Estimation device and control method thereof, data shaping device and control method thereof, and program

Abstract

To provide an estimation device that efficiently performs radio wave propagation estimation using a neural network.SOLUTION: An estimation device that estimates radio wave propagation characteristics includes: acquisition means that acquires first information on a ground surface and a building and second information on positions of a radio wave transmission point and a reception point; range determination means that determines a geographical range of the first information and the second information acquired by the acquisition means; series data generation means that generates, on the basis of the first information and the second information in the geographical range, M pieces of series data along each of M pieces of geographical routes set in the geographical range; derivation means that respectively inputs the M pieces of series data into pre-trained RNN encoders and derives M pieces of feature vectors corresponding to the M pieces of series data; generation means that generates a feature map by combining the M pieces of feature vectors; and estimation means that inputs the feature map into a pre-trained DNN and estimates the radio wave propagation characteristics.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a technique for estimating radio wave propagation characteristics. [Background technology]

[0002] There is known a technique for estimating radio wave propagation characteristics using a neural network (NN) and map data as input. For example, Patent Document 1 discloses a method for estimating radio wave propagation characteristics by extracting a first parameter representing a structure that may become an obstacle to radio waves from map data and applying a fully connected neural network (FNN) to the extracted first parameter and a second parameter representing the configuration of a wireless communication system. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-122008 Summary of the Invention [Problem to be solved by the invention]

[0004] In the method disclosed in Patent Document 1, map data with a fixed image size (e.g., 64 x 64 pixels) must be input to the NN. However, since the image size changes when the distance between the transmitter and receiver (the distance between the transmitting point and the receiving point) changes, when map data of a fixed range is extracted with a fixed resolution, information deficiency or duplication may occur depending on the distance between the transmitter and receiver. On the other hand, when the map extraction range is variable and scaling (resolution change) to a fixed image size occurs, the physical meaning of one pixel changes due to this scaling. As such, it is generally difficult to prepare data to input to the NN.

[0005] The present invention has been made in view of the above problems, and has an object to provide a technique for efficiently performing radio wave propagation estimation using a neural network. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, an estimation device for estimating radio wave propagation characteristics according to the present invention has the following configuration. an acquisition means for acquiring first information about the earth's surface and buildings and second information about the positions of transmitting and receiving points of radio waves; a range determination means for determining a geographical range of the first information and the second information acquired by the acquisition means; a sequence data generating means for generating M pieces of sequence data along M geographical routes set in the geographical range based on the first information and the second information in the geographical range acquired by the acquiring means; a derivation means for inputting the M pieces of sequential data into a pre-trained recurrent neural network (RNN) encoder, respectively, and deriving M pieces of feature vectors corresponding to the M pieces of sequential data; a generating means for generating a feature map by combining the M feature vectors; an estimation means for inputting the feature map into a pre-trained deep neural network (DNN) to estimate radio wave propagation characteristics; It has. [Effects of the Invention]

[0007] According to the present invention, it is possible to provide a technique for efficiently performing radio wave propagation estimation using a neural network. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an RNN encoder. [Figure 2] FIG. 2 is a diagram illustrating a hardware configuration of a radio wave propagation estimation device. [Figure 3] FIG. 2 is a diagram illustrating a functional configuration and a processing flow of a radio wave propagation estimating device. [Figure 4] FIG. 2 is a diagram illustrating a multidimensional vector x in the first embodiment. [Figure 5] 10 is a flowchart for generating building height data for a target area. [Figure 6] 10A and 10B are diagrams illustrating feature maps generated from map data with different transmission and reception distances; [Figure 7] FIG. 10 is a diagram for explaining another example of a multidimensional vector x (in a Cartesian coordinate system). [Figure 8] FIG. 10 is a diagram for explaining another example of a multidimensional vector x (polar coordinate system). [Figure 9] FIG. 10 is a diagram for explaining another example of a multidimensional vector x (elliptical coordinate system). DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the scope of the invention claimed. Although multiple features are described in the embodiments, not all of these multiple features are necessarily essential to the invention, and multiple features may be combined arbitrarily. Furthermore, in the accompanying drawings, the same reference numerals are used to designate the same or similar components, and redundant explanations will be omitted.

[0010] (First embodiment) As a first embodiment of the radio wave propagation estimation device according to the present invention, a radio wave propagation estimation device using a deep neural network (DNN) will be described below as an example. As the DNN, a fully connected neural network (FNN) or a convolutional neural network (CNN) can be used.

[0011] <Summary> As mentioned in the challenges section, preparing the data to be input to a radio wave propagation estimation device that uses a DNN is generally difficult. This is because the image size changes when the distance between the transmitter and receiver (the distance between the transmitting point and the receiving point) changes, making it difficult to generate a large amount of consistent data.

[0012] In the field of natural language processing, processing systems using recurrent neural networks (RNNs) have been proposed. In particular, Literature A describes a system that uses RNNs to convert input sentences into fixed-length feature vectors for processing. Document A: K. Cho et al., "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation", arXiv:1406.1078v1, June 2014

[0013] Figure 1 is a diagram explaining the RNN encoder-decoder model shown in Figure 1 of Reference A. The RNN encoder outputs a fixed-length feature vector (a vector representing the meaning of the input sentence) for an input data sequence of length T. The RNN decoder then generates and outputs a data sequence of length T' based on the fixed-length feature vector. In other words, the RNN encoder has the property of converting data sequences of various lengths (variable length) into consistent fixed-length data.

[0014] Therefore, in the first embodiment, the characteristics of this RNN encoder are utilized to input map data corresponding to various lengths of transmission-reception distances into the RNN encoder and convert it into consistent fixed-length data. The fixed-length data output from the RNN encoder is then combined to generate data (feature map) to be input to the DNN. This makes it possible to convert environmental information (such as distance) that changes depending on the transmission-reception distance into fixed-length data while maintaining its physical meaning. This makes it possible to efficiently generate data (feature map) of a fixed size to be input to the DNN.

[0015] <Hardware configuration> 2 is a diagram showing the hardware configuration of radio wave propagation estimating apparatus 200. In one example, radio wave propagation estimating apparatus 200 includes a processor 201, a ROM 202, a RAM 203, a storage unit 204, an operation unit 205, and a display unit 206. Note that various interfaces including wired / wireless communication units (not shown) may also be included.

[0016] The processor 201 is a computer configured to include one or more processing circuits, such as a general-purpose CPU (Central Processing Unit) or ASIC (Application Specific Integrated Circuit), and performs overall processing of the device and each process described below by reading and executing programs stored in the ROM 202 or storage device 204. Note that a GPU (Graphics Processing Unit) or an NPU (Neural Processing Unit) may be configured to be used for processing the neural network (NN).

[0017] The ROM 202 is a read-only memory that stores information such as programs and various parameters related to processes executed by the UE 101. The RAM 203 functions as a workspace when the processor 201 executes a program, and is a random access memory that stores temporary information. The storage unit 204 is configured, for example, by a removable external storage device.

[0018] The operation unit 205 is a functional unit that accepts operation input from the user. For example, it is realized by devices such as a keyboard, a mouse, and a touch panel. The display unit 206 is a functional unit that presents information to the user and displays a GUI (graphical user interface). For example, it is realized by devices such as a liquid crystal / organic EL display. It is also possible to use a touch panel display in which the operation unit 205 and the display unit 206 are integrated.

[0019] <Functional configuration and processing flow> 3A is a diagram showing the functional configuration of the radio wave propagation estimation device 200. The radio wave propagation estimation device 200 includes an elevation / building database (DB) 301, a transmission / reception point DB 302, a range determination unit 303, a sequence data generation unit 304, an RNN encoder 305, an integration unit 306, and a DNN 307. Note that, although the following description will be given assuming that all of the above-mentioned functional units are realized by the single hardware device shown in FIG. 2, a configuration in which all of the functional units are realized by sharing the functions among multiple hardware devices may also be used. For example, a three-unit configuration may be used in which the elevation / building DB 301 and the transmission / reception point DB 302 are a database device, the range determination unit 303 to the integration unit 306 are a data shaping device, and the DNN 307 is a radio wave propagation estimation device.

[0020] The elevation / building DB 301 is a database that stores, for example, elevation data of the earth's surface associated with longitude and latitude, and shape data of buildings (hereinafter also referred to as map data). The transmission / reception point DB 302 is a database that stores position information (including height, etc.) of transmission points and reception points. The elevation / building DB 301 and transmission / reception point DB 302 are realized, for example, by the storage unit 204. However, they may also be realized by an external database device (not shown) on the network. In that case, the radio wave propagation estimation device 200 is configured to be able to acquire these data via a communication unit (not shown). For example, PLATEAU, a 3D city model provided by the Ministry of Land, Infrastructure, Transport and Tourism, may be used as the elevation / building DB 301.

[0021] The range determination unit 303 is a functional unit that determines the geographical range that is the target of radio wave propagation estimation. The range determination unit 303 divides the determined range (target area) into a mesh / grid pattern. FIG. 4 shows an example in which map data of the target area is divided into multiple grids with M rows and n columns. Then, multiple pieces of environmental information (three items in FIG. 4: "building height," "building occupancy rate," and "transmission point distance") are determined for each unit area (grid). A method for generating the environmental information will be described later with reference to FIG. 5.

[0022] The sequence data generator 304 generates sequence data along a given geographical route (the horizontal axis direction (from left to right) in FIG. 4). In this example, M sequence data generators 304 are provided, each generating sequence data for one row of map data for the target area. In other words, the sequence data is a data string of multidimensional vectors along a given geographical route.

[0023] FIG. 4 is a diagram illustrating a multidimensional vector x in the first embodiment. The map data shown in the lower part of FIG. 4 corresponds to the target area determined by the range determination unit 303. Here, an example is shown in which the target area is divided into a grid of 5 rows and n columns. The sequence data corresponding to the first row (the top row) of the target area is exemplarily shown in the upper part of FIG. 4. As described above, three pieces of environmental information, "building height," "building occupancy rate," and "transmission point distance," are included in one grid, and n pieces of this information are linked along a route to form one piece of sequence data (x M 1~x M n ) is equivalent to

[0024] The RNN encoder 305 performs an operation on the sequence data generated and input by the corresponding sequence data generator 304, and generates a hidden layer vector h M 1~h M n This process is shown in FIG. 3(B). As shown in FIG. 3(A), M sets of sequence data generators 304 and RNN encoders 305 are prepared. The integrator 306 derives the hidden layer vector h generated and input by the RNN encoder 305. M 1~h M n are combined to generate one feature map. As shown in FIG. 3(C), the feature map is a feature map of a predetermined size (a tensor with L rows and M columns), where L is the dimension of h. The learning process of the RNN encoder 305 can use existing technology (such as reference A), so a detailed description will be omitted.

[0025] FIG. 6 is a diagram illustrating feature maps generated from map data with different transmitter-receiver distances. Here, the transmitter-receiver distance in FIG. 6(A) is longer than the transmitter-receiver distance in FIG. 6(B). However, as described above, the RNN encoder can convert environmental information (such as distance) that changes depending on the transmitter-receiver distance into fixed-length data (fixed-length feature vectors) while maintaining its physical meaning. Therefore, the feature maps generated in FIG. 6(A) and FIG. 6(B) are both data that maintain the same physical meaning.

[0026] The DNN 307 is a neural network that has been trained in advance to output radio wave propagation characteristics for an input feature map. As described above, an FNN, a CNN, or the like can be used as the DNN 307. Note that existing technologies (such as Patent Document 1) can be used for the learning process and estimation process of the DNN 307, and therefore detailed explanations thereof will be omitted.

[0027] <How to generate environmental information> 5 is a flowchart for generating building height data for a target area. This process is executed by the range determination unit 303.

[0028] In S501, the range determination unit 303 reads three-dimensional data (shape data of buildings) from the elevation / building DB 301. In S502, the range determination unit 303 divides the target area into multiple grids. The geographical size of one grid is, for example, 5 m square, 1 m square, etc. Then, the following processes of S503 to S504 are executed for each grid.

[0029] In S503, the range determination unit 303 lists all buildings in the currently focused grid based on the three-dimensional data (building positions). In S504, the range determination unit 303 extracts the maximum building height in the currently focused grid based on the three-dimensional data (building height). Building heights (or altitudes if no buildings exist) are extracted for all grids, and the "building height" data shown in the upper part of Figure 4 is generated.

[0030] As the "building height," a "relative building height" may be calculated by subtracting the transmitting antenna height (or the receiving antenna height) from the maximum building height.

[0031] "Building occupancy rate" data can also be generated in a similar manner (the processing in S504 is the calculation processing of the building occupancy rate based on the three-dimensional data (building positions)). Also, the "transmission point distance" is calculated (for example, as a straight-line distance (m)) based on the position information of the transmission points in the target area read from the transmission / reception point DB302 and the position information of each grid. Note that the logarithmic value of the distance may be calculated. Also, it may be configured to calculate the reception point distance instead of the transmission point distance.

[0032] As described above, according to the first embodiment, map data corresponding to various lengths of transmission-reception distances are input to an RNN encoder and converted into consistent fixed-length data. The fixed-length data output from the RNN encoder are then combined to generate data to be input to a DNN (which performs radio wave propagation estimation). This allows environmental information that changes depending on the transmission-reception distance to be converted into fixed-length data while maintaining its physical meaning. This makes it possible to efficiently generate data (feature maps) of a fixed size to be input to the DNN, thereby enabling efficient radio wave propagation estimation. Furthermore, because more feature maps are available, training the DNN using these feature maps is expected to improve the accuracy of modeling radio wave propagation characteristics.

[0033] (Variation) In the first embodiment described above, the multidimensional vector (x M n ) is a multidimensional vector containing multiple environmental information (building height, building occupancy rate, distance) in one unit area (grid). However, x M n For example, a multidimensional vector including one piece of environmental information (building height) in a plurality of unit areas along the route may be used.

[0034] 7 is a diagram illustrating another example of the multidimensional vector x. Specifically, this shows an example in which a multidimensional vector including building heights in multiple grids along a route in a Cartesian coordinate system is used.

[0035] As in the first embodiment described above, the target area is divided vertically (column direction, latitude direction) and horizontally (row direction, longitude direction) into a plurality of grids. However, unlike the first embodiment, the building height data in a plurality of horizontally consecutive grids (three in this case) is divided into x M n where x M 1 contains three grids of data corresponding to the first row, first to third columns, and x M 2 contains data for three grids corresponding to the second to fourth columns of the first row. Similarly, the three grids are slid to the right in sequence to generate n multidimensional vectors x.

[0036] By inputting the n multidimensional vectors x generated in this way into an RNN encoder, n fixed-length feature vectors h are obtained. By combining the n feature vectors h and inputting them into a pre-trained DNN, an estimated value of the radio wave propagation characteristics can be obtained, as in the first embodiment. Note that the configuration of the radio wave propagation estimation device is the same as in the first embodiment (FIGS. 2 and 3), and therefore a description thereof will be omitted.

[0037] 8 is a diagram illustrating yet another example of the multidimensional vector x. Specifically, this shows an example in which a multidimensional vector including building heights in multiple grids along a route in a polar coordinate system is used.

[0038] Unlike the first embodiment described above, the target area is divided into multiple grids defined by polar coordinates with the transmitting point (Tx) as the center. Here, the transmitting point is set as the center, and the direction of the receiving point (Rx) is set as the reference (θ=0) and the deflection angle θ is defined counterclockwise. The distance r in the radial direction is equal to the distance from the transmitting point. Here, x M 1 contains data for three grids corresponding to the first to third grids from the center in a given declination direction, and x M2 contains data for three grids corresponding to the second to fourth grids. Similarly, the three grids are slid outward in sequence to generate n multidimensional vectors x.

[0039] In this way, it is possible to define n multidimensional vectors x without using grid division in an orthogonal coordinate system. By inputting the n multidimensional vectors x to an RNN encoder, n fixed-length feature vectors h are obtained. By combining the n feature vectors h and inputting them to a pre-trained DNN, an estimated value of radio wave propagation characteristics can be obtained, as in the first embodiment.

[0040] Furthermore, grid division according to various coordinate systems is possible, but it is also possible to use an elliptical coordinate system in which the transmitting point (Tx) and the receiving point (Rx) are defined as the foci of two ellipses.

[0041] FIG. 9 is a diagram for explaining yet another example of a multidimensional vector x. Specifically, it shows an example using a multidimensional vector including building heights in multiple grids along a path in an elliptical coordinate system. From the definition of an ellipse, for any point P on the ellipse, (distance between Tx and P) + (distance between Rx and P) = constant. Here, x M 1 contains data for three grids corresponding to the first to third grids from the center in a given direction, where r = (the distance between Tx and P). Then, the three grids are slid outward in sequence to generate n multidimensional vectors x.

[0042] Alternatively, the free space propagation loss (dB), which can be calculated from distance data and frequency, or parameters (UMa, UMi, RMa, etc.) of the propagation loss model formula of ITU-R M.2412 may be used as the feature.

[0043] Furthermore, this invention enables efficient estimation of radio wave propagation characteristics using neural networks, which can contribute to achieving Goal 9 of the United Nations' Sustainable Development Goals (SDGs), which is to "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0044] The invention is not limited to the above-described embodiment, and various modifications and variations are possible within the scope of the gist of the invention. [Explanation of symbols]

[0045] 200 Radio wave propagation estimation device; 301 Elevation and building DB; 302 Transmission and reception point DB; 303 Range determination unit; 304 Sequence data generation unit; 305 RNN encoder; 306 Integration unit; 307 DNN

Claims

1. an acquisition means for acquiring first information relating to the ground surface and buildings and second information relating to the positions of transmitting points and receiving points of radio waves; a range determination means for determining a geographical range of the first information and the second information to be acquired by the acquisition means; a sequence data generating means for generating M pieces of sequence data along M geographical routes set in the geographical range based on the first information and the second information in the geographical range acquired by the acquiring means; a derivation means for inputting the M pieces of sequential data into a pre-trained recurrent neural network (RNN) encoder, respectively, and deriving M pieces of feature vectors corresponding to the M pieces of sequential data; a generating means for generating a feature map by combining the M feature vectors; an estimation means for inputting the feature map into a pre-trained deep neural network (DNN) to estimate radio wave propagation characteristics; An estimation device comprising:

2. The sequence data generating means generates M pieces of sequence data for each of at least two different geographical ranges determined by the range determining means.

2. The estimation device according to claim 1 .

3. The M feature vectors are fixed-length feature vectors.

2. The estimation device according to claim 1 .

4. The geographic area includes at least one transmission point.

2. The estimation device according to claim 1 .

5. The first information includes information about the elevation of the earth's surface and the location and height of buildings.

2. The estimation device according to claim 1 .

6. The sequence data generating means Dividing the geographical area into a plurality of grids based on a Cartesian coordinate system; calculating a building height and a building occupancy rate in each grid from the first information; Calculating a transmission point distance in each grid from the second information; each of the M geographical routes is a route along the same axis of the Cartesian coordinate system; Each of the M pieces of sequence data is generated as a data string of a multidimensional vector including the building height, building occupancy rate, and transmission point distance in each grid.

2. The estimation device according to claim 1 .

7. The sequence data generating means Dividing the geographical area into a plurality of grids based on a Cartesian coordinate system; calculating a building height in each grid from the first information; each of the M geographical routes is a route along the same axis of the Cartesian coordinate system; Each of the M pieces of sequence data is generated as a data string of a multidimensional vector including building heights in a plurality of consecutive grids along the corresponding geographical route.

2. The estimation device according to claim 1 .

8. The sequence data generating means Dividing the geographical area into a plurality of grids based on a polar coordinate system; calculating a building height in each grid from the first information; Each of the M geographical routes is a route along a radial direction, Each of the M pieces of sequence data is generated as a data string of a multidimensional vector including building heights in a plurality of consecutive grids along the corresponding geographical route.

2. The estimation device according to claim 1 .

9. The sequence data generating means Dividing the geographical area into a plurality of grids based on an ellipsoidal coordinate system with the transmitting point and the receiving point as foci; calculating a building height in each grid from the first information; Each of the M geographical routes is a route along a radial direction, Each of the M pieces of sequence data is generated as a data string of a multidimensional vector including building heights in a plurality of consecutive grids along the corresponding geographical route.

2. The estimation device according to claim 1 .

10. A control method for an estimation device that estimates radio wave propagation characteristics, comprising: the estimation device is configured to be able to acquire first information regarding the ground surface and buildings and second information regarding positions of transmitting points and receiving points of radio waves; The control method includes: a range determination step of determining a geographic range of the first information and the second information to be acquired; acquiring the first information and the second information in the geographical range; a sequence data generating step of generating M pieces of sequence data along M geographical routes set in the geographical range based on the first information and the second information acquired in the acquiring step; a derivation step of inputting the M pieces of sequential data into a pre-trained recurrent neural network (RNN) encoder, respectively, and deriving M pieces of feature vectors corresponding to the M pieces of sequential data; a generating step of generating a feature map by combining the M feature vectors; an estimation step of inputting the feature map into a pre-trained deep neural network (DNN) to estimate radio wave propagation characteristics; A control method comprising:

11. A data shaping device that generates a feature map to be input to an estimation device that estimates radio wave propagation characteristics, the estimation device is configured to estimate radio wave propagation characteristics by inputting a feature map of a predetermined size into a pre-trained deep neural network (DNN); The data shaping device an acquisition means for acquiring first information relating to the ground surface and buildings and second information relating to the positions of transmitting points and receiving points of radio waves; a range determination means for determining a geographical range of the first information and the second information to be acquired by the acquisition means; a sequence data generating means for generating M pieces of sequence data along M geographical routes set in the geographical range based on the first information and the second information in the geographical range acquired by the acquiring means; a derivation means for inputting the M pieces of sequential data into a pre-trained recurrent neural network (RNN) encoder, respectively, and deriving M pieces of feature vectors corresponding to the M pieces of sequential data; a generating means for generating a feature map of the predetermined size by combining the M feature vectors; A data shaping device comprising:

12. 1. A method for controlling a data shaping device that generates a feature map to be input to an estimation device that estimates radio wave propagation characteristics, comprising: the estimation device is configured to estimate radio wave propagation characteristics by inputting a feature map of a predetermined size into a pre-trained deep neural network (DNN); the data shaping device is configured to be able to acquire first information regarding the earth's surface and buildings and second information regarding the positions of transmitting points and receiving points of radio waves; The control method includes: a range determination step of determining a geographic range of the first information and the second information to be acquired; acquiring the first information and the second information in the geographical range; a sequence data generating step of generating M pieces of sequence data along M geographical routes set in the geographical range based on the first information and the second information acquired in the acquiring step; a derivation step of inputting the M pieces of sequential data into a pre-trained recurrent neural network (RNN) encoder, respectively, and deriving M pieces of feature vectors corresponding to the M pieces of sequential data; a generating step of generating a feature map of the predetermined size by combining the M feature vectors; A control method comprising:

13. A program for causing a computer to execute the control method according to claim 10 or 12.

Images