Position Estimation Device, Position Estimation Method, and Position Estimation Program

Abstract

A position estimation device 1 comprises: wireless communication units 1-1 to 1-R which each receive a wireless signal transmitted from a wireless communication unit of a wireless terminal and acquire, from the wireless signal, channel information pertaining to radio wave propagation; an input feature amount generation unit 1-2 which converts the channel information into an input feature amount that can be input to a position estimation model, stores the input feature amount, and outputs, to the position estimation model, a plurality of input feature amounts corresponding to a plurality of times; and a position estimation model utilization unit 1-3 which calculates and estimates the position of the wireless terminal by inputting the plurality of input feature amounts into the position estimation model that has been obtained by modeling, by machine learning, a relationship between the channel information pertaining to the radio wave propagation and positional information pertaining to the wireless terminal.

Description

Technical Field

[0001] The present invention relates to a position estimation device, a position estimation method, and a position estimation program.

Background Art

[0002] The realization of the Internet of Things (IoT) in which various devices are connected to the Internet is progressing. Various devices such as automobiles, drones, and construction machinery vehicles are being wirelessly connected. With regard to wireless communication standards, wireless communication standards supported by wireless LAN (Local Area Network) defined by the standard IEEE 802.11, Bluetooth (registered trademark), cellular communication by LTE or 5G, LPWA (Low Power Wide Area) communication for IoT, ETC (Electronic Toll Collection System) used for vehicle communication, VICS (Vehicle Information and Communication System), ARIB-STD-T109, etc. are also evolving and are expected to spread in the future.

[0003] In order to ensure high throughput and reliable performance, wireless communication devices have introduced MIMO (Multiple Input Multiple Output) communication technology using multiple antennas. The MIMO communication technology can improve throughput and reliable performance by using channel information indicating how radio waves propagate between the transmission side and the reception side. For example, the wireless communication device on the transmission side is supported with a function of transmitting a feedback signal for transmitting channel information to the wireless communication device on the reception side (see Non-Patent Document 1).

[0004] Also, a technique of using channel information regarding radio wave propagation for estimating the position of a wireless communication device is known (see Non-Patent Documents 2 and 3). For example, the position of a wireless communication device is specified based on the arrival time, level, etc. of wireless signals wirelessly communicated with a plurality of base stations.

Prior Art Documents

Non-Patent Literature

[0005]

Non-Patent Literature 1

Non-Patent Literature 2

Non-Patent Literature 3

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, conventional position estimation techniques have a problem in that in wireless communication devices, since they require a device that can perform wireless communication with a plurality of base stations simultaneously and has high-performance time resolution performance, a large cost is required for estimating the position of an object.

[0007] The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a technique capable of estimating the position of an object with high accuracy at low cost.

Means for Solving the Problem

[0008] A position estimation device according to an aspect of the present invention includes 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, and converts the channel information into an input feature amount that can be input to a position estimation model, stores it, 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 that estimates and calculates the position of the wireless terminal by inputting the plurality of input feature amounts into a position estimation model in which the relationship between the channel information related to radio wave propagation and the position information of the wireless terminal is modeled by machine learning.

[0009] A position estimation method according to an aspect of the present invention is a position estimation method performed by a position estimation device, and includes receiving a wireless signal transmitted from a wireless communication unit of a wireless terminal and acquiring channel information related to radio wave propagation from the wireless signal. Converting the channel information into an input feature amount that can be input to a position estimation model, storing it, and outputting a plurality of input feature amounts corresponding to a plurality of times to the position estimation model; and the plurality of input feature amounts are related to the radio wave propagation. Estimating and calculating the position of the wireless terminal by inputting the position estimation model in which the relationship between the channel information and the position information of the wireless terminal is modeled by machine learning.

[0010] A position estimation program according to an aspect of the present invention causes a computer to function as the above-described position estimation device.

Advantages of the Invention

[0011] According to the present invention, it is possible to provide a technique for estimating the position of an object at low cost.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

MODE FOR CARRYING OUT THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the description of the drawings, the same parts are denoted by the same reference numerals and the description thereof is omitted.

[0014] [SUMMARY OF THE INVENTION] The present invention estimates the position of a specific object by using general-purpose wireless communication information that is usually used, that is, by using channel information related to radio wave propagation included in the wireless signal of a wireless communication terminal and a position estimation model. Specifically, a position estimation model in which the relationship between channel information included in a wireless signal and the position information of a specific object is modeled by machine learning is generated in advance, and the channel information (input feature quantity obtained by converting the channel information so as to be inputtable to the position estimation model) obtained from the wireless signal is input to the above position estimation model, thereby estimating the position of the specific object in the real world at the time of position estimation.

[0015] As described above, since the present invention estimates the position of a specific object by using channel information related to radio wave propagation included in the wireless signal of a wireless communication terminal and a position estimation model, it is possible to perform the position estimation of the specific object by general-purpose wireless communication, and a technique capable of estimating the position of the specific object at low cost can be provided.

[0016] Note that the channel information is information regarding how radio waves are propagating between the wireless communication terminal on the transmission side and the wireless communication terminal on the reception side, and the communication quality of the wireless communication. For example, in the received power and radio wave propagation coefficient in wireless communication, in MIMO (Multiple Input Multiple Output) communication technology, it is the channel matrix representing the state of radio wave propagation between a plurality of antennas provided in the wireless communication terminal on the transmission side and a plurality of antennas provided in the wireless communication terminal on the reception side, and information regarding signal-to-noise interference power.

[0017] The input feature quantity is a feature quantity of the channel information obtained by converting the channel information so as to be inputtable to the position estimation model. For example, the input feature quantity is the channel information itself without conversion, or a numerical value obtained by performing various operations on the channel information.

[0018] The specific object is a movable wireless terminal located in the same environment as the wireless communication terminal. The position of the specific object is, for example, the position on the path along which the specific object is moving, the position in a two-dimensional space (such as a map), or the position in a three-dimensional space. In addition to this position information, more detailed physical states such as orientation and speed may be further estimated.

[0019] [Overall Configuration of Wireless Communication System] FIG. 1 is a diagram showing the overall configuration of the wireless communication system according to the present embodiment.

[0020] The wireless communication system includes a position estimation device 1 and wireless terminals 2-1 to 2-Q. Here, Q is an integer of 1 or more.

[0021] The position estimation device 1 estimates the position of the wireless terminal by collecting the channel information of the wireless signals transmitted from the wireless terminals 2-1 to 2-Q.

[0022] The wireless terminals 2-1 to 2-Q include wireless communication units 2-1-1 to 2-Q-1. The wireless communication units 2-1-1 to 2-Q-1 each transmit a pilot signal known in transmission and reception, or a wireless signal including channel information between any wireless communication unit. The any wireless communication unit is a wireless communication unit 1-1 to 1-R provided in the position estimation device 1 or another wireless communication unit.

[0023] The position estimation device 1 receives the above wireless signals via the wireless communication units 1-1 to 1-R. The position estimation device 1 acquires the 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 from the received wireless signals.

[0024] Then, the position estimation device 1 inputs the acquired channel information to the input feature quantity generation unit 1-2. The input feature quantity generation unit 1-2 converts the channel information into input feature quantities suitable for input to the position estimation model, and inputs the converted input feature quantities to the position estimation model utilization unit 1-3.

[0025] After that, the position estimation model utilization unit 1-3 inputs the input feature amount of the channel information collected from the wireless terminals 2-1 to 2-Q into the position estimation model that models the relationship between the position information of the wireless terminal and the input feature amount of the channel information by machine learning, thereby estimating the position of the wireless terminal 2-i (1 ≤ i ≤ Q) during movement or stop.

[0026] The position estimation device 1 is, for example, a device installed in a base station or a cloud installed in a main place. The position 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 a wireless signal transmitted from the wireless terminal 2-i. The position estimation device 1 does not necessarily need to have a function of transmitting a wireless signal, and may be a device that only performs reception.

[0027] As shown in FIG. 1, the position estimation device 1 includes, for example, wireless communication units 1-1 to 1-R, an input feature amount generation unit 1-2, a position estimation model utilization unit 1-3, and a position estimation model training unit 1-4.

[0028] 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 correspond to a plurality of frequencies, a plurality of frequency bands, or a plurality of wireless communication systems. In FIG. 1, the wireless communication units 1-1 to 1-R are configured to receive wireless signals transmitted from the wireless terminal 2-i.

[0029] Any one 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 an arbitrary base station not shown in FIG. 1 and receive a wireless signal transmitted by itself to the arbitrary base station. Further, as will be described later, the wireless terminal 2-i can also handle the case where the wireless terminal 2-i moves while being connected to a plurality of base stations by switching the position estimation model based on the ID of the base station to which the wireless terminal 2-i is communicatively connected. The receiving-side wireless communication units 1-1 to 1-R may collect the channel information of the plurality of wireless communication units 2-i-1 by transmitting wireless signals on channels of the same frequency.

[0030] Each of the wireless communication units 1-1 to 1-R receives a wireless signal, and from the received wireless signal, acquires 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 the position estimation device 1 and a wireless communication device other than the wireless terminal 2-i. When the wireless communication units 1-1 to 1-R acquire the channel information corresponding to the wireless terminal 2-i, they input the acquired channel information to the input feature quantity generation unit 1-2.

[0031] The input feature quantity generation unit 1-2 has a function of converting the input channel information into input feature quantities that can be input to the position estimation model and storing them. The input feature quantity generation unit 1-2 inputs a plurality of input feature quantities corresponding to a plurality of times, which are the channel information after conversion, to the position estimation model utilization unit 1-3.

[0032] 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 a plurality of input feature quantities corresponding to a plurality of times to the position estimation model, and outputting the estimated position information. The position estimation model utilization unit 1-3 may use a pre-generated position estimation model, or may use a position estimation model generated and updated by the position estimation model training unit 1-4.

[0033] The position estimation model is a model generated by training the relationship between the position information of the wireless terminal 2-i and the channel information (≈ input feature quantity) regarding radio wave propagation acquired from the wireless terminal 2-i by machine learning, and is a model that outputs position information using the input feature quantity. In addition, the position estimation model may be generated by using digital twin technology or the like to generate a space equivalent to the real space in the simulation space and using the relationship between the virtually generated wireless terminal and the channel information calculated by simulation. The position estimation model may use a position estimation model created from the relationship between the channel information measured by another position estimation unit and the wireless terminal.

[0034] The position estimation model training unit 1-4 has a function of generating a position estimation model and a function of updating the position estimation model.

[0035] For example, the position estimation model training unit 1-4 separately acquires data related to the position information of the wireless terminal 2-i, and trains 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, thereby generating or updating the position estimation model.

[0036] Alternatively, instead of separately acquiring data related to the position information, the position estimation model training unit 1-4 is located at a predetermined known position, and the wireless terminal 2-i performs a predetermined known operation. Then, the position estimation model may be generated or updated from the input feature amount generated from the channel information obtained at that time.

[0037] The position information of the wireless terminal 2-i to be separately acquired, any means for acquiring the position measurement data can be realized, for example, by using sensors, cameras, wireless positioning, SLAM (Simultaneous Localization And Mapping), GPS (Global Positioning System) installed in the wireless terminal 2-i, and sensors, cameras, etc. installed in the environment.

[0038] The position measurement data is, for example, the position information and time information of the wireless terminal 2-i. After being stored in the wireless terminal 2-i, it is input to the position estimation model training unit 1-4 or stored in the position estimation model training unit 1-4 for use. Using the input feature amount corresponding to the same time as the position measurement data generated in the input feature amount generation unit 1-2 and the training data composed of the position measurement data, the position estimation model is updated or generated in the position estimation model training unit 1-4.

[0039] The position estimation model training unit 1-4 can train the position estimation model by learning the relationship between the stored position and time information of the wireless terminal 2-i and the input feature amounts and time information also stored in the storage unit, using training data that arranges both on the same time axis. If the training data forms diverse combinations in a short time, the performance of the position estimation model can be improved. The method for generating the training data will be described later.

[0040] That is, the position estimation model training unit 1-4 generates a training data set (a training data set including a plurality of training data) by aligning a plurality of position information of the wireless terminal 2-i corresponding to a plurality of times obtained by a method other than the method (the method of the present invention) estimated and calculated by the position estimation model utilization unit 1-3 and a plurality of input feature amounts corresponding to the plurality of times at the same time, and uses the training data set to generate or update the position estimation model so as to improve the accuracy of the relationship between the input feature amounts and the position information, and has a function of updating the weights, biases, etc. of the position estimation model.

[0041] In addition, the position estimation model training unit 1-4 has a function of generating or updating a training data set using the input feature amounts related to the 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 position or at the time when it is detected that the wireless terminal 2-i is performing a predetermined known operation, and the predetermined known position information of the wireless terminal 2-i.

[0042] In addition, the position estimation model training unit 1-4 has a function of generating or updating the position estimation model using a plurality of training data sets obtained by extracting training data at different time intervals from among the plurality of training data included in the training data set, and updating the weights, biases, etc. of the position estimation model.

[0043] [Operation of Wireless Communication System] [Position Estimation Method of Wireless Terminal (First Example)] FIG. 2 is a diagram showing a position estimation flow (first example) for estimating the position of a wireless terminal.

[0044] 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).

[0045] 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).

[0046] Next, the input feature quantity generation unit 1-2 converts the acquired channel information regarding radio wave propagation into input feature quantities suitable for input to the position estimation model, which are a plurality of input feature quantities corresponding to a plurality of times, and stores them (step S1-3).

[0047] The input feature quantity is, for example, channel information that is a coefficient of radio wave propagation between antennas, or a numerical value obtained by performing various operations on the channel information. Received power, signal power, signal-to-noise power ratio, signal-to-noise interference power ratio, coefficients related to a channel matrix having radio wave propagation coefficients between a plurality of transmitting and receiving antennas as elements, channel vectors that are column or row vectors of the channel matrix, and those obtained by dividing them by a certain value. A correlation matrix obtained by multiplying a channel matrix by a Hermitian matrix of the channel matrix, a channel matrix, an absolute value of a channel vector, a Frobenius norm, etc., which are normalized to be constant. Those obtained by converting a correlation matrix, a channel matrix, or a channel vector into angle information. Those obtained by dividing the absolute value or Frobenius norm of a correlation matrix, a channel matrix, or a channel vector by a certain value. Those obtained by converting a correlation matrix, a channel matrix, or a channel vector into decibels and subtracting a certain value. Also, using a moving average, an average value of values obtained from power information, power ratio information, and channel information in a recent certain period is acquired, and the value is normalized with that value. Also, at least one time series information among those information can be used. The specific calculation method of the input feature quantity will be described later.

[0048] Further, the position estimation device 1 may separately generate auxiliary information other than the channel information input to the position estimation model, and add the generated auxiliary information to the input feature amount (step S1-4). The auxiliary information is, for example, specific information for identifying the wireless terminal 2-i, the mounted object / owner information of the wireless terminal 2-1, temperature / position information of a camera / sensor, etc., state information such as position / velocity / settings of the device to be estimated and its surrounding reflection structures, air temperature, humidity, congestion situation, and setting information of the wireless communication system.

[0049] Finally, the position estimation model utilization unit 1-3 inputs a plurality of input feature amounts corresponding to a plurality of times, which are the converted channel information, to the position estimation model, thereby estimating and outputting the position information of the wireless terminal 2-i (step S1-5).

[0050] [Training method of position estimation model (first example)] FIG. 3 is a diagram showing a training flow (first example) for training the position estimation model.

[0051] First, the wireless terminal 2-i transmits a wireless signal capable of acquiring channel information while being located at a known position at a specific time or while performing a known operation pattern (step S2-1).

[0052] 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 detected time (step S2-2). Here, the wireless terminal 2-i may be all the wireless terminals 2-1 to 2-Q that are the targets of position estimation, or one or more wireless terminals selected from among them.

[0053] Next, the input feature amount generation unit 1-2 generates an input feature amount from the acquired channel information (step S2-3).

[0054] At this time, the position estimation device 1 may separately generate auxiliary information other than the channel information input to the position estimation model, and add the generated auxiliary information to the input feature amount (step S2-4).

[0055] Finally, the position estimation model training unit 1-4 generates a training data set using the generated input feature amounts and the position or movement of the wireless terminal 2-i at the time corresponding to the time of the input feature amounts, and generates or updates a position estimation model using the training data set (step S2-5).

[0056] [Method for training position estimation model (second example)] FIG. 4 is a diagram showing a training flow (second example) for training a position estimation model.

[0057] First, the position estimation model training unit 1-4 generates or acquires position information from another position estimation technique or measurement result that is not the present invention, and generates and acquires training data in which the position information corresponding to the same time is associated with the input feature amounts (step S3-1).

[0058] For example, in the wireless terminal 2-i such as an autonomous vehicle or a robot, a log regarding position information is stored and notified to the position estimation device 1 at a certain frequency, the input feature amounts corresponding to the time of the log are stored in the position estimation device 1, and by arranging both according to time, training data can be obtained.

[0059] At this time, the position estimation device 1 may separately generate auxiliary information other than the channel information input to the position estimation model, and add the generated auxiliary information to the input feature amounts to generate training data (step S3-2).

[0060] Finally, the position estimation model training unit 1-4 generates or updates a position estimation model using the obtained training data (step S3-3). For example, the position estimation model training unit 1-4 updates the weights or biases of the position estimation model so as to improve the accuracy of the relationship between the input feature amounts and the position information using the training data set.

[0061] [Method for expanding training data] FIG. 5 is a diagram showing an expansion flow for expanding training data. This expansion method is effective when the position estimation model uses time series data including a plurality of time information as input features.

[0062] First, the position estimation model training unit 1-4 acquires training data for updating or generating the position estimation model (step S4-1). The training data has position information and input features with respect to time.

[0063] Next, the position estimation model training unit 1-4 increases the number of valid data by changing the time series arrangement of the acquired training data or arbitrarily setting a plurality of training data to be used (step S4-2). For example, the position estimation model training unit 1-4 generates a plurality of training data sets in which training data is extracted from a plurality of training data at different extraction frequencies in the time axis direction. By doing so, the number of valid data can be increased and the accuracy of position estimation can be further improved.

[0064] At this time, the position estimation device 1 may separately generate auxiliary information other than the channel information input to the position estimation model, add the generated auxiliary information to the input features, and generate training data (step S4-3).

[0065] Finally, the position estimation model training unit 1-4 generates or updates the position estimation model using the generated training data (step S4-4). For example, the position estimation model training unit 1-4 updates the weights or biases of the position estimation model.

[0066] Note that the position estimation model training unit 1-4 determines the order of using a plurality of training data sets according to the acquisition frequency of the input features input to the position estimation model or the speed of the position detection target to be targeted, and uses the plurality of training data sets in the determined order. The position estimation model may be generated or updated.

[0067] Generally, the entire time-series training data used for training is divided into batch data, and the process of updating the coefficients and weights of the entire model from the output error is performed for all the data as one epoch.

[0068] From among the plurality of training data sets generated according to this embodiment, those with a large deviation from the frequency of the input information obtained from the position estimation target to be estimated are selected in order, and model updates are performed using the plurality of training data sets in the selected order. Finally, model update for one epoch is completed using the time-series training data with the smallest deviation from the frequency at the time of actual position prediction, and this process is repeated multiple times to improve the position estimation accuracy of the position estimation model.

[0069] In this way, by increasing the amount of effective data per epoch and making the conditions for re-model update closest to the time-series conditions at the time of use, it is possible to improve the position estimation performance of the position estimation model obtained for each epoch.

[0070] In particular, if the speed conditions of the position estimation target are the same during training and inference, the time-series training data with the time interval closest to the acquisition interval of the input feature amount during inference is used last. When the expected speed of the position estimation target at the time of performing position estimation is V times the speed at the time of training data generation, if the acquisition time interval of the input feature amount during inference is Tp and the time interval of the time series acquired during training is Tt, the time-series training data of the time interval that is assumed to move the distance closest to the distance that the position estimation target moves at the time interval Tp during inference may be used last. Here, when three time-series training data of time intervals Tt, 2Tt, and 3Tt are generated from the training data generated at the interval Tt, training for each epoch can be performed so as to use the time-series training data with a value close to VTt last.

[0071] FIG. 6 is a diagram showing an example of expanding training data (Example 1). The training data is such that the input feature amounts In-1 to In-L and the position information (X, Y) of the wireless terminal to be estimated are arranged with respect to time. The acquisition times do not have to match exactly. In this example, it is in a format organized with a time width of 100 ms. It may be temporally complemented, or missing values may be copied from the previous value in time. In addition, complementation techniques used in machine learning can be used.

[0072] In the example of FIG. 6, the training data originally obtained at intervals of 100 ms is re-extracted at a period of 200 ms. In the case of a 200 ms period, it can be seen that two patterns of training data sets can be obtained depending on the start point of re-extraction. If the total number of time steps of the training data included in the original training data set is D, the number of time steps in the case of a 200 ms period is D / 2, and two training data sets can be created. It can be understood that if both the 100 ms period and the 200 ms period are used as training data, it becomes D+(D / 2)×2, and the training data is doubled.

[0073] If the training data is reconfigured for different time widths such as a 300 ms period and a 400 ms period, the training data can be increased by the number of types of time widths. Resetting the time width of the time step to be longer means that it is simulating the case where the moving speed of the position estimation target is increased. If the channel information is not greatly affected by the moving speed of the wireless terminal, it means that pseudo data simulating various moving speeds of the wireless terminal is increased from the same training data set.

[0074] FIG. 7 is a diagram showing an example of expanding training data (Example 2). Training data in a pattern where the training data is randomly removed is generated. Different from the case where the time width is simply set to a certain multiple of the minimum time width, it will pseudo-generate more complex movements of the wireless terminal. In FIG. 6, the speed of the wireless terminal is equivalently increased, but with this method, a large number of training data can be generated without extremely increasing the speed.

[0075] FIG. 8 is a diagram showing an example of expanding training data (Example 3). It is a method of reversing the temporal order of the training data. Training data is generated by rearranging from the future to the past. By doing so, pseudo data equivalent to the case where the wireless terminal moves in the opposite direction to the actual movement can be generated. Also, by using together the training data sets shown in FIGS. 6 and 7 along with the training data set generated in FIG. 8, the training data can be further increased.

[0076] [Method for Estimating Position of Wireless Terminal (Second Example)] Next, a method for estimating the position when the movement distance of the wireless terminal 2-i is long, such as moving through the communication areas of a plurality of base stations, will be described.

[0077] When arranging the wireless communication units 1-1 to 1-R in a wide area and generating input feature amounts from the channel information estimated by the pilot signal, a plurality of position estimation models are prepared corresponding to the wireless communication units 1-1 to 1-R, or a plurality of position estimation models are prepared for combinations of the wireless communication units 1-1 to 1-R, and by inputting to an appropriate position estimation model, the position of the wireless terminal 2-i is estimated.

[0078] Also, when generating input feature amounts by feedback of the channel information included in the wireless signal from the wireless terminal 2-i, since the content of the channel information does not change regardless of which wireless communication unit receives it, it is necessary to determine from which base station the wireless terminal 2-i has fed back the wireless signal. The flow in this case is shown in FIG. 9.

[0079] First, the wireless terminal 2-i communicates with an appropriate base station among the installed plurality of base stations and transmits a wireless signal including channel information (step S5-1).

[0080] Next, the wireless communication unit 1-1 of the position estimation device 1 receives the wireless signal and acquires the channel information and the ID of the connected base station (step S5-2).

[0081] Next, the position estimation model utilization unit 1-3 selects a position estimation model corresponding to the base station ID from among a plurality of position estimation models (step S5-3).

[0082] Next, the input feature amount generation unit 1-2 converts the channel information into an input feature amount (step S5-4).

[0083] At this time, the position estimation device 1 may separately 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 feature amount (step S5-5).

[0084] Finally, the position estimation model utilization unit 1-3 inputs the input feature amount into the selected position estimation model, and outputs the position of the wireless terminal 2-i (step S5-6).

[0085] [Channel Information Collection Method, Input Feature Amount Calculation Method] Examples of the channel information collection method and the input feature amount calculation method will be described below.

[0086] [Channel Information Collection Method, Input Feature Amount Calculation Method (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 the channel matrix between the antenna of its own wireless communication unit 1-r (number of receiving antennas: Mr) and the antenna of the wireless communication unit 2-i-1 that transmitted the pilot signal (number of transmitting antennas: Ni). In the case of OFDM (orthogonal frequency division multiplexing method) used in various wireless communication systems, it is possible to obtain the channel matrix of subcarriers corresponding to a plurality of frequencies.

[0087] From the channel matrix H of "number of transmitting antennas Ni × number of receiving antennas Mr" thus obtained, an input feature amount to be input to the position estimation model utilization unit 1-3 is generated. For example, when obtaining a channel matrix for a plurality of subcarriers by OFDM, the channel matrix of the η-th subcarrier is H ηis defined as. And as a method for converting into input feature quantities, first, as in Equation (1), the channel matrix H η is separated into a normalized channel matrix G η normalized with a predetermined norm, η and amplitude information γ η 2 or power information γ.

[0088]

Equation

[0089] G η For example, can be set so that ||G η || F = 1. ||·|| F represents the Frobenius norm. γ η Generally has a large amplitude and can have changes in values of the fifth power of 10 or more. Therefore, γ η or γ η 2 can be converted to dB, its maximum and minimum values can be defined, and the converted values expressed as normalized values within the range of 0 to 1 can be used. Values obtained by selecting or averaging a plurality of these for different frequency conditions and antenna conditions, γ all can also be used.

[0090] Alternatively, as in Equation (2), the amplitude information γ η is separated for each antenna, and each column vector g 1,η ~g Mr,η normalized to a certain value for the norm, 1,η ~γ Mr,η and its amplitude value γ

[0091]

Equation

[0092] g a,η For example, can be set as a defined vector so that ||g a,η || F = 1. γa,η and γ a,η 2 It is also possible to convert them to dB, define their maximum and minimum values, and use values obtained by converting them so as to be expressed as values normalized within a range of 0 to 1. For the γ corresponding to the a-th column vector obtained by selecting a plurality of these for η or averaging them, a,all it is also possible to use it.

[0093] Alternatively, as in Equation (3), the amplitude information γ η is separated for each antenna, and each row vector g’ 1,η ~g’ Ni,η whose norm value is normalized to a certain value, and its amplitude value γ’ 1,η ~γ’ Ni,η are obtained.

[0094]

Number

[0095] g’ b,η For example, g’ b,η || F can be set as a defined vector such that ||g’ b,η and γ’ b,η 2 are converted to dB, define their maximum and minimum values, and use values obtained by converting them so as to be expressed as values normalized within a range of 0 to 1. For the γ’ corresponding to the b-th column vector obtained by selecting a plurality of these for η or averaging them, b,all it is also possible to use it.

[0096] The channel matrix H η the normalized channel matrix G η the normalized vector g a,η the normalized vector g’ b,η can have the real and imaginary parts of each element as input feature amounts, can have the input information as imaginary numbers as they are, can be converted to another format such as angular information, or can be quantized.

[0097] Also, as described above, without obtaining a channel matrix such as H from the received signal, simply when something is received, the power of the received signal can be obtained as γ η 2 This kind of 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, can also be used.

[0098] Also, since the power information varies with time due to factors such as the humidity and temperature of the radio wave propagation environment, the temperature of the terminal and the base station, etc., it is also effective to use moving average or comparison information for the amplitude value information. That is, if the values such as the amplitude information, power information, and RSSI obtained at a certain time are denoted as γ[t], as the power ratio information ω[t], a calculated value as shown in Equation (4) may be used.

[0099]

Equation

[0100] The obtained value may be used as it is, or it may be in dB units.

[0101] Also, the correlation matrix H η generated using the channel matrix H η H η H 、H η H H η can be used. The correlation matrix G η generated using the normalized channel matrix G η G η H 、G η H G η can be used. The channel matrix H η 、the normalized channel matrix G η 、their correlation matrices Hη H η H 、H η H H η 、G η G η H 、G η H G η For a plurality of frequencies, matrices ΣH η and ΣG η can be used, where ΣH is obtained by taking the sum or average of matrices related to H for multiple frequencies. Similarly, for ΣG related to G. Also, eigenvalues, diagonal matrices, unitary matrices, etc. obtained by performing QR decomposition, SVD (Singular value decomposition), eigenvector decomposition, etc. of these matrices can be used. η H η H 、ΣH η H H η 、ΣG η G η H 、ΣG η H G η can be used. Eigenvalues, diagonal matrices, unitary matrices, etc. obtained by performing QR decomposition, SVD (Singular value decomposition), eigenvector decomposition, etc. of these matrices can be used.

[0102] Furthermore, using the above 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 The power characteristics with respect to the direction of arrival for the communication device obtained by the direction-of-arrival estimation technology can be used as input feature quantities. For example, for each vector component, values obtained by multiplying by (1, exp(jdθ), exp(j2dθ),..., exp(jNdθ)) can be calculated with respect to θ. θ from 0 to 2π can be generated at arbitrary angular intervals, and the outputs for a plurality of θ can also be used as input feature quantities. d is a predetermined constant. N is the number of elements of the vector.

[0103] That is, the input feature quantity generation unit 1-2 generates, as input feature quantities, the received power of a radio signal, the signal power, power ratio information obtained from the moving average of the received power and the signal power, a channel matrix composed of radio wave propagation coefficients between a plurality of antennas, a correlation matrix or a pseudo-correlation matrix of the channel matrix, an operation matrix obtained by signal-processing the channel matrix, an operation matrix obtained by signal-processing the correlation matrix or the pseudo-correlation matrix, an operation matrix obtained by signal-processing a channel matrix, a correlation matrix, or a pseudo-correlation matrix corresponding to a plurality of frequencies, a unitary matrix obtained by linearly operating on the channel matrix, a unitary matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a unitary matrix obtained by linearly operating on the operation matrix, a diagonal matrix obtained by linearly operating on the channel matrix, a diagonal matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a diagonal matrix obtained by linearly operating on the operation matrix, a triangular matrix obtained by linearly operating on the channel matrix, a triangular matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a triangular matrix obtained by linearly operating on the operation matrix, and values obtained by normalizing the phase, amplitude, real component, imaginary component values of one or more of these feature quantities, and the range of coefficients of these one or more values. The input feature quantity generation unit 1-2 stores such input feature quantities as time-series data and outputs the input feature quantities corresponding to a plurality of times from the past to the present to the position estimation model.

[0104] In a wireless communication system using an equalization technique, it is possible to estimate the arrival time, power, and phase conditions of the paths of incoming electrical signals. Even for such channel information obtained in time series, it can be used as an input feature quantity of the position estimation model by normalizing the power, converting it into frequency components, and using feature quantities and angle information extracted by existing arrival wave direction techniques.

[0105] [Channel Information Collection Method, Input Feature Quantity Calculation Method (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 and received from the specific wireless base station, quantizes it in some form to generate feedback information, 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 acquires the channel information between the specific wireless base station included in the received wireless signal and the wireless terminal 2-i.

[0106] First, a pilot signal known for transmission and reception is transmitted from a specific known wireless base station. In advance, by sending a known pattern, the wireless communication unit 2-i-1 of the wireless terminal 2-i can acquire the channel matrix between its own receiving antenna (number of receiving antennas: Ni) and the antenna of the specific wireless base station that transmitted the pilot signal (number of transmitting antennas: Mt). In the case of OFDM used in various wireless communication systems, the channel matrix of subcarriers corresponding to a plurality of frequencies can be obtained.

[0107] The channel matrix H of "number of transmitting antennas Mt × number of receiving antennas Ni" thus obtained α is used to generate an input feature amount to be input to the position estimation model utilization unit 1-3. For example, when obtaining a channel matrix for a plurality of subcarriers by OFDM, the channel matrix of the η-th subcarrier is defined as H α,η Here, when the 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. Also, in this case, the channel matrix H α,η corresponds to the transposed matrix of the channel matrix H η

[0108] Therefore, as a method for converting into an input feature amount, similar to the first method, as shown in Equation (5), the channel matrix H α,η is separated into a normalized channel matrix G α,η normalized with a predetermined norm, and amplitude information γ α,η or power information γ η 2 ​

[0109] [Number]

[0110] G α,η For example, ||G α,η || F can be set so that ||G|| = 1. ||·|| F represents the Frobenius norm. γ α,η Or γ α,η 2 After converting γ and γ to dB, defining their maximum and minimum values, and converting them to values expressed as normalized values within the range of 0 to 1, values obtained by such conversion may be used. For the values γ obtained by selecting or averaging multiple of these, α,all γ may be used. α,all When averaging γ, it may be averaged with the true value, or averaged after converting to dB, or the result of averaging with the true value may be converted to dB units.

[0111] Alternatively, as in Equation (6), the amplitude information γ α,η is separated for each antenna, and each column vector g α,1,η ~g α,Ni,η with the norm value normalized to a certain value, and its amplitude value γ α,1,η ~γ α,Ni,η may be obtained.

[0112] [Number]

[0113] g α,a,η For example, ||g α,a,η || F can be set as a defined vector so that ||g|| = 1. γ α,a,η Or γ α,a,η 2 After converting γ and γ to dB, defining their maximum and minimum values, and converting them to values expressed as normalized values within the range of 0 to 1, values obtained by such conversion may be used. For the value γ corresponding to the a-th column vector obtained by selecting or averaging multiple of these with respect to η, α,a,allIt may also be used.

[0114] Alternatively, as in Equation (7), the amplitude information γ α,η is separated for each antenna, and each row vector g’ α,1,η ~g’ α,Mt,η where the norm value is normalized to a certain value, and its amplitude value γ’ α,1,η ~γ’ α,Mt,η may be obtained.

[0115]

Number

[0116] g’ α,b,η can be set as a defined vector such that, for example, ||g’ α,b,η || F = 1. γ’ α,b,η and γ’ α,b,η 2 are converted to dB, the maximum and minimum values are defined, and the converted values expressed as normalized values within the range of 0 to 1 may be used. Values corresponding to the b-th column vector obtained by selecting or averaging a plurality of these for η, γ’ α,b,all may also be used.

[0117] The channel matrix H α,η , the normalized channel matrix G α,η , the normalized vector g α,a,η , and the normalized vector g’ α,b,η can have the real and imaginary parts of each element as input feature amounts, remain as imaginary numbers as input information, be converted to another format such as angle information, or be quantized.

[0118] Also, the correlation matrix H α,η generated using the channel matrix H α,η H α,η H 、H α,η H H α,η can be used. The correlation matrix G α,η generated using the normalized channel matrix G α,η G α,ηH , G α,η H G α,η can be used. The channel matrix H α,η , the normalized channel matrix G α,η , their correlation matrices H α,η H α,η H , H α,η H H α,η , G α,η G α,η H , G α,η H G α,η are used to calculate the sum or average of the matrices ΣH α,η , ΣG α,η , ΣH α,η H α,η H , ΣH α,η H H α,η , ΣG α,η G α,η H , ΣG α,η H G α,η can be used. Eigenvalues, diagonal matrices, unitary matrices obtained by performing QR decomposition, SVD, eigenvector decomposition, etc. on these matrices can be used.

[0119] α,η As an example, the case of using the channel information feedback used in the wireless LAN standards IEEE 802.11n / ac / ax is shown. This is the case where the right singular matrix obtained by the SVD of the channel matrix H

[0120] [Number]

[0121] Here, U η is the left singular matrix of Ni×Ni, and Σ η is the eigenvalue λ 1,η ~λ Ni,ηAn Ni×Ni diagonal matrix with diagonal elements, 0 is a matrix with elements of Ni×(Mt - Ni) zeros, (V S,η V N,η ) H is an Mt×Mt right singular matrix, V S,η is a vector corresponding to the eigenvalue, and V N,η corresponds to the zero 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 for spatial multiplexing. Hereinafter, it is described assuming that feedback is performed for the number of receiving antennas Ni, but it may be any number less than or equal to Ni. What is feedback is Ni vectors V η among the right singular matrices, and the value of SNR obtained from the eigenvalues. SNR is not for each subcarrier but is averaged, for example, λ 1,η ~λ Ni,η are averaged over the generated subcarriers, and the value obtained by dividing λ 1, ~λ Ni by the thermal noise level, S1~S Ni can be fed back. V η and, as in Equation (9), can be converted into angle information (φ, ψ) and fed back.

[0122]

Equation

[0123] This expression is written focusing on a certain frequency. The V matrix in Equation (9) exists for the specified number of subcarriers, and angle information is generated for each subcarrier. Further, it is quantized with the 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 further can obtain the RSSI information of the wireless signal.

[0124] The angle information may be used directly as the input feature amount. The sine and cosine components calculated from the angle information may be used as the input feature amount. The matrix obtained by returning the angle information to the right singular matrix using Equation (9) may be used. After returning the angle information to the 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 applying signal processing such as QR decomposition to the averaging matrix may be used. For example, when returning to the correlation matrix, V corresponding to the η-th subcarrier η is obtained, while the SNR is only a representative value for the whole. Therefore, the pseudo-correlation matrix obtained as an approximation is calculated, for example, as shown in Equation (10).

[0125] [Number]

[0126] Here, Θ represents the group of subcarriers for which V η is obtained, and N Θ is the number of elements of Θ. Since the lower triangular part of the non-diagonal elements and the upper triangular part of the correlation matrix obtained here are in a complex conjugate relationship with each other, when used as an input feature amount in machine learning, among the pseudo-correlation matrices R obtained here, the diagonal elements, and the lower left half or the upper right half of the non-diagonal elements can be used. Since the non-diagonal elements are imaginary numbers, by using the real part and the imaginary part as real numbers respectively as the input feature amount, a general machine learning algorithm can be applied.

[0127] Also, in Equation (10), even if the number of columns of the fed-back V η and the number of SNR (eigenvalues) change over time, in order for the correlation matrix of Equation (10) not to be affected, among V η , only the first vector and the first SNR are used to obtain the correlation matrix by Equation (10), or a number of vector numbers N Θ smaller than the currently obtained number of columns (in the case of Equation (10), N Θ ) of the columns and the number of SNR are selected, and the correlation matrix is obtained and used as the input feature amount. Also, at this time,

[0128] [Mathematics]

[0129] By calculating as such, the elements of R calculated can be in the range of 0 to 1 or -1 to 1, and the normalization process when inputting to deep learning can also be omitted.

[0130] Furthermore, as described above, when obtaining the correlation matrix by the method of selecting the first vector, since the first diagonal element is 1, simply

[0131] [Mathematics]

[0132] the correlation matrix can be calculated as such. Here, V η,1 represents the first column vector of V η . To normalize the range of R as described above, the denominators of formulas (10) to (12) may be omitted and divided by the coefficient for normalization.

[0133] Also, when restoring the right eigenmatrix V from the angle information, the imaginary part of the last element of each column vector will always be 0. Therefore, assuming that the right singular matrix is obtained as a matrix of M×Ni, from the numerical values of the real part and the imaginary part of each element, the numerical value of 2×Mt×Ni - Ni becomes meaningful information. For example, when 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. When 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 may not be used.

[0134] [Experimental Results] The position estimation method and its effects of this embodiment will be described with specific examples and indoor experimental results.

[0135] In the indoor experimental environment area shown in FIG. 10, two wireless communication units 1-1 and 1-2 were installed to estimate the position of the wireless terminal 2-1. The wireless terminal 2-1 is running in an eight-shaped pattern in the center of the experimental environment area and is transmitting feedback of a wireless signal including channel information to the base station (AP: Access Point).

[0136] The wireless terminal 2-1 is required to report channel information every 100 ms from the base station AP, and is transmitting feedback of the angle information of the channel information by the channel information feedback method defined in the wireless LAN standardization specification IEEE 802.11ac. The number of antennas of the base station AP is four. The number of antennas of the wireless terminal 2-1 and the wireless communication units 1-1 and 1-2 is two each.

[0137] The two wireless communication units 1-1 and 1-2 can obtain the angle information, SNR(λ1,λ2) generated from the right singular matrix of the channel matrix between the wireless terminal 2-1 and the base station AP. In the wireless communication unit 1-1, the RSSI values (γ 11 ,γ 12 ) can also be obtained, and the wireless communication unit 1-2 can also obtain the RSSI values (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.

[0138] The input feature amount generation unit 1-2 calculates a pseudo-correlation matrix from the angle information according to the above formula, averages the calculated pseudo-correlation matrix in frequency, calculates the pseudo-correlation matrix R in Equation (10), divides R by the sum of the diagonal terms of the obtained R, and performs normalization. As the pseudo-correlation matrix R, a total of 16 input feature amounts were generated, including 4 diagonal terms and 12 real and imaginary parts of 6 elements in the lower left half of the non-diagonal terms. These are the pseudo-correlation matrix features (R1, …, R 16) and denote it as R. The dB values of the two SNRs (λ1, λ2) included in the feedback signal are divided by a value exceeding the value observable in the experimental environment, and the coefficient after normalization is denoted as SNR. Also, the RSSI (γ 11 , γ 12 ) whose dB value is normalized to be distributed in the range from 0 to 1 is denoted as RSSI, and RSSI uses the result obtained by using the wireless communication unit 1-1.

[0139] Here, a method for generating the position estimation model will be described.

[0140] In order to generate the position estimation model according to this embodiment, an autonomous mobile robot as the wireless terminal 2-1 was made to travel in the experimental environment area indoors for 4 days. The line remaining in the figure-eight shape shown in FIG. 10 is the actual path of the autonomous mobile robot. To complicate the operation, the robot heads to a goal point defined by a square of 60 cm on each side, but within this square, the actual target point is made random and is set to skip with a probability of 50%.

[0141] At this time, the position estimation model training unit 1-4 generates training data in which the high-precision position information of the autonomous mobile robot obtained from the LIDAR (light detection and ranging) and the tire control information provided in the wireless terminal 2-1 and the input feature amount related to the above-described propagation are arranged in the same time series at a cycle of 200 ms. As described above, the generation cycle of the input feature amount is 100 ms, but the training data was generated by selecting and using the latest information obtained within the time width divided by the 200 ms cycle.

[0142] Then, using the generated training data for 8 hours, a position estimation model based on a deep neural network using GRU (Gated Recurrent Unit) and direct connections was trained. The learning rate was 0.0002, and the optimization algorithm used was ADAM. The GRU had 1 hidden layer with a dimension of 400, and two direct connection layers with an input of 400 and an output of 400, and one direct connection layer with an input of 400 and an output of 2. The weights and biases of the position estimation model were updated by backpropagation so as to output the information of the X coordinate and the Y coordinate within the experimental environment area. Using the 30-minute data for validation, the updated position estimation model with the minimum MSE was used.

[0143] Figure 11 is a diagram plotting the amount of positional deviation as an error in a cumulative probability distribution (CDF: Cumulative Distribution Function) when estimating the position of the wireless terminal 2-1 using the position estimation model generated according to this embodiment.

[0144] The graph of "SNR / RSSI / R" is the error result of position estimation by the position estimation model generated by using a total of 20 (=16 + 2 + 2) input features of the aforementioned pseudo-correlation matrix R, SNR, and RSSI, and inputting 20 samples of each time-series data into an RNN (Recurrent Neural Network) for training.

[0145] The graph of "SNR / RSSI / R (immediately after training)" is the error result when using the position estimation model for the test data measured at a time different from the training data after running for 4 days, and a lower position estimation error than "SNR / RSSI / R" can be confirmed.

[0146] The other error results (graphs of "R", "SNR / RSSI", "single-time SNR / RSSI") are the cumulative probability distributions of the errors when estimating the position by measuring the same features 30 days later using the position estimation model generated using the 4-day training data.

[0147] First, regarding the result using the total of 20 input features of "SNR / RSSI / R", although the positioning error increases compared to the characteristics used immediately after training, with the error that was 7.7 cm at the median increasing to 12.7 cm, it can still be seen that a high-precision positioning effect with a positioning error of about 10 cm is obtained. This is significantly higher than the estimation result of "SNR / RSSI" in which a total of 4 input features of SNR and RSSI were trained using 20 samples in time series and applied to the training data 30 days later. "Single-time SNR / RSSI" that uses only a single sample in the time direction without using time series data has an even more significantly larger positioning error. Therefore, it can be confirmed that a high positioning accuracy can be achieved with the pseudo-correlation matrix R.

[0148] Figure 12 is a diagram plotting the cumulative probability distributions of the estimation errors of "SNR / RSSI / R", "R", and "SNR / RSSI" for test data with a different time from the training within 4 days, data 34 days later, and data 6 days later, after increasing the dimensions of the aforementioned deep learning positioning model, setting the dimension of GRU to 2000, the first two fully connected layers to input 2000, output 2000, and the last layer to input 2000, output 2, and training with the data for 4 days in the same manner as the aforementioned case.

[0149] With this result, the result of the pseudo-correlation matrix R is further improved and is approaching the estimation error of "SNR / RSSI / R". Although the characteristics of "SNR / RSSI" are greatly improved when used immediately after training, it can be confirmed that the positioning accuracy decreases as one month and two months pass.

[0150] Figures 13 and 14 are diagrams showing graphs verifying the aforementioned training method. Here, considering training the positioning model for the following cases using the training data with the feedback signal obtained at 100 ms intervals, with the dimension of GRU set to 500, the first two fully connected layers to input 500, output 500, and the last layer to input 500, output 2 for evaluation.

[0151] Case 1 is a case where a position estimation model for estimating the position of a robot is trained by inputting input feature amounts every 100 ms for 40 samples in time series into a deep learning position estimation model based on a feedback signal obtained every 100 ms.

[0152] Case 2 is a case where a position estimation model for estimating the position of a robot is trained by inputting input feature amounts every 200 ms for 40 samples in time series into a deep learning position estimation model based on a feedback signal obtained every 200 ms.

[0153] In FIG. 13, when training the position estimation model, the case where only the original training data obtained every 100 ms is used is described as "data only at 100 ms intervals".

[0154] Also, using the data expansion method shown in FIG. 6, a data set at 100 ms intervals is reconfigured into two data series at 200 ms intervals. When training, the data of the two data series are divided into batch data, the output error for all batch data is evaluated as one epoch, and the coefficients and weights of the entire position estimation model are updated. Also, the model is updated from the output for each batch data from the training data at 200 ms intervals. After considering all the data, batch data is further generated from the training data at 100 ms intervals, and all the data therein is considered to update the position estimation model as one epoch. The estimation error in this case is described as "data at 200 and 100 ms intervals".

[0155] Furthermore, "data at 400, 300, 200, 100 ms intervals" is the result of repeating training for one epoch in order using four data series at 400 ms intervals, three data at 300 ms intervals, two data at 200 ms intervals, and data at 100 ms intervals. When applied using the 200 ms training data generated by the data expansion of the present invention, the estimation error is reduced by 7%.

[0156] That is, in FIG. 13, after the position estimation model training unit 1-4 updates the position estimation model using a plurality of training data sets such as "data at intervals of 400, 300, and 200 ms", it finally updates the position estimation model using the original training data set, which is the "data at intervals of 100 ms" generated in advance. As a result, as described above, the position estimation error can be further reduced.

[0157] In FIG. 14, in the position estimation model that estimates the position using the input feature amounts in a time series at intervals of 200 ms, the effect of the data augmentation method in FIG. 6 is similarly verified. First, for the training data of "only data at intervals of 200 ms", as in the data augmentation method shown in FIG. 6, two pieces of training data at intervals of 200 ms are extracted from the training data existing at intervals of 100 ms, and training data equivalent to the training data at intervals of 100 ms is used for learning. Comparing FIG. 14 with FIG. 13, it can also be confirmed that the position estimation performance is almost the same as that in FIG. 13 even though the usage frequency of the data used is halved.

[0158] Also, the case where three pieces of data at intervals of 300 ms are generated from the original training data at intervals of 100 ms, the position estimation model is trained with the training data at intervals of 300 ms, and then the position estimation model is trained with the training data at intervals of 200 ms is shown as "data at intervals of 300 and 200 ms". Similar to the case of "only data at intervals of 100 ms" shown in FIG. 13, it can be confirmed that the position estimation accuracy is improved and the error is reduced.

[0159] Furthermore, for the results of "data at intervals of 400, 300, and 200 ms" by the position estimation model obtained by training using the data at intervals of 400, 300, and 200 ms in this order, similarly high effects are obtained. In this way, by expanding the limited training data through data augmentation, the position estimation accuracy of the position estimation model can be improved.

[0160] That is, in FIG. 14, the position estimation model training unit 1-4 updates the position estimation model using, among a plurality of training data sets such as "data at intervals of 400, 300, and 200 ms", the training data set closer to the original training data set which is the "data at intervals of 100 ms" whose time interval of the training data was generated in advance, later. Thereby, as described above, the position estimation error can be further reduced.

[0161] To confirm the effect of the pseudo-correlation matrix, instead of the 16 feature amounts of the pseudo-correlation matrix R, the V matrix restored from the feedback information is averaged by the number of sub-carriers, and FIG. 15 shows the difference in position estimation accuracy when using the real part and the imaginary part information V of the resulting 4×2 matrix. From FIG. 15, it can be confirmed that the position estimation performance is higher when using the pseudo-correlation matrix R, SNR, and RSSI than when averaging the V matrix.

[0162] FIG. 16 shows, in the experimental environment area shown in FIG. 10, the value of RSSI (γ 11 ,γ 12 ) in the wireless communication unit 1-1, and two values of RSSI (γ 21 ,γ 22 ) in the wireless communication unit 1-2. Using these, the dimension of GRU is set to 400, the first two layers of the fully connected layer have an input of 400, an output of 400, and the last layer has an input of 400 and an output of 2. Training is performed with data for 4 days in the same manner as in the previous case, and it is a figure comparing the estimation accuracy when using only the value of RSSI (γ 11 ,γ 12 ) in the wireless communication unit 1-1. The effect of adding the wireless communication unit 1-2 is not large, but the error of 6.48 cm at the median is improved to 5.97 cm.

[0163] [Effects of the present embodiment] According to this embodiment, the position estimation device 1 includes radio communication units 1-1 to 1-R that receive radio signals transmitted from the radio communication unit of a wireless terminal and acquire channel information related to radio wave propagation from the radio signals, an input feature quantity generation unit 1-2 that converts the channel information into input feature quantities that can be input to a position estimation model, stores them, and outputs a plurality of input feature quantities 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 quantities to a position estimation model in which the relationship between the channel information related to radio wave propagation and the position information of the wireless terminal is modeled by machine learning, thereby estimating and calculating the position of the wireless terminal. That is, since the position of the wireless terminal is estimated using the channel information related to radio wave propagation included in the radio signal of the wireless terminal and the position estimation model, it becomes possible to perform the position estimation of the wireless terminal by general-purpose wireless communication, and a technique capable of estimating the position of the wireless terminal at low cost and with high accuracy can be provided.

[0164] [Others] The present invention is not limited to the above embodiment. The present invention can be variously modified within the scope of the gist of the present invention.

[0165] The position estimation device 1 described above in this embodiment can be realized, for example, by 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 as shown in FIG. 16. The memory 902 and the storage 903 are storage devices. In the computer system, each function of the position estimation device 1 is realized by the CPU 901 executing a predetermined program loaded on the memory 902.

[0166] The position estimation device 1 may be implemented by one computer. The position estimation device 1 may be implemented by a plurality of computers. The position estimation device 1 may be a virtual machine implemented on a computer. The program for the position estimation device 1 can be stored in a computer-readable recording medium such as an HDD, an SSD, a USB memory, a CD, or a DVD. The program for the position estimation device 1 can also be distributed via a communication network.

Explanation of Symbols

[0167] 1: Position estimation device 1-1~1-R: Wireless communication unit 1-2: Input feature quantity generation unit 1-3: Position estimation model utilization unit 1-4: Position estimation model training unit 2-1~2-Q: Wireless terminal 2-1-1~2-Q-1: Wireless communication unit

Claims

1. A wireless communication unit that receives a wireless signal transmitted from a wireless communication unit of a wireless terminal and obtains channel information regarding radio wave propagation from the wireless signal; An input feature amount generation unit that converts the channel information into an input feature amount that can be input to a position estimation model, stores the converted information, and outputs a plurality of input feature amounts corresponding to a plurality of times to the position estimation model; A position estimation model utilization unit that estimates and calculates the position of the wireless terminal by inputting the plurality of input feature amounts into a position estimation model that models the relationship between the channel information regarding radio wave propagation and the position information of the wireless terminal by machine learning; A position estimation model training unit that updates the position estimation model, and The position estimation model training unit generates a training data set in which a plurality of position information of the wireless terminal corresponding to a plurality of times obtained by a method other than the estimation calculation method and a plurality of input feature amounts corresponding to the plurality of times are aligned at the same time, and uses the training data set to update the position estimation model so as to improve the accuracy of the relationship between the input feature amount and the position information.

2. The position estimation model training unit uses the input feature amount related to the channel information acquired from the wireless terminal at the time when it is detected that the wireless terminal is located at a predetermined known position or at the time when it is detected that the wireless terminal is performing a predetermined known operation, and the predetermined known position information of the wireless terminal to generate the training data set. The position estimation device according to claim 1.

3. The position estimation model training unit updates the position estimation model using a plurality of training data sets obtained by extracting training data at different time intervals from the training data of the training data set. The position estimation device according to claim 1 or 2.

4. The position estimation model training unit after updating the position estimation model using the plurality of extracted training data sets, updates the position estimation model using the finally generated training data set, or among the plurality of extracted training data sets, the closer the time interval of the training data is to the generated training data set, the later the training data set is used to update the position estimation model. The position estimation device according to claim 3.

5. The input feature amount generation unit As the input feature quantity, power ratio information obtained from the received power, signal power, moving average of the received power or signal power of the wireless signal, a channel matrix composed of radio wave propagation coefficients between a plurality of antennas, a correlation matrix or a pseudo-correlation matrix of the channel matrix, an operation matrix obtained by signal-processing the channel matrix, an operation matrix obtained by signal-processing the correlation matrix or the pseudo-correlation matrix, an operation matrix obtained by signal-processing the channel matrix, the correlation matrix or the pseudo-correlation matrix corresponding to a plurality of frequencies, a unitary matrix obtained by linearly operating on the channel matrix, a unitary matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a unitary matrix obtained by linearly operating on the operation matrix, a diagonal matrix obtained by linearly operating on the channel matrix, a diagonal matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a diagonal matrix obtained by linearly operating on the operation matrix, a triangular matrix obtained by linearly operating on the channel matrix, a triangular matrix obtained by linearly operating on the correlation matrix or the pseudo-correlation matrix, a triangular matrix obtained by linearly operating on the operation matrix, among the feature quantities, generating a normalized value of the phase, amplitude, real component, imaginary component value of one or more feature quantities, and the range of coefficients of the one or more values. The position estimation device according to any one of claims 1 to 4.

6. In a position estimation method performed by a position estimation device, Receiving a wireless signal transmitted from a wireless communication unit of a wireless terminal, and obtaining channel information regarding radio wave propagation from the wireless signal; Converting the channel information into an input feature quantity that can be input to a position estimation model, storing it, and outputting a plurality of input feature quantities corresponding to a plurality of times to the position estimation model; Estimating and calculating the position of the wireless terminal by inputting the plurality of input feature quantities into a position estimation model that models the relationship between the channel information regarding radio wave propagation and the position information of the wireless terminal by machine learning; Updating the position estimation model; and performing In the step of updating, Generating a training data set in which a plurality of position information of the wireless terminal corresponding to a plurality of times obtained by a method other than the method of estimating and calculating and the plurality of input feature quantities corresponding to the plurality of times are aligned at the same time, and using the training data set to improve the accuracy of the relationship between the input feature quantity and the position information. A position estimation method for updating the position estimation model.

7. A position estimation program that causes a computer to function as the position estimation device according to any one of claims 1 to 5.

Images