Learning method, propagation characteristic estimation method, propagation characteristic estimation device, propagation characteristic estimation system, propagation characteristic estimation program

Abstract

A method for training a machine learning model that estimates a propagation characteristic between a transmitting station and a receiving station in a wireless communication system, the training method including: executing an input image generation process to generate an input image in accordance with each of one or a plurality of receiving-station positions, which are supplied as training data from map data; accepting the input image as an input to the machine learning model and acquiring an output pertaining to the result of estimating a propagation characteristic; and updating parameters of the machine learning model on the basis of the output of the machine learning model. The input image generation process includes a process for generating target range height data that supplies a distribution of structure heights in an estimation target range for which a propagation characteristic is to be estimated, and a process for reading in some of the target range height data to thereby generate an input image for each of the one or plurality of receiving-station positions.

Description

Technical Field

[0001] The present invention relates to a technique for estimating propagation characteristics between a transmitting station and a receiving station in a wireless communication system.

Background Art

[0002] Techniques for estimating propagation characteristics (propagation loss) between a transmitting station (Transmitter: Tx) and a receiving station (Receiver: Rx) in a wireless communication system are known. For example, Patent Document 1 and Patent Document 2 disclose such propagation characteristic estimation techniques.

[0003] In particular, Patent Document 2 discloses a technique for estimating propagation characteristics using a machine learning model. More specifically, an image representing the distribution of building heights around the Rx is prepared. By using the image (input image) as the input to a convolutional neural network (CNN), feature amounts of the input image are extracted. Then, by inputting the extracted feature amounts into a fully-connected neural network (FNN), the propagation characteristics are estimated. Similarly, learning of the machine learning model is performed so as to optimize the estimation result with the input image as the input to the machine learning model.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] As described above, a technique for estimating propagation characteristics by using, as an input to a machine learning model, an input image representing the distribution of building heights around an Rx is known. Here, it is conceivable to generate, according to the position of the Rx, an input image representing the distribution of building heights around the Rx from map data including information on the positions and heights of buildings on the map. Thereby, there is an advantage that the position of the Rx may be input when estimating the propagation characteristics. Also, when training the machine learning model, the position of the Rx may be given as training data. However, conventionally, there has been a problem that the amount of processing becomes enormous when generating an input image from map data.

[0006] One object of the present disclosure is to provide a technique capable of generating an input image from map data at high speed.

Means for Solving the Problem

[0007] A first aspect relates to a method for training a machine learning model for estimating propagation characteristics between a transmitting station and a receiving station in a wireless communication system. Here, the machine learning model includes a feature extraction layer that extracts features by using, as an input, an input image giving the distribution of the heights of structures around the receiving station, and an estimation layer that estimates the propagation characteristics by using at least the features as an input. The learning method according to the first aspect includes setting learning data including one or more receiver positions, executing an input image generation process for generating the input image according to each of the one or more receiver positions from map data including information on the structure on the map, obtaining the output of the machine learning model with the input image generated by the input image generation process as the input to the feature extraction layer, and updating the parameters of the machine learning model based on the output of the machine learning model. Here, the input image generation process includes a process of specifying an estimation target range on the map that is the target of the estimation of the propagation characteristics based on the learning data, a process of generating target range height data that gives the distribution of the heights of the structures in the estimation target range from the map data, and a reading process of generating the input image by reading a part of the target range height data for each of the one or more receiver positions.

[0008] The second aspect relates to a propagation characteristic estimation method for estimating the propagation characteristics between a transmitter and a receiver in a wireless communication system. The propagation characteristic estimation method according to the second aspect includes setting estimation data including one or more receiver positions, executing an input image generation process for generating an input image that gives the distribution of the heights of the structures around the receiver according to each of the one or more receiver positions from map data including information on the structures on the map, and estimating the propagation characteristics by inputting the input image into a machine learning model. Here, the machine learning model includes a feature extraction layer that extracts feature amounts with the input image as the input, and an estimation layer that estimates the propagation characteristics with at least the feature amounts as the input. Further, the input image generation process includes a process of specifying an estimation target range on the map that is the target of the estimation of the propagation characteristics based on the estimation data, a process of generating target range height data that gives the distribution of the heights of the structures in the estimation target range from the map data, and a reading process of generating the input image by reading a part of the target range height data for each of the one or more centroid positions.

[0009] The third aspect relates to a propagation characteristic estimation apparatus that estimates propagation characteristics between a transmitting station and a receiving station in a wireless communication system. The propagation characteristic estimation apparatus according to the third aspect is configured to execute a process of acquiring estimation data including one or more receiving station positions, an input image generation process of generating an input image that gives a distribution of heights of structures around the receiving station according to each of the one or more receiving station positions from map data including information on structures on a map, and a process of estimating the propagation characteristics by inputting the input image into a machine learning model. Here, the machine learning model includes a feature extraction layer that extracts features by using the input image as an input, and an estimation layer that estimates the propagation characteristics by using at least the features as an input. Further, the input image generation process includes a process of specifying an estimation target range on the map that is a target of estimation of the propagation characteristics based on the estimation data, a process of generating target range height data that gives a distribution of heights of the structures in the estimation target range from the map data, and a process of generating the input image by reading a part of the target range height data for each of the one or more receiving station positions.

[0010] The fourth aspect relates to a propagation characteristic estimation system that estimates propagation characteristics between a transmitting station and a receiving station in a wireless communication system. The propagation characteristic estimation system according to the fourth aspect includes one or more processors and a data server that manages map data including information on structures on a map as a map database. The one or more processors are configured to execute a process of acquiring estimation data including one or more receiver positions, an input image generation process of generating an input image that gives a distribution of heights of structures around the receiver according to each of the one or more receiver positions from the map data, and a process of estimating the propagation characteristics by inputting the input image into a machine learning model. Here, the machine learning model includes a feature extraction layer that extracts features using the input image as an input, and an estimation layer that estimates the propagation characteristics using at least the feature extraction layer as an input. Further, the input image generation process includes a process of specifying an estimation target range on the map that is a target of estimation of the propagation characteristics based on the estimation data, a process of generating target range height data that gives a distribution of heights of the structures in the estimation target range from the map data, and a process of generating the input image by reading a part of the target range height data for each of the one or more receiver positions.

[0011] The fifth aspect relates to a propagation characteristic estimation program for estimating propagation characteristics between a transmitter and a receiver in a wireless communication system. The propagation characteristic estimation program according to the fifth aspect causes a computer to execute a process of acquiring estimation data including one or more receiver positions, an input image generation process of generating an input image that gives a distribution of heights of the structures around the receiver according to each of the one or more receiver positions from map data including information on structures on the map, and a process of estimating the propagation characteristics by inputting the input image into a machine learning model. Here, the machine learning model includes a feature extraction layer that extracts features using the input image as an input, and an estimation layer that estimates the propagation characteristics using at least the features as an input. Further, the input image generation process includes a process of specifying an estimation target range on the map that is a target of estimation of the propagation characteristics based on the estimation data, a process of generating target range height data that gives a distribution of heights of the structures in the estimation target range from the map data, and a process of generating the input image by reading a part of the target range height data for each of the one or more receiver positions.

Effect of the Invention

[0012] According to the present invention, target range height data is generated in the input image generation process. Then, for each of one or more receiving stations given by learning data or estimation data, an input image is generated by reading a part of the target range height data. Thereby, an input image can be generated at high speed. Consequently, it is possible to achieve the effects of improving the calculation speed and reducing the processing load.

Brief Description of the Drawings

[0013]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Embodiments for Carrying Out the Invention

[0014] Embodiments of the present invention will be described with reference to the accompanying drawings.

[0015] Consider the estimation of propagation characteristics (propagation loss) between a transmitter (Tx) and a receiver (Rx) in a wireless communication system. For example, the transmitter is a base station (BS), and the receiver is a mobile station (MS).

[0016] 1. Overview The overview of the propagation characteristic estimation apparatus according to the present embodiment will be described. In the propagation characteristic estimation apparatus according to the present embodiment, propagation characteristics are estimated using a machine learning model based on a convolutional neural network (CNN). FIG. 1 is a conceptual diagram for explaining the overview of the machine learning model used in the propagation characteristic estimation apparatus according to the present embodiment. As is well known, CNN is a useful tool that can automatically extract feature amounts from images. CNN has a structure in which convolutional layers and pooling layers are repeatedly arranged. CNN can also be referred to as a "feature extraction layer".

[0017] A fully-connected neural network (FNN) is arranged at the subsequent stage of the CNN. The FNN outputs the propagation loss L by taking as input at least the feature quantities extracted by the CNN. The FNN can also be referred to as an “estimation layer” regarding propagation characteristics. Note that the FNN may be configured to take as input system parameters related to the wireless communication system together with the feature quantities extracted by the CNN. Examples of the system parameters include the frequency of the transmitted radio wave, the antenna height of the Tx, the antenna height of the Rx, and the like.

[0018] In the propagation characteristic estimation according to the present embodiment, an overhead image (for example, colored according to height) giving the height distribution of structures such as buildings is used as input image 1 (input to the machine learning model) to the CNN. The overhead image is a two-dimensional image viewed from above and can also be referred to as an aerial view image. In particular, in order to consider the radio wave propagation around the Rx, as input image 1, an image giving the height distribution of structures in a predetermined range around the Rx is used. Here, the predetermined range is, for example, a range on a rectangle centered on the Rx. In the propagation characteristic estimation according to the present embodiment, such input image 1 is input to the CNN, and the feature quantities of input image 1 are extracted.

[0019] Here, input image 1 giving the height of structures around the Rx is generally managed as matrix data. In this case, each component of the matrix corresponds to each position in a predetermined range around the Rx, and the value of each component gives the height of the structure at the corresponding position. Here, each position corresponding to each component of the matrix is, for example, one cell when the predetermined range around the Rx is divided into a grid.

[0020] With reference to FIG. 2, an example of the case where input image 1 is managed as matrix data will be described. In the example shown in FIG. 2, input image 1 is a range on a rectangle around the Rx. Also, the range given by input image 1 includes structure B1 and structure B2. Here, the ranges of structure B1 and structure B2 are indicated by solid lines.

[0021] Figure 2 shows how the range given by the input image 1 is divided into a grid. By dividing into a grid in this way, the position of each cell can be made to correspond to each component of the matrix. For example, in the example shown in Figure 2, the position of the cell at the upper left corner of the input image 1 is made to correspond to the component in the first row and first column of the matrix (hereinafter, the component in the m-th row and n-th column of the matrix is represented as the "(m,n) component"), and the position of the cell in the downward direction of the drawing is specified by the row number, and the position of the cell in the rightward direction of the drawing is specified by the column number so as to correspond. In this case, position P1 corresponds to the (3,3) component, position P2 corresponds to the (5,4) component, and position P3 corresponds to the (6,5) component, respectively.

[0022] And when the position of the cell is included in the range of the structure, by substituting the height of the structure into the component of the matrix corresponding to the position of the cell, the input image 1 can be managed as matrix data. For example, since position P1 is included in structure B1, the height of structure B1 is substituted into the (3,3) component of the matrix. Also, since position P3 is included in structure B2, the height of structure B2 is substituted into the (6,5) component of the matrix. On the other hand, since position P2 is not included in any structure, no substitution is performed.

[0023] In the following description, it is assumed that the input image 1 is managed as such matrix data. However, the matrix data may be given as array data when inputting to arithmetic processing or a machine learning model.

[0024] 2. Example of functional configuration Hereinafter, with reference to FIGS. 3 and 4, an example of the functional configuration of the propagation characteristic estimation device 100 according to the present embodiment will be described.

[0025] First, with reference to FIG. 3, an example of the functional configuration of the propagation characteristic estimation device 100 in the learning stage will be described. The propagation characteristic estimation device 100 in the learning stage includes, as functional blocks, an input image generation unit 110, a system parameter generation unit 120, a model unit 130, an error calculation unit 140, and a model update unit 150. The model unit 130 includes, as a machine learning model, a CNN 131 and an FNN 132.

[0026] In the learning stage, first, learning data is set. The learning data includes at least the positions of one or more Rx (receiver positions) that are the targets for estimating propagation characteristics. The learning data also includes data for evaluating the output results of the machine learning model with respect to the learning data (data that is regarded as the correct answer for the input), for example, including the transmission power of the Tx, the reception power of the Rx, propagation loss, etc. These can use values obtained from measured values or simulation data by ray tracing. In addition, as learning data that is a system parameter, the position of the Tx, the frequency of the transmitted radio wave, the antenna height of the Tx, the antenna height of the Rx, etc. may be set. Note that the learning data may similarly include test data for evaluating the learning state and generalization performance of the machine learning model and verification data for adjusting hyperparameters.

[0027] The input image generation unit 110 executes a process of generating the input image 1 from the map data according to each of the one or more receiver positions given in the learning data (hereinafter, also referred to as "input image generation process"). Here, the input image generation unit 110 acquires information on the positions and heights of structures on the map by designating a specific range as the map data. Here, the map data can be acquired from a map database that manages information on the positions and heights of structures on the map. The input image 1 generated in the input image generation unit 110 serves as the input to the CNN 131.

[0028] The propagation characteristic estimation device 100 according to the present embodiment is characterized by the input image generation process. Details of the input image generation process executed in the propagation characteristic estimation device 100 according to the present embodiment will be described later.

[0029] The system parameter generation unit 120 executes a process of generating system parameters that are input to the FNN 132 based on the learning data. That is, it executes a process of converting the learning data related to the system parameters so as to be input to the FNN 132. For example, it executes a process of converting the learning data into array data. Note that when system parameters are not given as the input to the FNN 132, the system parameter generation unit 120 may be configured not to be included as a functional block.

[0030] The model unit 130 takes the input image 1 generated by the input image generation unit 110 and the system parameters generated by the system parameter generation unit 120 as inputs, and outputs an estimation result of the propagation characteristics. More specifically, the CNN 131 extracts feature amounts with the input image 1 as the input, and the FNN 132 outputs an estimation result of the propagation characteristics with the feature amounts of the output of the CNN 131 and the system parameters as inputs. Here, the CNN 131 and the FNN 132 may adopt a suitable configuration according to the environment to which the propagation characteristic estimation apparatus 100 according to the present embodiment is applied.

[0031] The error calculation unit 140 acquires the estimation result output from the model unit 130. Then, the error calculation unit 140 calculates the estimation error by the model unit 130 with reference to the learning data. For example, when the model unit 130 outputs the propagation loss as the estimation result, the error calculation unit 140 calculates the estimation error based on the difference between the received power calculated from the estimated propagation loss and the received power given as the correct data in the learning data. Alternatively, the error calculation unit 140 calculates the value of a predetermined loss function as the estimation error.

[0032] The model update unit 150 updates the parameters of the machine learning model included in the model unit 130 until the estimation error converges to a certain level or below. Typically, the model update unit 150 updates the parameters of the machine learning model so that the estimation error becomes smaller by the gradient descent method using the error backpropagation method. Examples of the parameters of the machine learning model include the filter parameters and bias values related to the CNN 131, and the weight parameters and bias values related to the FNN 132.

[0033] Next, referring to FIG. 4, a functional configuration example of the propagation characteristic estimation apparatus 100 in the propagation characteristic estimation stage will be described. The propagation characteristic estimation apparatus 100 in the propagation characteristic estimation stage includes, as functional blocks, an input image generation unit 110, a system parameter generation unit 120, a model unit 130, and a result output unit 160.

[0034] In the propagation characteristic estimation stage, first, estimation data is set. The estimation data provides data for which the propagation characteristics are to be estimated. Generally, the estimation data is data having the same content as the data given as learning data in the learning stage (excluding the data that is considered correct for the input). That is, the estimation data includes at least one or more receiving station positions for which the propagation characteristics are to be estimated. Also, the estimation data may include data that becomes system parameters.

[0035] The input image generation unit 110 and the system parameter generation unit 120 are the same as the functional blocks in the learning stage. The model unit 130 has been learned through the learning stage. The learned model unit 130 takes the input image 1 generated by the input image generation unit 110 and the system parameters generated by the system parameter generation unit 120 as inputs and outputs an estimation result of the propagation characteristics. The result output unit 160 stores the data of the estimation result output from the model unit 130 in a storage device or presents it to the user.

[0036] 3. Input Image Generation Process The propagation characteristic estimation apparatus 100 according to the present embodiment is characterized by the input image generation process executed in the input image generation unit 110. Hereinafter, details of the input image generation process executed in the propagation characteristic estimation apparatus 100 according to the present embodiment will be described.

[0037] First, as a comparative example, the conventional input image generation process will be described. FIG. 5 is a conceptual diagram for explaining the outline of the input image generation process according to the comparative example.

[0038] In the input image generation process according to the comparative example, first, a range 2 on the map (hereinafter also referred to as the "estimation target range 2") that is the target of propagation characteristic estimation as shown in Fig. 5(A) is specified. The estimation target range 2 is specified based on the set learning data or estimation data. For example, a range including one or more transmitter positions and one or more receiver positions given as learning data or estimation data is specified as the estimation target range 2.

[0039] As shown in Fig. 5(A), it is assumed that the estimation target range 2 includes a plurality of structures Bj (j = 1, 2, ···, Nb). Here, the subscript j of the symbol is given to distinguish each of the plurality of structures. Therefore, Nb is an integer and indicates the number of structures Bj included in the estimation target range 2.

[0040] Next, in the input image generation process according to the comparative example, map data (including position and height information) of each of the plurality of structures Bj included in the estimation target range 2 is acquired. That is, Nb pieces of map data for each of the plurality of structures Bj are acquired.

[0041] Then, in the input image generation process according to the comparative example, the input image 1 is generated as follows according to each of the one or more receiver positions given as learning data or estimation data. Now, consider the components of the matrix corresponding to the position Pi of the cell as shown in Fig. 5(B). Here, the subscript i (i = 1, 2, ···, N) of the symbol is given to distinguish each of the cell positions. Also, N indicates the total number of cells represented by the input image 1, and by extension, the number of matrix components. At this time, in the input image generation process according to the comparative example, for each of the plurality of structures Bj (j = 1, 2, ···, Nb), a determination as to whether the cell position Pi is included in the range of the structure Bj (hereinafter also referred to as "inside / outside determination") is made. And when the cell position Pi is included in the range of the structure Bj, the height of the structure Bj is substituted into the component of the matrix corresponding to the cell position Pi.

[0042] Therefore, in the input image generation process according to the comparative example, in order to generate one input image 1, it is necessary to perform inside / outside determination for all components of the matrix corresponding to each cell position Pi (i = 1, 2, ···, N) with respect to a plurality of structures Bj (j = 1, 2, ···, Nb). That is, in order to generate one input image 1, a maximum of N × Nb search points are required.

[0043] FIG. 6 shows the process executed when generating one input image 1 in the input image generation process according to the comparative example. The flowchart shown in FIG. 6 starts, for example, when generating the input image 1 according to a certain receiving station position. Then, it is repeatedly executed for each of one or more receiving station positions given as learning data or estimation data. With reference to FIG. 6, the specific process of the input image generation process according to the comparative example will be described.

[0044] First, in step S100, the components of the matrix to be substituted are initialized, and in step S101, the structure Bj to be the target of inside / outside determination is initialized. Then, in step S102, the inside / outside determination is performed.

[0045] When the result of the inside / outside determination is affirmative (step S102; Yes), the height of the structure Bj that is the target of the inside / outside determination is substituted into the component of the matrix that is the target of substitution (step S103), and the process proceeds to step S106. When the result of the inside / outside determination is negative (step S102; No), the process proceeds to step S104.

[0046] In step S104, it is determined whether the inside / outside determination has been performed for all of the plurality of structures Bj for which the map data has been acquired. If the inside / outside determination has not been performed for all of the plurality of structures Bj yet (step S104; Yes), j is incremented (step S105), and the inside / outside determination (step S102) is performed again. If the inside / outside determination has been performed for all of the plurality of structures Bj (step S104; No), the process proceeds to step S106.

[0047] In step S106, it is determined whether the inside / outside determination for a plurality of structures Bj has been performed for all components of the matrix. If the inside / outside determination for a plurality of structures Bj has not been performed yet for all components of the matrix (step S106; Yes), increment i (step S107), return to step S101 again, and repeat the process. If the inside / outside determination for a plurality of structures Bj has been performed for all components of the matrix (step S106; No), the generation process of one input image 1 is terminated.

[0048] By referring to FIG. 6, it can also be understood that in the input image generation process according to the comparative example, in order to generate one input image 1, a maximum of N×Nb search points are required. In particular, when generating the input image 1 for each of one or a plurality of receiving station positions, the number of search points becomes even more enormous. For example, assuming that the number of cells when the estimation target range 2 is divided into grids in the same manner as the input image 1 is Np, consider the case where a receiving station position is given to each position of the cells in the estimation target range 2. In this case, since Np input images 1 are to be generated, the number of search points is at most Np×N×Nb. Here, assuming that the distribution of the structure Bj is uniform in the estimation target range 2 and Nb = (1 / α)·Np (α is a constant indicating the density of the distribution), the number of search points is (1 / α)·N×Np^2. That is, as the estimation target range 2 expands, the number of search points increases in an order of square. The increase in the number of search points becomes a factor in the decrease in the calculation speed and the increase in the processing load. In a radio wave environment that changes every moment, it is not desirable to require a lot of time for the estimation of propagation characteristics. Also, the high processing load is not desirable from the cost aspect and practicality.

[0049] The input image generation process according to the present embodiment can generate the input image 1 at high speed compared to the comparative example. Hereinafter, the input image generation process according to the present embodiment will be described. FIG. 7 is a conceptual diagram for explaining the outline of the input image generation process according to the present embodiment.

[0050] First, similar to the comparative example, in the input image generation process according to the present embodiment, an estimation target range 2 as shown in FIG. 7(A) is specified. Next, in the input image generation process according to the present embodiment, target range height data 3 that gives the height distribution of the structure Bj in the estimation target range 2 is generated. The target range height data 3 can be managed as matrix data in the same manner as the input image 1. That is, the estimation target range 2 is divided into grids, and the position Qk (k = 1, 2, ···, Np) of each cell is made to correspond to each component of the matrix. Then, the value of each component of the matrix is set as the height of the structure at the position Qk of each corresponding cell. However, the grid size related to the target range height data 3 may be given so as to be different from the grid size related to the input image 1.

[0051] Referring to FIG. 7(B), the generation of the target range height data 3 in the input image generation process according to the present embodiment will be described. First, map data (including position and height information) of each of the plurality of structures Bj included in the estimation target range 2 is acquired. Next, referring to the map data of each of the plurality of structures Bj in sequence, the height of the structure Bj is substituted into the component of the matrix corresponding to the position Qk of the cell included in the range of the structure Bj. As a result, by referring to the map data of all the plurality of structures Bj, the generation of the target range height data 3 is completed. In particular, in such generation, the number of search points in the substitution process performed by referring to one structure Bj is sufficient for the search from the minimum coordinate to the maximum coordinate of the position information indicating the range of the structure Bj in the map data. Therefore, generally considering that the plurality of structures Bj do not interfere with each other, it can be considered that the number of search points until the generation of the target range height data 3 is completed is at most about a constant multiple of Np.

[0052] FIG. 8 shows the process executed when generating the target range height data 3 in the input image generation process according to the present embodiment. Referring to FIG. 8, the specific process executed when generating the target range height data 3 will be described.

[0053] First, in step S200, the structure Bj to be referenced is initialized. After step S200, the process proceeds to step S201.

[0054] In step S201, the height of the structure Bj is substituted into the components of the corresponding matrix at the positions Qk of the cells included in the range of the structure Bj. After step S201, the process proceeds to step S202.

[0055] In step S202, it is determined whether references have been made for all of the plurality of structures Bj for which map data has been acquired. If references have not yet been made for all of the plurality of structures Bj (step S202; Yes), the increment of j is performed (step S203), and the process returns to step S201 again to repeat the process. If references have been made for all of the plurality of structures Bj (step S202; No), the process ends.

[0056] After generating the target range height data 3 in this way, in the input image generation process according to the present embodiment, the input image 1 is generated by reading a part of the target range height data 3 (see (B) of FIG. 7). This can be done by specifying the receiving station position and the range of the input image 1 (the range around Rx).

[0057] FIG. 9 shows the process executed when generating one input image 1 in the input image generation process according to the present embodiment. The flowchart shown in FIG. 9 starts, for example, when generating the input image 1 according to a certain receiving station position. With reference to FIG. 9, the specific process executed when generating one input image 1 in the input image generation process according to the present embodiment will be described.

[0058] First, in step S300, the components of the matrix to be substituted are initialized. After step S300, the process proceeds to step S301.

[0059] In step S301 (loading process), the target range height data 3 at the position corresponding to the component of the matrix to be substituted is loaded. After step S301, the process proceeds to step S302.

[0060] In step S302, it is determined whether the target range height data 3 has been loaded for all components of the matrix. If the target range height data 3 has not been loaded for all components of the matrix yet (step S302; Yes), i is incremented (step S303), and the process returns to step S301 again to repeat the process. If the target range height data 3 has been loaded for all components of the matrix (step S302; No), the generation process of one input image 1 ends.

[0061] Thus, in the input image generation process according to this embodiment, the number of search points until the generation of one input image 1 is completed is N. Here, in the input image generation process according to this embodiment, even when generating the input image 1 according to each of a plurality of receiving station positions given by the set learning data or estimation data, the input image 1 can be generated by loading a part of the same target range height data 3. That is, it is sufficient to generate the target range height data 3 once for the set learning data or estimation data.

[0062] From the above description, in the input image generation process according to this embodiment, considering the case where the receiving station position is given to each cell position of the estimation target range 2, the number of search points is at most Np×N + β·Np (β is a constant). Generally, considering that β << N, the number of search points is approximately Np×N, which is 1 / Nb times that of the comparative example.

[0063] Let's consider this with a specific example. Assume that the target estimation range 2 is in the range of 1000m × 1000m. Also, assume that the distribution of the structure Bj is uniform within the target estimation range 2, with 50 per 100m square, and the average size of the structure Bj is 10m square. Further assume that the position information of the structure Bj in the map data is given in a grid every 0.5m. Also, assume that the grid of the input image 1 and the target estimation range 2 is given every 1m, and the size of the matrix related to the input image 1 is 64×64. At this time, Nb = 5000, Np = 1×10^6, and N = 4096. Therefore, in the input image generation process according to the comparative example, the maximum number of search points is Np×N×Nb = 2.048×10^13.

[0064] On the other hand, since the range of the structure Bj is composed of 20×20 cells, the number of search points until the generation of the target range height data 3 is completed in the input image generation process according to the present embodiment is 20×20×Nb = 2×10^6. And in the input image generation process according to the present embodiment, the maximum number of search points until the generation of the input image 1 is completed is Np×N = 4.096×10^9. This is much larger than the number of search points until the generation of the target range height data 3 is completed. Therefore, in the input image generation process according to the present embodiment, the maximum number of search points is approximately Np×N = 4.096×10^9, which is approximately 1 / Nb times that of the comparative example.

[0065] In this way, the input image generation process according to the present embodiment can generate the input image 1 at high speed compared to the comparative example. Consequently, it can achieve the effects of improving the calculation speed and reducing the processing load.

[0066] Furthermore, in the input image generation process according to the present embodiment, the reading of the target range height data 3 for each component of the matrix related to the input image 1 (step S301 in FIG. 9) can be performed in parallel and efficiently by matrix processing using affine transformation. With reference to FIG. 10, the reading of the target range height data 3 by matrix processing using affine transformation will be described.

[0067] Consider reading the target range height data 3 for the components of the matrix related to the input image 1 corresponding to the position P1 shown in (A) of FIG. 10. Here, assume that the coordinates of the position P1 can be represented as (X0, Y0) in the coordinates on the input image 1 (hereinafter referred to as "reference coordinates"). The reference coordinates are points from -2 / d to 2 / d.

[0068] Next, calculate the coordinates (Xout, Yout) in the position coordinate system of the position P1 (the coordinate system defined by the x-axis and y-axis shown in (B) of FIG. 10) by affine transformation. At this time, considering the case where the input image 1 is rotated by the azimuth angle θ with respect to the estimated target range 2, the affine matrix M related to the affine transformation can be expressed by the following equation (1). Here, (xp, yp) are the coordinates in the position coordinate system of the center point P0 of the input image 1.

[0069]

Equation

[0070] Therefore, the coordinates (Xout, Yout) in the position coordinate system of the position P1 can be calculated by the following equation (2).

[0071]

Equation

[0072] Next, the conversion to the coordinate system (u, v) (hereinafter also referred to as "image coordinate system") for referring to the target range height data 3 corresponding to the position P1 can be calculated by the following equation (3). Here, (Xin, Yin) are the coordinates of the structure Bj of the estimated target range 2, and dout is the difference in the grid resolution between the input image 1 and the target range height data 3.

[0073]

Equation

[0074] Then, by referring to the target range height data 3 based on the coordinates in the image coordinate system, it is possible to read the target range height data 3 for the components of the matrix related to the input image 1 corresponding to the position P1.

[0075] Such matrix processing using an affine transformation can be performed independently for each of the components of the matrix related to the input image 1. That is, the reading of the target range height data 3 for each of the components of the matrix related to the input image 1 can be performed in parallel. By performing parallel processing for the reading of the target range height data 3 in this way, the input image generation process according to the present embodiment can be further accelerated.

[0076] In the input image process according to the comparative example as well, it is possible to consider performing parallel processing for each of the components of the matrix related to the input image 1. However, since it is necessary to sequentially perform inside / outside determination for each of the plurality of structures Bj and the processing load increases according to the number of structures Bj for which map data has been acquired, it is not suitable for parallel processing. On the other hand, in the input image process according to the present embodiment, matrix processing using the above-described affine transformation is possible, and in this case, the processing of reading the target range height data 3 is completed in several steps as described above. Thus, in the input image process according to the present embodiment, it is possible to configure to efficiently perform parallel processing by matrix processing using an affine transformation.

[0077] 4. Configuration example of propagation characteristic estimation device FIG. 11 is a block diagram showing a configuration example of a propagation characteristic estimation device 100 according to the present embodiment. The propagation characteristic estimation device 100 includes one or more processors 103 (hereinafter simply referred to as "processor 103"), one or more storage devices 104 (hereinafter simply referred to as "storage device 104"), a user interface 101, and an I / O interface 102.

[0078] Processor 103 performs various information processes. For example, processor 103 includes a CPU (Central Processing Unit). The storage device 104 stores various information necessary for the processes by processor 103. Examples of the storage device 104 include a volatile memory, a non-volatile memory, an HDD (Hard Disk Drive), an SSD (Solid State Drive), and the like.

[0079] The propagation characteristic estimation program 105 is a computer program executed by processor 103. By processor 103 executing the propagation characteristic estimation program 105, the functions of processor 103 (propagation characteristic estimation device 100) are realized. That is, by processor 103 executing the propagation characteristic estimation program 105, the functional configuration of the propagation characteristic estimation device 100 shown in FIGS. 3 and 4 is realized. The propagation characteristic estimation program 105 is stored in the storage device 104. The propagation characteristic estimation program 105 may be recorded on a computer-readable recording medium. The propagation characteristic estimation program 105 may be provided to the propagation characteristic estimation device 100 via a network.

[0080] Processor 103 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array).

[0081] The user interface 101 provides information to the user and also receives information input from the user. The user interface 101 includes an input device and a display device.

[0082] The I / O interface 102 is communicably connected to the data server 200. The data server 200 manages at least the position and height information of the structures on the map as the map database 201. And the I / O interface 102 is configured to communicate with the data server 200 and acquire desired map data from the map database 201. The processor 103 can acquire necessary information from the database via the I / O interface 102.

[0083] Note that FIG. 11 can also be considered to show a propagation characteristic estimation system composed of the propagation characteristic estimation device 100 and the data server 200.

[0084] FIG. 12 is a flowchart schematically showing the processing by the propagation characteristic estimation device 100 according to the present embodiment. In particular, FIG. 12 shows the processing related to the learning of the estimation model.

[0085] In step S400, the propagation characteristic estimation device 100 specifies the estimation target range 2 according to the set learning data, and acquires the map data of each of the plurality of structures Bj included in the estimation target range 2 from the map database 201.

[0086] In step S401, the propagation characteristic estimation device 100 executes an input image generation process.

[0087] In step S402, the propagation characteristic estimation device 100 estimates the propagation characteristics using a machine learning model. More specifically, the propagation characteristic estimation device 100 extracts feature amounts by inputting the input image 1 generated in step S401 to the CNN 131. Further, the propagation characteristic estimation device 100 estimates the propagation characteristics by inputting the feature amounts of the output of the CNN 131 and the system parameters to the FNN 132.

[0088] In step S403, the propagation characteristic estimation device 100 calculates an estimation error by comparing the estimation result of the propagation characteristic with the received power given as the correct answer in the learning data.

[0089] In step S404, the propagation characteristic estimation device 100 determines whether the estimation error has converged below a certain level. If the estimation error exceeds the certain level (step S404; No), the process proceeds to step S405.

[0090] In step S405, the propagation characteristic estimation device 100 updates the parameters of the machine learning model so that the estimation error decreases. Then, the process returns to step S401.

[0091] If the estimation error has converged below a certain level (step S404; Yes), the learning of the machine learning model is completed.

[0092] In the propagation characteristic estimation stage, steps S400 to S402 are the same. However, in step S402, the propagation characteristic estimation device 100 estimates the propagation characteristic using the learned machine learning model. Then, the propagation characteristic estimation device 100 stores the estimation result of the propagation characteristic in the storage device 104. Also, the propagation characteristic estimation device 100 presents the estimation result of the propagation characteristic to the user via the user interface 101.

[0093] 5. Effects As described above, according to the present embodiment, the target range height data 3 is generated in the input image processing. Then, for each of one or a plurality of receiving stations given by the set learning data or estimation data, the input image 1 is generated by reading a part of the target range height data 3 around Rx. Thereby, the input image 1 can be generated at high speed. Consequently, the effects of improving the calculation speed and reducing the processing load can be achieved.

[0094] Furthermore, according to the present embodiment, parallel processing is possible for reading the target range height data 3 for each component of the matrix related to the input image 1 by matrix processing using an affine transformation. Thereby, the input image generation process according to the present embodiment can be further accelerated.

Explanation of Signs

[0095] 1 Input image 2 Estimation target range 3 Target range height data 100 Propagation characteristic estimation device 101 User interface 102 I / O interface 103 Processor 104 Storage device 105 Propagation characteristic estimation program 110 Input image generation unit 120 System parameter generation unit 130 Model unit 131 CNN (Feature extraction layer) 132 FNN (Estimation layer) 140 Error calculation unit 150 Model update unit 160 Result output unit 200 Data server 201 Map database Bj Structure

Claims

1. A learning method for a machine learning model that estimates propagation characteristics between a transmitting station and a receiving station in a wireless communication system, The machine learning model includes a feature extraction layer that extracts features by taking as input an input image giving a distribution of heights of structures around the receiving station, and an estimation layer that estimates the propagation characteristics by taking at least the features as input. The learning method includes setting learning data including one or more receiving station positions, performing an input image generation process of generating the input image according to each of the one or more receiving station positions from map data including information on the structures on a map, obtaining an output of the machine learning model by using the input image generated in the input image generation process as input to the feature extraction layer, updating parameters of the machine learning model based on the output of the machine learning model. The input image generation process includes a process of specifying an estimation target range on the map that is the target of estimation of the propagation characteristics based on the learning data, a process of generating target range height data giving a distribution of heights of the structures in the estimation target range from the map data, a reading process of generating the input image by reading a part of the target range height data for each of the one or more receiving station positions. A learning method characterized by the above.

2. The learning method according to claim 1, The reading process includes performing in parallel by matrix processing the reading of the target range height data for each position around the receiving station. A learning method characterized by the above.

3. A propagation characteristic estimation method for estimating propagation characteristics between a transmitting station and a receiving station in a wireless communication system, setting estimation data including one or more receiving station positions, Executing an input image generation process for generating an input image that gives a distribution of heights of the structures around the receiving station according to each of the one or more receiving station positions from map data including information on the structures on the map; Estimating the propagation characteristics by inputting the input image into a machine learning model, The machine learning model includes a feature extraction layer that extracts features using the input image as an input, and an estimation layer that estimates the propagation characteristics using at least the features as an input. The input image generation process includes a process of specifying an estimation target range on the map that is the target of the estimation of the propagation characteristics based on the estimation data, a process of generating target range height data that gives a distribution of heights of the structures in the estimation target range from the map data, and a reading process of generating the input image by reading a part of the target range height data for each of the one or more receiving station positions. A propagation characteristic estimation method characterized by the above.

4. The propagation characteristic estimation method according to claim 3, The reading process includes parallelly reading the target range height data for each position around the receiving station by matrix processing. A propagation characteristic estimation method characterized by the above.

5. A propagation characteristic estimation device for estimating propagation characteristics between a transmitting station and a receiving station in a wireless communication system, configured to execute a process of acquiring estimation data including one or more receiving station positions, an input image generation process of generating an input image that gives a distribution of heights of the structures around the receiving station according to each of the one or more receiving station positions from map data including information on the structures on the map, and a process of estimating the propagation characteristics by inputting the input image into a machine learning model. The machine learning model includes a feature extraction layer that extracts features using the input image as input, and an estimation layer that estimates the propagation characteristics using at least the features as input. The input image generation process includes a process of specifying an estimation target range on a map that is the target of the propagation characteristic estimation based on the estimation data, a process of generating target range height data that gives the height distribution of the structures in the estimation target range from the map data, and a process of generating the input image by reading a part of the target range height data for each of the one or more receiver positions. A propagation characteristic estimation apparatus characterized by the above.

6. A propagation characteristic estimation system for estimating the propagation characteristics between a transmitter and a receiver in a wireless communication system, including one or more processors, and a data server that manages map data including information on structures on a map as a map database. The one or more processors are configured to acquire estimation data including one or more receiver positions, perform an input image generation process of generating an input image that gives the height distribution of the structures around the receiver according to each of the one or more receiver positions from the map data, and perform a process of estimating the propagation characteristics by inputting the input image into a machine learning model. The machine learning model includes a feature extraction layer that extracts features using the input image as input, and an estimation layer that estimates the propagation characteristics using at least the features as input. The input image generation process includes a process of specifying an estimation target range on a map that is the target of the propagation characteristic estimation based on the estimation data, a process of generating target range height data that gives the height distribution of the structures in the estimation target range from the map data, and a process of generating the input image by reading a part of the target range height data for each of the one or more receiver positions. For each of the one or more receiver positions, a process of generating the input image by reading a part of the target range height data, and A propagation characteristic estimation system characterized by the above.

7. A propagation characteristic estimation program for estimating the propagation characteristics between a transmitter and a receiver in a wireless communication system, A process of acquiring estimation data including one or more receiver positions, An input image generation process of generating an input image that gives the height distribution of the structures around the receiver according to each of the one or more receiver positions from map data including information on the structures on the map, A process of estimating the propagation characteristics by inputting the input image into a machine learning model, and causing a computer to execute, The machine learning model includes a feature extraction layer that extracts features using the input image as an input, and an estimation layer that estimates the propagation characteristics using at least the features as an input, The input image generation process includes A process of specifying an estimation target range on the map that is the target of the estimation of the propagation characteristics based on the estimation data, A process of generating target range height data that gives the height distribution of the structures in the estimation target range from the map data, For each of the one or more receiver positions, a process of generating the input image by reading a part of the target range height data, and A propagation characteristic estimation program characterized by the above.

Images