented in a computer. The program for the position estimation device 1 may be stored in a computer-readable recording medium such as a HDD, SSD, USB memory, CD, or DVD. The program for the position estimation device 1 may also be distributed via a communication network.

1: Position estimation device
1-1 to 1-R: Wireless communication unit
1-2: Input feature generation unit
1-3: Location estimation model utilization section
1-4: Location estimation model training section
2-1 to 2-Q: Wireless terminals
2-1-1 to 2-Q-1: Wireless communication section

Claims (7)

a wireless communication unit that receives a wireless signal transmitted from a wireless communication unit of a wireless terminal and acquires channel information related to radio wave propagation from the wireless signal;
an input feature generating unit that converts the channel information into an input feature that can be input to a position estimation model, stores the converted input feature, and outputs a plurality of input feature values corresponding to a plurality of times to the position estimation model;
a location estimation model utilization unit that estimates and calculates a location of the wireless terminal by inputting the plurality of input feature amounts into a location estimation model that models a relationship between channel information related to the radio wave propagation and location information of the wireless terminal by machine learning;
a location estimation model training unit that updates the location estimation model,
The location estimation model training unit is
A location estimation device that generates a training data set in which multiple location information of the wireless terminal corresponding to multiple times obtained by a method other than the estimation calculation method and multiple input features corresponding to the multiple times are aligned at the same time, and updates the location estimation model using the training data set to improve accuracy of the relationship between the input features and the location information . The location estimation model training unit is
2. The position estimation device according to claim 1, wherein the training data set is generated using input features relating to channel information acquired from the wireless terminal at a time when the wireless terminal is detected to be located at a predetermined known position or when the wireless terminal is detected to be performing a predetermined known operation, and predetermined known position information of the wireless terminal. The location estimation model training unit is
The position estimation device according to claim 1 or 2 , wherein the position estimation model is updated using a plurality of training data sets each obtained by extracting training data at different time intervals from the training data set. The location estimation model training unit is
4. The position estimation device according to claim 3, wherein, after updating the position estimation model using the extracted plurality of training data sets, the generated training data set is used last to update the position estimation model, or, among the extracted plurality of training data sets, a training data set having a time interval closer to the generated training data set is used later to update the position estimation model. The input feature generation unit
As the input feature amount, the received power of the wireless signal, the signal power, power ratio information obtained from a moving average of the received power or the signal power, a channel matrix consisting of radio wave propagation coefficients between a plurality of antennas, a correlation matrix or a pseudo-correlation matrix of the channel matrix, an operational matrix obtained by signal processing the channel matrix, an operational matrix obtained by signal processing the correlation matrix or the pseudo-correlation matrix, an operational matrix obtained by signal processing the channel matrix or the correlation matrix or the pseudo-correlation matrix corresponding to a plurality of frequencies, a unitary matrix obtained by linearly operating the channel matrix, and a unitary matrix obtained by linearly operating the correlation matrix or the pseudo-correlation matrix. 5. The position estimation device according to claim 1 , further comprising: a unitary matrix obtained by linearly operating the operation matrix; a diagonal matrix obtained by linearly operating the channel matrix; a diagonal matrix obtained by linearly operating the correlation matrix or the pseudo correlation matrix; a diagonal matrix obtained by linearly operating the operation matrix; a triangular matrix obtained by linearly operating the channel matrix; a triangular matrix obtained by linearly operating the correlation matrix or the pseudo correlation matrix; and a triangular matrix obtained by linearly operating the operation matrix, the unitary matrix obtained by linearly operating the operation matrix; A position estimation method performed by a position estimation device,
receiving a radio signal transmitted from a radio communication unit of a radio terminal and acquiring channel information related to radio wave propagation from the radio signal;
converting the channel information into an input feature quantity that can be input to a position estimation model, storing the input feature quantity, and outputting a plurality of input feature quantities corresponding to a plurality of times to the position estimation model;
a step of inputting the plurality of input feature amounts into a position estimation model that models a relationship between channel information regarding radio wave propagation and position information of the wireless terminal by machine learning, thereby estimating and calculating a position of the wireless terminal;
updating the location estimation model;
In the updating step,
A location estimation method that generates a training data set in which multiple pieces of location information of the wireless terminal corresponding to multiple times obtained by a method other than the estimation calculation method and multiple input features corresponding to the multiple times are aligned at the same time, and updates the location estimation model using the training data set to improve accuracy of the relationship between the input features and the location information . A position estimation program that causes a computer to function as the position estimation device according to any one of claims 1 to 5 .