JP7569072B2 - Radio wave propagation estimation system, radio wave propagation estimation method, and method for manufacturing a generation unit - Google Patents

Description

Article 30, paragraph 2 of the Patent Act applies. Nobumasa Sakauchi and Katsuya Sudo, "A Study on Radio Wave Propagation Estimation Method Using Deep Generative Models," Proceedings of the 2020 Society Conference, B-17-20, p. 306, Institute of Electronics, Information and Communication Engineers, September 1, 2020.

Application of Article 30, Paragraph 2 of the Patent Act Nobumasa Sakauchi and Katsuya Sudo, "Spatial Extrapolation with Generative Adversarial Networks for Radio Map Construction," Proceedings of 2020 International Conference on Emerging Technologies for Communications, p. 1, The Institute of Electronics, Information and Communication Engineers, December 2, 2020

Article 30, paragraph 2 of the Patent Act applies. Nobumasa Sakauchi and Katsuya Sudo, "A method for constructing radio wave maps using deep generative models with 3D maps," Technical Report of the Institute of Electronics, Information and Communication Engineers, Vol. 120, No. 238, pp. 114-119, Institute of Electronics, Information and Communication Engineers, November 11, 2020.

The present invention relates to a radio wave propagation estimation system, a radio wave propagation estimation method, and a method for manufacturing a generation unit for extrapolating radio wave propagation characteristics, which are the distribution of power received in a specified area for a radio signal transmitted from a specified transmitting station.

Next-generation wireless systems and position estimation systems are expected to realize spatial radio wave propagation estimation through crowdsensing using mobile devices and sensors as radio wave environment observation terminals. However, observation information cannot be obtained from areas where mobile devices and sensors cannot observe, or from areas where the received power value is below the noise floor. Estimating radio wave propagation in areas where there is no observation information is called extrapolation estimation. On the other hand, when mobile devices and sensors are partially present and observation information can be obtained in gaps, estimating radio wave propagation at the gaps is called interpolation estimation.

In relation to the above, Patent Document 1 (JP 2019-080144 A) discloses a wireless environment status prediction system. This wireless environment status prediction system is for predicting the wireless environment status of a first point, which is a predetermined position in space. This wireless environment status prediction system includes a transmitter, a receiver, a channel usage status observation unit, a channel state data acquisition unit, and a prediction unit. Here, the transmitter generates a wireless signal by performing a transmission process on data. The receiver acquires data carried by the wireless signal by performing a reception process on the wireless signal. The channel usage status observation unit observes the usage status of a channel used for communication of the wireless signal by controlling at least one of the transmitter and the receiver. The channel state data acquisition unit acquires first data related to the channel state of the first point during a first period and second data correlated with the first data based on a signal acquired by at least one of the transmitter, the receiver, and the channel usage status observation unit. The prediction unit performs a prediction process to predict future data of the first data at the first location based on the first data and the second data, and to predict future data of the second data at the first location based on the first data and the second data.

Patent Document 1 considers regression analysis using the distance between the transmitting point and the receiving point as a feature in order to estimate distance attenuation, but it is not possible to estimate short-term fluctuations in power values that are influenced by nearby structures.

JP 2019-080144 A

The present invention provides a radio wave propagation estimation system, a radio wave propagation estimation method, and a method for manufacturing a generator that perform extrapolation estimation of radio wave propagation characteristics, which is the distribution of power received in a specified area for a radio signal transmitted from a specified transmitting station, with higher accuracy. Other objects and novel features will become apparent from the description of this specification and the accompanying drawings.

Below, the means for solving the problem are explained using the numbers used in (Mode for carrying out the invention). These numbers are added to clarify the correspondence between the description in (Claims) and (Mode for carrying out the invention). However, these numbers should not be used to interpret the technical scope of the invention described in (Claims).

According to one embodiment, the radio wave propagation estimation system (1) includes a first estimation means (311, 312, 313), a simple regression analysis unit (314), an adder (315), and an output device (5). The first estimation means (311, 312, 313) generates short-term fluctuation map data (93) representing components estimated to originate from structures among fluctuations in received power of a radio signal transmitted from a predetermined transmitting station received at each of a plurality of first points (92) included in the first region (9) based on structure map data (90) representing planar spatial information of a structure present in a first region (9) included in the predetermined region. The simple regression analysis unit (314) generates distance attenuation map data (96) representing a component estimated to be derived from the distance from a predetermined transmitting station to the first points (92) among the fluctuations of the received power received at the first points (92) based on second area radio wave map data (84) obtained by observing the received power of the radio signal at a plurality of second points included in a second area (8) that is included in a predetermined area and different from the first area (9). The adder (315) generates first area radio wave map data (97) by extrapolating the fluctuations of the received power of the radio signal received at the first points (92) based on the short-term fluctuation map data (93) and the distance attenuation map data (96). The output device (5) outputs the first area radio wave map data (97) to the outside. The first estimation means (311, 312, 313) includes a first generation unit (312). The first generation unit (312) performs machine learning based on the correlation between the first divided short-interval variation map data of the second region (8) based on the second region radio wave map data (84) and the second divided short-interval variation map data of the second region (8) based on the second region structure map data (80) representing planar spatial information of structures existing in the second region (8), and generates short-interval variation map data for each of the multiple first locations (9) based on the structure map data (90) for each of the multiple first locations (92).

According to one embodiment, the radio wave propagation estimation method includes generating (S11, S12, S13) short-term fluctuation map data (93) representing components estimated to originate from structures among fluctuations in received power of a radio signal transmitted from a predetermined transmitting station received at each of a plurality of first points (92) included in the first region (9) based on structure map data (90) representing planar spatial information of structures present in a first region (9) included in the predetermined region, and generating (S14) distance attenuation map data (96) representing components estimated to originate from distances from the predetermined transmitting station to the plurality of first points (92) among fluctuations in received power received at the plurality of first points (92) based on second region radio wave map data (84) in which the received power of the radio signal is observed at a plurality of second points included in a second region (8) included in the predetermined region and different from the first region (9). The radio wave propagation estimation method further includes generating (S15) first area radio wave map data (97) by extrapolating the fluctuation of the received power of the radio signal received at a plurality of first points (92) based on the short-range fluctuation map data (93) and the distance attenuation map data (96), and outputting (S16) the first area radio wave map data (97) to the outside. Generating the short-range fluctuation map data (S11, S12, S13) includes generating (S12) short-range fluctuation map data of the first area (9) based on the structure map data (90) of the first area (9) by a generating unit (312) that has performed machine learning based on the correlation between the first divided short-range fluctuation map data of the second area (8) based on the second area radio wave map data (84) and the second divided short-range fluctuation map data of the second area (8) based on the second area structure map data (80) representing the planar space information of the structure existing in the second area (8).

According to one embodiment, the manufacturing method of the generating unit includes generating (S21) a plurality of first observation area division short-interval fluctuation map data each representing a fluctuation in received power other than attenuation of received power resulting from the distance from a predetermined transmitting station to each of a plurality of partial areas included in the predetermined observation area (8) based on radio wave map data (84) observing the distribution of received power of a radio signal transmitted from a predetermined transmitting station received at each of a plurality of points included in the predetermined observation area (8), and generating (S22) a plurality of second observation area division short-interval fluctuation map data each representing a fluctuation in received power resulting from structures existing in the plurality of partial areas based on observation area structure map data (80) representing planar spatial information of structures existing in the predetermined observation area (8) by a generating unit (324) having a first parameter. The manufacturing method of the generating unit further includes discriminating (S23) between the plurality of first observation area division short-interval fluctuation map data and the plurality of second observation area division short-interval fluctuation map data by a discriminating unit (325) having a second parameter. The manufacturing method of the generating unit further includes updating the first parameter so that the numerical value representing the discrimination result reaches a predetermined threshold value (S25), generating second observation area division short-interval variation map data using the updated first parameter (S22), updating the second parameter so that the numerical value does not reach the threshold value (S25), and making a discrimination using the updated second parameter (S23). The manufacturing method of the generating unit further includes storing the first parameter as a parameter of the generating unit (S12) that performed machine learning so that it can be used by an external estimation means (31) when the discrimination result reaches the threshold value (S26).

According to the embodiment described above, it is possible to perform extrapolation estimation of radio wave propagation characteristics in a specified area with higher accuracy.

FIG. 1 is a block circuit diagram showing an example of the configuration of a radio wave propagation estimating system according to an embodiment. FIG. 2 is a block circuit diagram showing an example of the configuration of the estimation means according to an embodiment. FIG. 3 is a flowchart showing an example of a configuration of a radio wave propagation estimation method according to an embodiment. FIG. 4 is a diagram showing an example of the positional relationship between observed and unobserved areas and radio wave propagation characteristics of the observed areas in one embodiment. FIG. 5 is a diagram showing an example of unobserved area structure map data and observed area structure map data according to one embodiment. FIG. 6A is a diagram for explaining the relationship between unobserved area structure map data and meshes according to one embodiment. FIG. 6B is a diagram for explaining the relationship between the unobserved area structure map data, the meshes, and the first divided structure map data according to one embodiment. FIG. 6C is a diagram for explaining the relationship between the unobserved area structure map data, the meshes, and the second divided structure map data according to one embodiment. FIG. 6D is a diagram for explaining the relationship between the unobserved area structure map data, the meshes, and the third divided structure map data according to one embodiment. FIG. 7 is a block circuit diagram showing an example of a configuration of a generating unit according to an embodiment. FIG. 8 is a diagram showing an example of unobserved area short-range variation map data according to one embodiment. FIG. 9 is a diagram showing an example of observation area radio wave map data according to one embodiment. FIG. 10 is a diagram showing an example of observed area distance map data and unobserved area distance map data according to one embodiment. FIG. 11 is a diagram illustrating an example of unobserved area distance attenuation map data according to one embodiment. FIG. 12 is a diagram illustrating an example of unobserved area radio wave map data according to an embodiment. FIG. 13 is a block circuit diagram showing an example of the configuration of a learning means according to an embodiment. FIG. 14 is a flowchart showing an example of a configuration of a manufacturing method for a generating unit according to an embodiment. FIG. 15 is a diagram showing an example of observed area distance attenuation map data and unobserved area distance attenuation map data according to one embodiment. FIG. 16 is a diagram showing an example of short-range observation area variation map data according to one embodiment. FIG. 17 is a block circuit diagram showing an example of the configuration of a determination unit according to an embodiment.

In one embodiment, the radio wave propagation estimation system extrapolates radio wave propagation characteristics, which are the distribution of power received in a specified area for a wireless signal transmitted from a specified transmitting station, using a radio wave propagation method. In one embodiment, before that, a generating unit included in the radio wave propagation estimation system is manufactured using machine learning or the like. With reference to the attached drawings, the following describes an embodiment of the radio wave propagation estimation system, the radio wave propagation estimation method, and the manufacturing method of the generating unit.

(Embodiment)
As shown in FIG. 1, the radio wave propagation estimation system 1 according to the embodiment includes a bus 2, a calculation device 3, a storage device 4, and an output device 5. The calculation device 3, the storage device 4, and the output device 5 are communicably connected to each other via the bus 2. The radio wave propagation estimation system 1 may be configured as a so-called computer. That is, the calculation device 3 realizes the function of the radio wave propagation estimation system 1 by executing a program stored in the storage device 4. In other words, the calculation device 3 and the program work together to realize the function of the radio wave propagation estimation system 1. The program may be read from the recording medium 401 and stored in the storage device 4. The output device 5 outputs the result of correcting the radio wave propagation characteristics to the outside. As an example, the output device 5 may include a display device.

The calculation device 3 includes an estimation means 31 and a learning means 32. The storage device 4 stores an estimation program 41 and a learning program 42. The estimation means 31 is a function of the radio wave propagation estimation system 1 that is realized by the calculation device 3 executing the estimation program 41. The learning means 32 is another function of the radio wave propagation estimation system 1 that is realized by the calculation device 3 executing the learning program 42. That is, for convenience, the estimation means 31 and the learning means 32 are shown in FIG. 1 as components of the calculation device 3, but in reality, they may be means for realizing the functions of the radio wave propagation estimation system 1.

The estimation means 31 extrapolates radio wave propagation characteristics, which are the distribution of power received in a specified area for a radio signal transmitted from a specified transmitting station. Extrapolation estimation is estimation of radio wave propagation characteristics in an area where there is no observational information, and is distinguished from interpolation estimation, which estimates propagation radio wave characteristics in gaps in an area where observational information is partially obtained. The learning means 32 implements machine learning in the generation unit provided in the estimation means 31.

As shown in FIG. 2, the estimation means 31 according to one embodiment includes a data expansion unit 311, a generation unit 312, a map data integration unit 313, a simple regression analysis unit 314, and an adder 315. The data expansion unit 311, the generation unit 312, the map data integration unit 313, the simple regression analysis unit 314, and the adder 315 are functions of the estimation means 31 that are realized by the calculation device 3 executing the estimation program 41.

The operation of the estimation means 31 according to one embodiment, that is, the radio wave propagation estimation method according to one embodiment, will be described with reference to the flowchart in FIG. 3.

In step S11, the data expansion unit 311 reads the unobserved area structure map data from the storage device 4, and generates divided structure map data for the unobserved area based on the unobserved area structure map data.

The unobserved area structure map data represents planar spatial information of structures existing in unobserved areas, and is prepared in advance so that it can be read from the storage device 4. An unobserved area is a desired area for which radio wave propagation characteristics are desired to be known, where observation of radio wave propagation characteristics has not been performed even partially. Conversely, a desired area where observation of radio wave propagation characteristics has been performed is referred to as an observed area. In other words, the part of the desired area other than the observed area is referred to as an unobserved area. Figure 4 shows an example of the positional relationship between observed area 8 and unobserved area 9, and the radio wave propagation characteristics of observed area 8.

As mentioned above, radio wave propagation characteristics are the distribution of power received in a specified area when a radio signal transmitted from a specified transmitting station is received. In one embodiment, the radio wave propagation characteristics are mainly observed and estimated in areas of the desired area, such as roads where no buildings exist.

The planar spatial information of the structure includes, for example, at least one of the following: the altitude of the structure, the altitude of the point where the structure is located, the planar distance from the transmitting station to the receiving point where the structure is located, the average altitude from the transmitting station to the receiving point, the maximum altitude of other structures existing between the transmitting station and the receiving point, the altitude of the other structure closest to the receiving point, the planar distance from the receiving point to the other structure closest to the receiving point, and the average altitude of other structures existing around the receiving point. The planar spatial information of the structure may be a combination of the above information summarized as a tensor. Note that the above combination is merely an example of the spatial information of the structure, and can be changed to a different combination of information. FIG. 5 is a diagram showing an example of unobserved area structure map data 90 and observed area structure map data 80 according to one embodiment. The observed area structure map data 80 represents planar spatial information of structures existing in the observed area 8.

The relationship between the unobserved area structure map data 90 and the divided structure map data will be explained with reference to Figures 6A, 6B, 6C, and 6D.

As shown in FIG. 6A, the data expansion unit 311 first divides the unobserved area structure map data 90 into a number of meshes 91 and 91A to 91P. When the multiple meshes 91 and 91A to 91P are not distinguished from one another, they are collectively referred to as meshes 91. As an example, the meshes 91 may be divided into a grid in orthogonal X and Y directions. Also, as an example, the lengths of each mesh 91 in the X and Y directions may be the same mesh unit length. As an example, the mesh unit length is 10 m (meters).

Next, the data expansion unit 311 defines the multiple meshes 91 as divided structure map data 92. As an example, as shown in FIG. 6B, the first divided structure map data 92A has lengths in the X and Y directions that are three times the mesh unit length, and is composed of a total of nine meshes 91 consisting of meshes 91A, 91B, 91C, 91E, 91F, 91G, 91I, 91J, and 91K.

As another example, as shown in FIG. 6C, the second divided structure map data 92B has a length in the X and Y directions that is three times the mesh unit length, and is composed of a total of nine meshes 91 consisting of meshes 91B, 91C, 91D, 91F, 91G, 91H, 91J, 91K, and 91L.

As another example, as shown in FIG. 6D, the third divided structure map data 92C has a length in the X and Y directions that is three times the mesh unit length, and is composed of a total of nine meshes 91 consisting of meshes 91E, 91F, 91G, 91I, 91J, 91K, 91M, 91N, and 91O.

When there is no distinction between divided structure map data 92A, 92B, and 92C, they are collectively referred to as divided structure map data 92.

In this way, the data expansion unit 311 generates multiple divided structure map data 92 having the same shape and the same area based on the unobserved area structure map data 90. At this time, each divided structure map data 92 is composed of multiple meshes 91, and each mesh 91 is included in the multiple divided structure map data 92. As an example, meshes 91B, 91C, 91F, 91G, 91J, and 91K are shared by the first divided structure map data 92A and the second divided structure map data 92B arranged in the X direction. In addition, meshes 91E, 91F, 91G, 91I, 91J, and 91K are shared by the first divided structure map data 92A and the third divided structure map data 92C arranged in the Y direction. The data expansion unit 311 may store the multiple divided structure map data 92 in the storage device 4.

In step S12, the generating unit 312 generates divided short-interval change map data based on the divided structure map data 92 of the unobserved area 9. The generating unit 312 may store the generated divided short-interval change map data in the storage device 4. The generating unit 312 is prepared in advance by the learning means 32 executing a manufacturing method for the generating unit 312, as described below.

The short-term variation map data represents the short-term variation in the corresponding divided structure map data 92. The short-term variation is a spatial variation of radio wave propagation characteristics in a section included in one divided structure map data 92 that is sufficiently short compared to the dimensions of the unobserved area 9. In other words, the short-term variation is a component of the variation in the received power of a radio signal transmitted from a specified transmitting station received at a point in the divided structure map data 92 that is estimated to be derived from a structure existing in the area of this divided structure map. The generation unit 312 performs machine learning based on the correlation between the received power of a radio signal transmitted from a specified transmitting station received in a specified area and the planar spatial information of a structure existing in the specified area. The generation unit 312 generates divided short-term variation map data for each divided structure map data 92 of the unobserved area 9, estimating the short-term variation of radio wave propagation characteristics derived from a structure included in the divided structure map data 92.

As shown in FIG. 7, the generation unit 312 according to one embodiment may sequentially perform a plurality of processes on the input divided structure map data 92. In the example of FIG. 7, the generation unit 312 includes convolution calculation units 61A, 61B, 61C, 61D, 61E, 61F, spectrum normalization calculation units 62A, 62B, 62C, 62D, 62E, batch normalization calculation units 63A, 63B, 63C, 63D, 63E, ReLU (Rectified Linear Unit) 64A, 64B, 64C, 64D, 64E, and Tanh (hyperbolic tangent) calculation unit 65.

When the convolution calculation units 61A, 61B, 61C, 61D, 61E, and 61F are not distinguished from one another, they are collectively referred to as the convolution calculation unit 61. When the spectral normalization calculation units 62A, 62B, 62C, 62D, and 62E are not distinguished from one another, they are collectively referred to as the spectral normalization calculation unit 62. When the batch normalization calculation units 63A, 63B, 63C, 63D, and 63E are not distinguished from one another, they are collectively referred to as the batch normalization calculation unit 63. When the ReLUs 64A, 64B, 64C, 64D, and 64E are not distinguished from one another, they are collectively referred to as the ReLU 64.

The convolution calculation unit 61 performs a convolution calculation using a predetermined kernel on input data having a 3x3 matrix structure. However, the above 3x3 is merely an example and does not limit the embodiment. The spectrum normalization calculation unit 62 performs spectrum normalization processing on the input data. The batch normalization calculation unit 63 performs batch normalization processing on the input data. The ReLU 64 performs normalization linear processing on the input data. The Tanh calculation unit 65 performs hyperbolic tangent calculation processing on the input data.

In the example of FIG. 7, the generation unit 312 according to one embodiment repeats five times a process set in which the input divided structure map data 92 is subjected to a convolution operation by the convolution operation unit 61, a spectral normalization process by the spectral normalization operation unit 62, a batch normalization process by the batch normalization operation unit 63, and a normalized linear process by the ReLU 64, and then further performs a convolution operation by the convolution operation unit 61E and a hyperbolic tangent operation process by the Tanh operation unit 65, and outputs the results of the calculation.

In step S13, the map data integration unit 313 generates unobserved area short-range variation map data based on the multiple divided short-range variation map data. FIG. 8 is a diagram showing an example of unobserved area short-range variation map data 93 according to one embodiment. The unobserved area short-range variation map data 93 represents short-range variation in the entire unobserved area 9, and represents short-range variation in each of all meshes 91 included in the unobserved area 9.

The map data integration unit 313 generates unobserved area short-range variation map data 93 by combining the multiple divided short-range variation map data obtained in step S12, in the opposite manner to the process performed by the data expansion unit 311 in step S11. At this time, for each of the multiple meshes 91 included in the unobserved area 9, the average value of the divided short-range variation map data corresponding to all the divided structure map data 92 including that mesh 91 is calculated as the short-range variation in that mesh 91. The map data integration unit 313 generates unobserved area short-range variation map data 93 by combining all the meshes 91 for which short-range variations have been calculated in this manner. The map data integration unit 313 may store the generated unobserved area short-range variation map data 93 in the storage device 4.

In step S14, the simple regression analysis unit 314 generates unobserved area distance attenuation map data based on the observed area radio wave map data. FIG. 9 is a diagram showing an example of observed area radio wave map data 84 according to one embodiment. The observed area radio wave map data 84 represents radio wave propagation characteristics observed in the observed area 8. The unobserved area distance attenuation map data represents an estimated value of attenuation of the power of a radio signal received at a specified point in the unobserved area 9 according to the distance from a specified transmitting station that transmits a radio signal to this point. In other words, the unobserved area distance attenuation map data represents a component of the fluctuation in the received power of a radio signal received at a specified point that is estimated to be derived from the distance from the specified transmitting station to the specified point.

The simple regression analysis unit 314, which uses the distance between the transmitting point and the receiving point as a variable, generates unobserved area distance attenuation map data by applying the relationship obtained by performing simple regression of the observation area radio wave map data 84 on the observation area distance map data to the unobserved area distance map data. The observation area distance map data represents the distance from a specific transmitting station to a specific point in the observation area 8. The unobserved area distance map data represents the distance from a specific transmitting station to a specific point in the unobserved area 9. FIG. 10 is a diagram showing an example of the observation area distance map data 85 and the unobserved area distance map data 95 according to one embodiment. FIG. 11 is a diagram showing an example of the unobserved area distance attenuation map data 96 according to one embodiment. The simple regression analysis unit 314 may store the generated unobserved area distance attenuation map data 96 in the storage device 4.

In step S15, the adder 315 generates unobserved area radio wave map data based on the unobserved area short-range variation map data 93 obtained in step S13 and the unobserved area distance attenuation map data 96 obtained in step S14. The unobserved area radio wave map data represents an estimate of radio wave propagation characteristics in the unobserved area 9. FIG. 12 is a diagram showing an example of unobserved area radio wave map data 97 according to one embodiment. Specifically, the adder 315 generates the unobserved area radio wave map data 97 by adding the unobserved area short-range variation map data 93 and the unobserved area distance attenuation map data 96. The adder 315 may store the generated unobserved area radio wave map data 97 in the storage device 4.

In step S16, the output device 5 outputs the unobserved area radio wave map data 97 to the outside. As an example, the output device 5 is provided with a display device (not shown), and the output device 5 displays the unobserved area radio wave map data 97 on the display device. As another example, the output device 5 is provided with a communication device (not shown), and the output device 5 may transmit the unobserved area radio wave map data 97 to an external computer via the communication device.

When step 16 is completed, the flowchart in Figure 3 ends.

By doing this, the estimation means 31 of the radio wave propagation estimation system 1 in one embodiment can extrapolate and estimate the radio wave propagation characteristics in the unobserved area 9.

As described above, the generator 312 that generates the divided short-interval variation map data in step S12 has undergone machine learning for this purpose in advance. This machine learning, that is, a method for manufacturing the generator 312 according to one embodiment, will be described below.

As shown in FIG. 13, the learning means 32 according to one embodiment includes a simple regression analysis unit 321, a subtractor 322, a data expansion unit 323, a generation unit 324, a discrimination unit 325, and a judgment unit 326. The simple regression analysis unit 321, the subtractor 322, the data expansion unit 323, the generation unit 324, the discrimination unit 325, and the judgment unit 326 are functions of the learning means 32 that are realized by the calculation device 3 executing the learning program 42.

The operation of the learning means 32 according to one embodiment, that is, the manufacturing method of the generation unit 312 according to one embodiment, will be described with reference to the flowchart in FIG. 14.

In step S21, the simple regression analysis unit 321, the subtractor 322, and the data expansion unit 323 generate divided short-interval fluctuation map data for the observation area 8 based on the observation area radio wave map data 84. Here, first, the simple regression analysis unit 321 generates the observation area distance attenuation map data based on the observation area radio wave map data 84. FIG. 15 is a diagram showing an example of the observation area distance attenuation map data 86 and the unobserved area distance attenuation map data 96 according to one embodiment. Next, the subtractor 322 generates the observation area short-interval fluctuation map data based on the observation area radio wave map data 84 and the observation area distance attenuation map data 86. Next, the data expansion unit 323 generates multiple divided short-interval fluctuation map data for the observation area 8 based on the observation area short-interval fluctuation map data.

The operation of the simple regression analysis unit 321 in step S21 will be described. The simple regression analysis unit 321 generates the observation area distance attenuation map data 86 by performing a simple regression of the observation area radio wave map data 84 on the observation area distance map data 85.

The operation of the subtractor 322 in step S21 will be described. The subtractor 322 subtracts the observation area radio wave map data 84 from the observation area distance attenuation map data 86 to generate observation area short-range fluctuation map data. Figure 16 is a diagram showing an example of the observation area short-range fluctuation map data 83 according to one embodiment.

The operation of the data expansion unit 323 in step S21 will be described. Like the data expansion unit 311 of the estimation means 31, the data expansion unit 323 divides the observation area short-interval variation map data 83 into multiple meshes, and generates divided short-interval variation map data for multiple areas, each of which is composed of multiple meshes.

In step S22, the data expansion unit 323 and the generation unit 324 generate multiple divided short-section variation map data for the observation area 8 based on the observation area structure map data 80.

The operation of the data expansion unit 323 in step S22 will be described. Similar to the data expansion unit 311 of the estimation means 31, the data expansion unit 323 divides the observation area structure map data 80 into multiple meshes, and generates multiple divided structure map data for the observation area 8, each of which is composed of multiple meshes.

The operation of the generation unit 324 in step S22 will be described. The generation unit 324 has the configuration shown in FIG. 7, similar to the generation unit 312 of the estimation means 31, and performs a process set of sequentially performing convolution calculation by the convolution calculation unit 61, spectrum normalization processing by the spectrum normalization calculation unit 62, batch normalization processing by the batch normalization calculation unit 63, and normalized linear processing by ReLU 64 five times on the input divided structure map data, and then performs convolution calculation by the convolution calculation unit 61E and hyperbolic tangent calculation processing by the Tanh calculation unit 65, and outputs the calculation results.

Note that the multiple divided short-interval variation map data in the observation area 8 generated in step S21 and the multiple divided short-interval variation map data in the observation area 8 generated in step S22 are generated based on different input data and by different calculations, and are therefore different data. For convenience, the divided short-interval variation map data in the observation area 8 generated in step S21 is referred to as the first divided short-interval variation map data, and the divided short-interval variation map data in the observation area 8 generated in step S22 is referred to as the second divided short-interval variation map data. As described below, in one embodiment, the generation unit 324 is adjusted so that the second divided short-interval variation map data generated by the generation unit 324 approaches the first divided short-interval variation map data based on the observed values of the radio wave propagation characteristics.

In step S23, the discrimination unit 325 discriminates between the first divided short-interval variation map data and the second divided short-interval variation map data in the observation area 8. The discrimination unit 325 discriminates whether each of the second divided short-interval variation map data is sufficiently close to the first divided short-interval variation map data of the corresponding area, and outputs a numerical value representing the result. As an example, when the discrimination unit 325 cannot discriminate between each of the second divided short-interval variation map data and the first divided short-interval variation map data of the corresponding area, that is, when the two are sufficiently close, it outputs "1" as the discrimination result. Conversely, when the discrimination unit 325 can discriminate between each of the second divided short-interval variation map data and the first divided short-interval variation map data of the corresponding area, that is, when the two are not sufficiently close, it outputs "-1" as the discrimination result.

As shown in FIG. 17, the discrimination unit 325 according to one embodiment may sequentially perform a plurality of processes on the input second divided short interval variation map data. In the example of FIG. 17, the discrimination unit 325 includes convolution calculation units 71A, 71B, 71C, 71D, and 71E, spectrum normalization calculation units 72A, 72B, 72C, and 72D, batch normalization calculation units 73B, 73C, and 73D, and Leaky LeRUs 74A, 74B, 74C, and 74D.

When the convolution calculation units 71A, 71B, 71C, 71D, and 71E are not differentiated from one another, they are collectively referred to as the convolution calculation unit 71. When the spectral normalization calculation units 72A, 72B, 72C, and 72D are not differentiated from one another, they are collectively referred to as the spectral normalization calculation unit 72. When the batch normalization calculation units 73B, 73C, and 73D are not differentiated from one another, they are collectively referred to as the batch normalization calculation unit 73. When the Leaky LeRUs 74A, 74B, 74C, and 74D are not differentiated from one another, they are collectively referred to as the Leaky LeRU 74.

The convolution calculation unit 71 performs a convolution calculation using a predetermined kernel on input data having a 3x3 matrix structure, similar to the convolution calculation unit 61 in FIG. 7. However, the above 3x3 is merely an example and does not limit the embodiment. The spectrum normalization calculation unit 72 performs a spectrum normalization process on the input data, similar to the spectrum normalization calculation unit 62 in FIG. 7. The batch normalization calculation unit 73 performs a batch normalization process on the input data, similar to the batch normalization calculation unit 63 in FIG. 7. The Leaky ReLU 74 performs normalized linear processing on the input data.

In the example of FIG. 17, the discrimination unit 325 according to one embodiment first performs a convolution operation, a spectrum normalization process, and a normalized linear process on the input second divided short interval variation map data. Next, the discrimination unit 325 repeats a sequential process set of a convolution operation, a spectrum normalization process, a batch normalization process, and a normalized linear process on the output of the Leaky LeRU 74A three times, and then performs a further convolution operation and outputs the result of the operation.

In step S24, the determination unit 326 determines whether the discrimination result reaches a predetermined threshold value. As an example, the determination unit 326 calculates the sum of values representing the result of discrimination performed by the determination unit 325 on all of the second divided short-interval variation map data. If the determination unit 326 determines that this sum does not reach the predetermined threshold value (No), that is, if the determination unit 326 determines that the multiple second divided short-interval variation map data generated by the generation unit 324 are not sufficiently close to the multiple first divided short-interval variation map data as a whole, the process proceeds to step S25. On the other hand, if the determination unit 326 determines that this sum has reached the predetermined threshold value (Yes), that is, if the determination unit 326 determines that the multiple second divided short-interval variation map data generated by the generation unit 324 are sufficiently close to the multiple first divided short-interval variation map data as a whole, the process proceeds to step S26.

In step S25, the determination unit 326 updates the parameters of the generation unit 324 and the discrimination unit 325. Here, the determination unit 326 updates the parameters of the generation unit 324 so that the discrimination result reaches a predetermined threshold when steps S23 and S24 are next executed. The parameters of the generation unit 324 updated here are, for example, the kernel used in the convolution operation of the convolution operation unit 61 shown in FIG. 6. In addition, the determination unit 326 updates the parameters of the discrimination unit 325 so that the discrimination result does not reach a predetermined threshold when steps S23 and S24 are next executed. The parameters of the discrimination unit 325 updated here are, for example, the kernel used in the convolution operation of the convolution operation unit 71 shown in FIG. 17. After step S25, the process proceeds to step S22, and steps S22, S23, and S24 are executed again.

In step S26, the determination unit 326 stores the generation unit 324 in the storage device 4. More specifically, the determination unit 326 stores a set of parameters that are the result of the machine learning performed by the generation unit 324 in the storage device 4. By doing so, the estimation means 31 according to one embodiment can read out this set of parameters from the storage device 4 and set it in the generation unit 312, and the generation unit 312 can perform the processing of step S12 in a state in which machine learning has been performed. When step S26 is completed, the flowchart in FIG. 14 ends.

As described above, in one embodiment of the radio wave propagation estimation system 1, the learning means 32 performs machine learning on the generation unit 324 in the manufacturing method of the generation unit 324, and the estimation means 31 including the generation unit 312 set with the results of this machine learning executes the radio wave propagation estimation method, thereby making it possible to perform extrapolation estimation of the radio wave propagation characteristics in the unobserved area 9 with higher accuracy.

As an example, the results of radio wave propagation estimation according to one embodiment were compared with the results of radio wave propagation estimation according to the related technology. Here, in the related technology, the unobserved area short-range variation map data 93 is not calculated, and the unobserved area distance attenuation map data 96 is used as the unobserved area radio wave map data 97. The root mean square error between the estimated value and the true value was 5.77 dB (decibels) in the one embodiment, whereas it was 10.82 dB in the related technology. This comparison was based on an observation experiment using the following specifications.
Observation area: Around Kudanshita Station Mesh size: 10m x 10m
Observation frequency band: 3,500MHz (megahertz)
Transmit power: 28.7 dBm
Antenna directivity: Omni Polarization: H polarization Transmitting station height: 17.5m
Height of receiving station: 1.5m

The invention made by the inventor has been specifically described above based on the embodiment, but it goes without saying that the present invention is not limited to the above embodiment and can be modified in various ways without departing from the gist of the invention. Furthermore, the features described in the above embodiment can be freely combined as long as they are not technically inconsistent.

As a variation of one embodiment, the recording medium 401 may be a non-transitory and tangible medium.

As a variation of the embodiment, the radio wave propagation estimation system 1 may further include a communication device (not shown), and the program stored in the storage device 4 may be supplied from the outside via this communication device and stored in the storage device 4.

As a modified example of an embodiment, the set of the data expansion unit 311, the generation unit 312, and the map data integration unit 313 shown in FIG. 2 may be referred to as a first estimation means. Also, the set of steps S11, S12, and S13 shown in the flowchart of FIG. 3 may be referred to as step S10. In this case, in step S10, the first estimation means generates unobserved area short-section variation map data 93 based on unobserved area structure map data 90.

As a variation of one embodiment, step S14 in the flowchart of FIG. 3 does not need to be executed after step S13, so long as it is executed before step S15. As an example, step S14 may be executed before any of steps S11 to S13, or may be executed in parallel with some or all of steps S11 to S13.

As a modified example of the embodiment, the learning means 32 may have two data extension units 323. That is, in FIG. 13, the data extension unit 323 operating in step S22 and the data extension unit 323 operating in step S21 are shown as the same data extension unit 323, but these may be two different data extension units 323. In this case, the set of the simple regression analysis unit 321, the subtractor 322, and the data extension unit 323 operating in step S21 shown in FIG. 13 may be referred to as the first learning means. At this time, in step S21 shown in the flowchart of FIG. 14, the first learning means generates the first divided short-interval fluctuation map data in the observation area 8 based on the observation area radio wave map data 84. Also, in this case, the set of the data extension unit 323 operating in step S22 shown in FIG. 13 and the generation unit 324 may be referred to as the second learning means. At this time, in step S22 shown in the flowchart of FIG. 14, the second learning means generates second divided short-interval variation map data for the observation area 8 based on the observation area structure map data 80.

REFERENCE SIGNS LIST 1 Radio wave propagation estimation system 2 Bus 3 Calculation device 31 Estimation means 311 Data expansion section 312 Generation section 313 Map data integration section 314 Simple regression analysis section 315 Adder 32 Learning means 321 Simple regression analysis section 322 Subtractor 323 Data expansion section 324 Generation section 325 Discrimination section 326 Judgment section 4 Storage device 401 Recording medium 41 Estimation program 42 Learning program 5 Output device 61A to 61F Convolution calculation section 62A to 62E Spectral normalization calculation section 63A to 63E Batch normalization calculation section 64A to 64E ReLU
65 Tanh calculation unit 71A to 71E Convolution calculation unit 72A to 72D Spectral normalization calculation unit 73B to 73D Batch normalization calculation unit 74A to 74D Leaky LeRU
8 Observation area 80 Observation area structure map data 83 Observation area short-range fluctuation map data 84 Observation area radio wave map data 85 Observation area distance map data 86 Observation area distance attenuation map data 9 Unobserved area 90 Unobserved area structure map data 91, 91A-91P Mesh 92, 92A-92C Divided structure map data 93 Unobserved area short-range fluctuation map data 95 Unobserved area distance map data 96 Unobserved area distance attenuation map data 97 Unobserved area radio wave map data

Claims (6)

a first estimation means for generating short-term fluctuation map data representing components estimated to be derived from structures among fluctuations in received power of a wireless signal transmitted from a predetermined transmitting station and received at each of a plurality of first points included in a first area based on structure map data representing planar spatial information of structures present in the first area included in the first area;
a simple regression analysis unit that generates distance attenuation map data representing a component of fluctuations in the received power received at the first points that is estimated to be derived from a distance from the predetermined transmitting station to the first points, based on second area radio wave map data obtained by observing the received power of the wireless signal at a plurality of second points included in a second area that is included in the predetermined area and different from the first area;
an adder that generates first regional radio wave map data by extrapolating fluctuations in the received power of the wireless signal received at the first locations based on the short-term fluctuation map data and the distance attenuation map data;
an output device that outputs the first regional radio wave map data to an outside device,
The first estimation means
A radio wave propagation estimation system comprising a first generation unit that performs machine learning based on a correlation between first divided short-interval variation map data of the second region based on the second region radio wave map data and second divided short-interval variation map data of the second region based on second region structure map data representing planar spatial information of structures existing in the second region, and generates the short-interval variation map data for each of the multiple first locations based on the structure map data for each of the multiple first locations. 2. The radio wave propagation estimation system according to claim 1,
The first estimation means
A data expansion unit that generates a plurality of divided structure map data corresponding to the plurality of first locations based on the unobserved area structure map data,
The first generation unit generates a plurality of divided short-range variation map data corresponding to the plurality of first locations based on the plurality of divided structure map data,
The first estimation means
The radio wave propagation estimation system further comprises a map data integration unit that generates unobserved area short-range variation map data based on the plurality of divided short-range variation map data. 3. The radio wave propagation estimation system according to claim 1,
The first generation unit further includes a learning unit for performing the machine learning,
The learning means includes:
a first learning means for generating a plurality of first divided short interval fluctuation map data each representing a fluctuation of the received power other than the attenuation of the received power resulting from a distance from the predetermined transmitting station to each of the plurality of second points, based on second regional radio wave map data obtained by observing a distribution of the received power of the wireless signal received at each of the plurality of second points;
a second learning means for generating, based on the second regional structure map data, a plurality of second divided short interval fluctuation map data each representing a fluctuation in the received power originating from the structures present at the plurality of second locations, using a first parameter;
A discrimination unit that discriminates between the plurality of first divided short interval variation map data and the plurality of second divided short interval variation map data using a second parameter;
a determination unit that updates the first parameter so that a numerical value representing the determination result reaches a predetermined threshold value until the numerical value reaches the threshold value, and updates the second parameter so that the numerical value does not reach the threshold value,
The second learning means generates the plurality of second divided short interval variation map data by using the updated first parameters,
The discrimination unit performs the discrimination using the updated second parameter,
When the result of the determination reaches the threshold, the determination unit stores the first parameter so as to be usable as a parameter of the first generation unit. The radio wave propagation estimation system according to any one of claims 1 to 3,
A radio wave propagation estimation system, wherein the planar spatial information of the structure represented by the structure map data includes at least one of the following: the altitude of the structure, the altitude of the point where the structure is located, the planar distance from the transmitting station to the receiving point where the structure is located, the average altitude from the transmitting station to the receiving point, the maximum altitude of other structures existing between the transmitting station and the receiving point, the altitude of the other structure closest to the receiving point, the planar distance from the receiving point to the other structure closest to the receiving point, and the average altitude of the other structures existing around the receiving point. generating short-term variation map data representing components estimated to be derived from the structures among fluctuations in received power of a wireless signal transmitted from a predetermined transmitting station and received at each of a plurality of first points included in the first area, based on structure map data representing planar spatial information of structures present in a first area included in a predetermined area;
generating distance attenuation map data representing a component of fluctuations in the received power received at the plurality of first points that is estimated to be derived from a distance from the predetermined transmitting station to the plurality of first points, based on second area radio wave map data obtained by observing the received power of the wireless signal at a plurality of second points included in a second area that is included in the predetermined area and different from the first area;
generating first regional radio wave map data by extrapolating fluctuations in the received power of the wireless signal received at the first locations based on the short-term fluctuation map data and the distance attenuation map data;
outputting the first regional radio wave map data to an external device;
The generating of the short-term variation map data includes:
A radio wave propagation estimation method comprising: a generation unit that performs machine learning based on a correlation between first-divided short-interval variation map data of the second region based on the second region radio wave map data and second-divided short-interval variation map data of the second region based on second region structure map data representing planar spatial information of structures existing in the second region, generating the short-interval variation map data of the first region based on the structure map data of the first region. generating a plurality of first observation area division short interval fluctuation map data each representing a fluctuation of the received power other than the attenuation of the received power resulting from a distance from a predetermined transmitting station to each of a plurality of partial areas included in the predetermined observation area, based on radio wave map data obtained by observing a distribution of the received power of a radio signal transmitted from the predetermined transmitting station received at each of a plurality of points included in the predetermined observation area;
generating, by a generating unit having a first parameter, a plurality of second observation area divided short interval variation map data each representing a variation of the received power originating from the structures existing in the plurality of partial regions, based on observation area structure map data representing planar spatial information of structures existing in the predetermined observation area;
a discrimination unit having a second parameter discriminating between the plurality of first observation area divided short interval variation map data and the plurality of second observation area divided short interval variation map data;
updating the first parameter so that a numerical value representing the result of the discrimination reaches a predetermined threshold value until the numerical value reaches the threshold value, generating the plurality of second observation area divided short interval variation map data using the updated first parameter, updating the second parameter so that the numerical value does not reach the threshold value, and performing the discrimination using the updated second parameter;
and when the result of the discrimination reaches the threshold, storing the first parameter as a parameter of the generation unit that has performed machine learning so that it can be used by an external estimation means.

Images