Trained model, estimation device, estimation system, and learning method

Abstract

A trained model is provided that can improve the accuracy of estimating electromagnetic waves. [Solution] The trained model is a trained model that estimates electromagnetic waves generated by an electromagnetic wave source in space, and is trained using as training data first distribution data determined by the position of the electromagnetic wave source in space, second distribution data related to electrical properties in space, simulation data that tracks electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is first calculated electromagnetic wave data at a predetermined time after the emission of the electromagnetic waves, and simulation data that tracks electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is second calculated electromagnetic wave data at a time after the predetermined time, and estimates electromagnetic wave data after the time of emission from the electromagnetic wave source or after an arbitrary time based on the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time.

Description

[Technical Field]

[0001] The present disclosure relates to a trained model, an estimation device, an estimation system, and a learning method. [Background technology]

[0002] The ray tracing method is one of the well-known methods for estimating radio wave strength. The ray tracing method can estimate radio wave strength with high accuracy based on geometric optics and geometric optics analysis theory. Non-Patent Document 1 listed below describes a method for speeding up calculations using the ray tracing method by using machine learning. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Ryotaro Okuda, Tomoaki Nonoume, Naoki Kato, Seiichi Harada, Takuto Sakuma, Shohei Kato, Yutaka Hirayama, "Basic Study of Radio Wave Intensity Estimation Method for Factories Using 3D CNN", IEICE Technical Report, A P2023-163, January 2024, pp.19-23 Summary of the Invention [Problem to be solved by the invention]

[0004] In the estimation method described in Non-Patent Document 1, the final calculation results obtained by the ray tracing method are given to the learning model. Therefore, there is a need to propose a more accurate estimation method that reflects the physical characteristics of radio waves in space.

[0005] The present disclosure describes a trained model, an estimation device, an estimation system, and a training method that can improve the estimation accuracy of electromagnetic waves. [Means for solving the problem]

[0006] A trained model according to one aspect of the present disclosure is a trained model that estimates electromagnetic waves generated by an electromagnetic wave source in space, and is trained using, as training data, first distribution data determined by the position of the electromagnetic wave source in space, second distribution data related to electrical properties in space, simulation data that tracks electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is first calculated electromagnetic wave data at a predetermined time after the emission of the electromagnetic waves, and simulation data that tracks electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is second calculated electromagnetic wave data at a time after the predetermined time, and estimates electromagnetic wave data after the time of emission from the electromagnetic wave source or after an arbitrary time based on the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time.

[0007] This trained model is trained using as training data first distribution data regarding the spatial position of the electromagnetic wave source, second distribution data regarding the electrical properties in the space, first calculated electromagnetic wave data tracking the electromagnetic waves at a predetermined time after the electromagnetic waves are emitted from the electromagnetic wave source for a predetermined period of time, and second calculated electromagnetic wave data tracking the electromagnetic waves at times after the predetermined time. As a result, the trained model can accurately estimate electromagnetic wave data after the emission time from the electromagnetic wave source or after the arbitrary time using the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time as input. As a result, it is possible to obtain estimation results that also reflect the physical properties of electromagnetic waves in space, thereby improving the estimation accuracy of electromagnetic waves in space.

[0008] The electromagnetic wave may be radio wave intensity or electric field intensity. In this case, the electromagnetic wave intensity in space or the electric field distribution in space can be estimated with high accuracy.

[0009] The first calculated electromagnetic wave data may include calculated electromagnetic wave data at multiple times after the predetermined time. In this case, the estimation accuracy of the electromagnetic wave data can be improved by training using the time-series calculated electromagnetic wave data.

[0010] The second calculated electromagnetic wave data may include calculated electromagnetic wave data at a plurality of times. In this case, by training using the time-series calculated electromagnetic wave data, it becomes possible to estimate the time-series electromagnetic wave data at once with high accuracy.

[0011] An estimation device according to another aspect of the present disclosure includes the trained model recited in claim 1, and calculates final output electromagnetic wave data indicating an electromagnetic wave distribution associated with continuous electromagnetic wave radiation from an electromagnetic wave source by integrating time-series electromagnetic wave data estimated by the trained model. In this estimation device, by integrating the time-series electromagnetic wave data, it becomes possible to accurately obtain the electromagnetic wave distribution associated with continuous electromagnetic wave radiation.

[0012] An estimation system according to another aspect of the present disclosure includes the trained model recited in claim 1, and calculates final output electromagnetic wave data indicating an electromagnetic wave distribution accompanying continuous electromagnetic wave radiation from an electromagnetic wave source by integrating time-series electromagnetic wave data estimated by the trained model. In this estimation device, in this estimation system, by integrating the time-series electromagnetic wave data, it becomes possible to accurately obtain the electromagnetic wave distribution accompanying continuous electromagnetic wave radiation.

[0013] A learning method according to another aspect of the present disclosure is a learning method for learning a model that estimates electromagnetic waves generated by an electromagnetic wave source in space, using, as training data, first distribution data determined by the position of the electromagnetic wave source in space, second distribution data related to electrical properties in the space, simulation data that tracks electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is first calculated electromagnetic wave data at a predetermined time after the emission of the electromagnetic waves, and simulation data that tracks the electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, which is second calculated electromagnetic wave data at a time after the predetermined time, and learning the model to estimate electromagnetic wave data after the emission time from the electromagnetic wave source or after the arbitrary time, based on the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time.

[0014] According to this learning method, a model can be trained using as training data first distribution data regarding the position of an electromagnetic wave source in space, second distribution data regarding electrical properties in space, first calculated electromagnetic wave data obtained by tracking electromagnetic waves at a predetermined time after the electromagnetic wave is emitted from the electromagnetic wave source for a predetermined period of time, and second calculated electromagnetic wave data obtained by tracking electromagnetic waves at times after the predetermined time. Thus, the trained model can accurately estimate electromagnetic wave data after the time of emission from the electromagnetic wave source or after the arbitrary time using as input the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time. As a result, estimation results that also reflect the physical properties of electromagnetic waves in space can be obtained, thereby realizing a model that can improve the estimation accuracy of electromagnetic waves in space. [Effects of the Invention]

[0015] According to the present disclosure, it is possible to improve the accuracy of estimating the electromagnetic wave distribution. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a schematic diagram of an estimation device according to an embodiment. [Figure 2] FIG. 2 is a diagram showing the layout of a space for which the estimation device of FIG. 1 estimates radio wave intensity. [Figure 3] FIG. 3 is a diagram showing the distribution of radio waves represented by the calculated electromagnetic wave data along the XY plane. [Figure 4] FIG. 4 is a diagram illustrating the configuration of a trained model constructed by the training unit of FIG. [Figure 5] FIG. 5 is a diagram illustrating the estimation process of a trained model according to an embodiment. [Figure 6] FIG. 6 is a flowchart showing the procedure of the learning process of the trained model performed by the estimation device of FIG. DETAILED DESCRIPTION OF THE INVENTION

[0017] A trained model, an estimation device, an estimation system, and a training method according to an embodiment will be described in detail below with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are designated by the same reference numerals, and redundant description will be omitted.

[0018] A schematic configuration of an estimation device according to one embodiment will be described with reference to FIG. 1. FIG. 1 is a schematic configuration diagram of an estimation device according to one embodiment. The estimation device 1 shown in FIG. 1 is a data processing device such as a personal computer, a server device, or a terminal device, which includes a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access Memory). For example, various functions of the estimation device 1 shown in FIG. 1 are realized by loading a program stored in the ROM onto the RAM and executing it on the CPU. Furthermore, the estimation device 1 may be an estimation system made up of a plurality of data processing devices configured to be able to transmit and receive data to and from each other.

[0019] The estimation device 1 is a system that calculates estimation data for estimating characteristic values ​​of electromagnetic waves generated by an electromagnetic wave source such as a transmitting antenna in a space, and includes, as functional components, a data acquisition unit 101, a learning unit 102, a model storage unit 103, an estimation unit 104, and a calculation unit 105. The characteristic value of the electromagnetic waves estimated by the estimation device 1 may be radio wave intensity, electric field intensity, or, if the electromagnetic wave source is a light source, light intensity. An example in which the characteristic value to be estimated is radio wave intensity will be described below.

[0020] FIG. 2 is a diagram showing a layout L of a space for which the estimation device 1 estimates radio wave intensity. The layout L has a size of, for example, 3 m in height, 12.8 m in length, and 12.8 m in width, and is made of concrete (with a relative dielectric constant ε r The room has a ceiling and floor formed by dielectric constant σ = 6, conductivity σ = 0.194 S / m, and no walls. In this layout L, there is an obstacle (relative permittivity ε r The estimation target was a layout L of the estimation target, in which a half-wave dipole antenna (transmission power 17 dBm) with a radiating frequency of 100 kHz and a conductivity of σ=27778 S / m) was placed at a height of 2.5 mm from the center of the floor. However, the layout L of the estimation target can be changed as appropriate in terms of size, materials of the floor or ceiling, number, positions, and materials of obstacles, and position, number, and configuration of the transmitting antenna.

[0021] Next, the function of each component of the estimation device 1 will be described.

[0022] The data acquiring unit 101 acquires input data to be input to a trained model that estimates the propagation of electromagnetic waves generated by a transmitting antenna in layout L. That is, four types of data are acquired as input data: relative permittivity distribution data (second distribution data), conductivity distribution data (second distribution data), transmitter-receiver distance data (first distribution data), and free space propagation data. Furthermore, the data acquiring unit 101 acquires calculated electromagnetic wave data when training the trained model.

[0023] The relative permittivity distribution data is three-dimensional voxel data representing the relative permittivity distribution in a three-dimensional XYZ space, with the corner of the floor in layout L as the origin O, the directions along the two sides of the floor as the X-axis and Y-axis, and the direction perpendicular to the floor as the Z-axis. For example, the relative permittivity distribution data is three-dimensional voxel data divided at 0.2 m intervals in the Z-axis direction and at 0.1 m intervals in the X-axis and Y-axis directions. Similarly, the conductivity distribution data is three-dimensional voxel data representing the conductivity distribution in the three-dimensional XYZ space. The transmitter-receiver distance data is three-dimensional voxel data representing the distance distribution between the transmitting antenna and the receiving antenna in the three-dimensional XYZ space (assuming that the receiving antenna is located at each voxel in the three-dimensional space). The free space propagation data is three-dimensional voxel data representing the radio wave intensity distribution when there are no obstacles in layout L. This transmitter-receiver distance data is data determined by the position of the transmitting antenna in space of layout L, the relative permittivity distribution data and the conductivity distribution data are data determined by the electrical characteristics in space of layout L, and the free space propagation data is data calculated by the Friis propagation formula. The relative permittivity distribution data, conductivity distribution data, transmitter-receiver distance data, and free space propagation data may be calculated and generated within the estimation device 1 based on the configuration of layout L, the positions and configuration of obstacles, and the positions and configuration of the transmitting antenna, or may be calculated and generated by an external device and acquired by the estimation device 1.

[0024] The calculated electromagnetic wave data is simulation data used in ray-tracing calculations based on the configuration of layout L, the positions and configurations of obstacles, and the positions and configurations of transmitting antennas. The ray-tracing method tracks pulsed radio waves emitted from a transmitting point (corresponding to a transmitting antenna) as they are geometrically reflected, transmitted, and diffracted by structures along the way, and adds up the tracked pulsed radio waves at each receiving point (corresponding to each point in layout L) to calculate the radio wave intensity at each receiving point when radio waves are continuously emitted from the transmitting point. The calculated electromagnetic wave data is three-dimensional voxel data that shows the distribution of radio waves calculated by tracking them at each receiving point in the layout after a predetermined arrival time based on radio waves emitted from a transmitting antenna within the layout for a predetermined time (e.g., 3.3 ns). In other words, the calculated electromagnetic wave data is data used for addition in ray-tracing calculations and is radio wave intensity data of pulsed radio waves at a predetermined time. The calculated electromagnetic wave data may be calculated and generated within the estimation device 1 based on the configuration of the layout L, the position and configuration of obstacles, and the position and configuration of the transmitting antenna, or may be calculated and generated by an external device and acquired by the estimation device 1.

[0025] 3 shows the distribution of radio waves represented by calculated electromagnetic wave data along the XY plane at the height of the transmitting antenna. The data acquisition unit 101 acquires a plurality of calculated electromagnetic wave data corresponding to a plurality of arrival time widths in a time series after emission of pulsed radio waves having a time width of 3.3 ns from the transmitting antenna. For example, assuming that the time when radio waves are emitted from the transmitting antenna is 0 s, the time series of calculated electromagnetic wave data is acquired, including calculated electromagnetic wave data representing the propagation state of radio waves for an arrival time width "Time 0" ranging from 0 ns immediately thereafter to 3.3 ns after the start of radio wave emission, calculated electromagnetic wave data for an arrival time width "Time 1" ranging from 3.3 ns after the start of radio wave emission to 6.6 ns after the start of radio wave emission, calculated electromagnetic wave data for an arrival time width "Time 2" ranging from 6.6 ns after the start of radio wave emission to 10.0 ns after the start of radio wave emission, calculated electromagnetic wave data for an arrival time width "Time 3" ranging from 10.0 ns after the start of radio wave emission to 13.3 ns after the start of radio wave emission, calculated electromagnetic wave data for an arrival time width "Time 4" ranging from 13.3 ns after the start of radio wave emission to 16.6 ns after the start of radio wave emission, and calculated electromagnetic wave data for an arrival time width "Time N" at any time after the start of radio wave emission. These time series of calculated electromagnetic wave data reflect the physical characteristics of radio waves obtained in the calculation process using the ray tracing method. Specifically, the shape of the shadow of the radio wave cast by an obstacle, which changes depending on the distance between the transmitting antenna and the obstacle, is reflected in the calculated electromagnetic wave data as a physical characteristic of the radio wave.

[0026] The learning unit 102 uses the input data and the time-series calculated electromagnetic wave data acquired by the data acquisition unit 101 as training data to construct, by machine learning, a trained model, which is an algorithm (program) for estimating the propagation of radio waves generated by a transmitting antenna in the layout L. As the model that is the target of machine learning by the learning unit 102, for example, a U-Net configured by a CNN (convolutional neural network) for semantic segmentation is used.

[0027] FIG. 4 is a diagram showing the configuration of a trained model constructed by the training unit 102. The trained model includes an encoder unit ENC and a decoder unit DEC. The encoder unit ENC convolves five channels of input data, which are 16 x 128 x 128 three-dimensional voxel data, multiple times to extract features of the input data. The decoder unit DEC receives the features extracted by the encoder unit ENC, performs deconvolution, and outputs estimated electromagnetic wave data representing the distribution of radio waves, which is three-dimensional voxel data of the same size as the input data.

[0028] In detail, the encoder unit ENC is configured to perform a convolution operation with a filter size of 3x3x3 using the convolution layer L1 and maximum value pooling within a range of 2x2x2 using the pooling layer L2 four times. The decoder unit DEC is configured to perform upsampling using the upsampling layer L3 and a convolution operation similar to that of the encoder unit ENC using the convolution layer L1 four times, and then perform a convolution operation with a filter number of 1 and a filter size of 1x1x1 using the convolution layer L4 to output estimated electromagnetic wave data. Because the encoder unit ENC has performed dimensionality reduction to extract features, the upsampling layer L3 of the decoder unit DEC operates to combine the data before dimensionality reduction with the upsampled data so as not to lose information.

[0029] When constructing the trained model, the learning unit 102 inputs five channels of three-dimensional voxel data, which combines four types of input data and calculated electromagnetic wave data (first calculated electromagnetic wave data) corresponding to a predetermined arrival time width (predetermined time), into the trained model configured as described above. The trained model is trained using a known learning method by adjusting parameters such as coefficients of the convolution operation in the trained model so that the estimated electromagnetic wave data, which is one channel of three-dimensional voxel data output from the trained model, approaches the calculated electromagnetic wave data (second calculated electromagnetic wave data) corresponding to an arrival time width after the predetermined arrival time width. At this time, the learning unit 102 may train the trained model using input data and calculated electromagnetic wave data acquired when the configuration of the layout L, the position and configuration of obstacles, and the position and configuration of the transmitting antenna are variously changed. The learning unit 102 stores the trained model, whose parameters have been adjusted by training, in the model storage unit 103.

[0030] The estimation unit 104 acquires estimated electromagnetic wave data after the radio wave emission start time based on four types of input data related to the layout L, which is the target of electromagnetic wave propagation estimation, and transmitted electromagnetic wave data related to radio waves emitted from the transmitting antenna on the layout L. The transmitted electromagnetic wave data is calculated electromagnetic wave data corresponding to the arrival time width immediately after the radio wave emission start time ("time 0"), which is calculated based on the type of transmitting antenna, the radio wave frequency, the transmission power, etc. In other words, the estimation unit 104 inputs the four types of input data and the transmitted electromagnetic wave data into the trained model read from the model storage unit 103, and causes the trained model to estimate estimated electromagnetic wave data for "time 1" immediately after the radio wave emission start time.

[0031] Furthermore, the estimation unit 104 sequentially inputs the four types of input data and the estimated electromagnetic wave data corresponding to the predetermined arrival time width output from the trained model to the trained model, and sequentially estimates estimated electromagnetic wave data corresponding to the arrival time width after the predetermined arrival time width. Here, the estimation unit 104 preferably uses the trained model to estimate estimated electromagnetic wave data corresponding to the arrival time width "time t+1" immediately after the predetermined arrival time width from estimated electromagnetic wave data corresponding to the predetermined arrival time width "time t" (t is a natural number) (FIG. 5). This allows sequential acquisition of time-series estimated electromagnetic wave data (e.g., N pieces of estimated electromagnetic wave data corresponding to consecutive arrival time widths from "time 1" to "time N").

[0032] The calculation unit 105 calculates final output electromagnetic wave data indicating the distribution of radio wave intensity accompanying continuous radiation of radio waves from the transmitting antenna by integrating voxel values ​​for the time-series estimated electromagnetic wave data acquired for the layout L and the transmitting antenna on the layout L. This final output electromagnetic wave data is three-dimensional voxel data with the same number of dimensions as the estimated electromagnetic wave data. The calculation unit 105 then outputs the calculated final output electromagnetic wave data. The output destination of the final output electromagnetic wave data may be an output device such as a display connected to the estimation device 1, or may be a device internal to the estimation device 1 such as RAM. The calculation unit 105 may also transmit the final output electromagnetic wave data to an external device such as a terminal device via a network.

[0033] Hereinafter, with further reference to FIG. 6, the learning process of the trained model performed by the estimation device 1, that is, the procedure of the learning method of this embodiment, will be described in detail.

[0034] 6, the estimation device 1 first acquires input data and time-series calculated electromagnetic wave data based on the configuration of a predetermined layout L, the positions and configuration of predetermined obstacles, and the positions and configuration of predetermined transmitting antennas (step S1). Next, the estimation device 1 uses the input data and the time-series calculated electromagnetic wave data as training data to train a learning model so that estimated electromagnetic wave data estimated from calculated electromagnetic wave data corresponding to a predetermined arrival time width (first calculated electromagnetic wave data) approaches calculated electromagnetic wave data corresponding to an arrival time width after the predetermined arrival time width (second calculated electromagnetic wave data) (step S2).

[0035] Then, the estimation device 1 determines whether to repeat learning of the trained model using calculated electromagnetic wave data corresponding to another arrival time width as input (step S3). If the result of the determination is to repeat learning (step S3; Yes), the estimation device 1 repeats the learning process of step S2 using calculated electromagnetic wave data corresponding to another arrival time width as input.

[0036] On the other hand, if the learning is not to be repeated (step S3; No), the estimation device 1 determines whether or not to repeat the learning based on a different configuration of the layout L, a different position and configuration of obstacles, or a different position and configuration of the transmitting antenna (step S4). If it is determined as a result of the determination that the learning is to be repeated (step S4; Yes), steps S1 to S3 are repeated to execute the learning process using new input data and new calculation electromagnetic wave data.

[0037] On the other hand, if it is determined not to repeat the learning (step S4; No), the estimation device 1 stores the learned model constructed by the learning process in the model storage unit 103 (step S5). This completes the learning process by the estimation device 1. The judgments in steps S3 and S4 may be, for example, whether all of the prepared training data has been learned, whether specific training data with low reliability has been learned, or whether the process has been repeated a predetermined number of times.

[0038] The trained model constructed by the estimation device 1 described above is trained using, as training data, first distribution data regarding the position of the transmitting antenna in the space of layout L, second distribution data regarding the electrical properties in the space of layout L, first calculated electromagnetic wave data obtained by tracking radio waves for a predetermined arrival time width after the radio waves are emitted from the transmitting antenna for a predetermined time, and second calculated electromagnetic wave data obtained by tracking radio waves for an arrival time width after the arrival time width. As a result, the trained model receives as input the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the transmitting antenna or the calculated electromagnetic wave data for an arbitrary arrival time width, and is capable of accurately estimating electromagnetic wave data after the time of transmission from the transmitting antenna or after the arbitrary arrival time width. As a result, the shape of the shadow of the radio waves cast by the obstacle, which changes depending on the distance between the transmitting antenna and the obstacle, is reflected in the calculated electromagnetic wave data as a physical characteristic of the radio waves. This allows for estimation results that also reflect the physical characteristics of the radio waves in the space of layout L, thereby improving the estimation accuracy of the radio waves in the space of layout L.

[0039] Furthermore, according to the estimation device 1 equipped with a trained model, by accumulating the time-series electromagnetic wave data estimated by the trained model, it is possible to accurately obtain the distribution of radio wave intensity associated with the continuous emission of radio waves.

[0040] Although one embodiment of the present disclosure has been described in detail above, the trained model, estimation device, and estimation system according to the present disclosure are not limited to the above embodiment.

[0041] In the above embodiment, the trained model is trained to estimate, from calculated electromagnetic wave data (first calculated electromagnetic wave data) corresponding to one arrival time width, calculated electromagnetic wave data (second calculated electromagnetic wave data) corresponding to an arrival time width later than the one arrival time width. In contrast, as a first modified example, the trained model may be trained to estimate, from a plurality of calculated electromagnetic wave data corresponding to each of a plurality of arrival time widths, calculated electromagnetic wave data (second calculated electromagnetic wave data) corresponding to an arrival time width later than the plurality of arrival time widths. Furthermore, as a second modified example, the trained model may be trained to estimate, from a plurality of calculated electromagnetic wave data (first calculated electromagnetic wave data) corresponding to one arrival time width, a plurality of electromagnetic wave data (second calculated electromagnetic wave data) corresponding to each of a plurality of arrival time widths later than the one arrival time width. Furthermore, as a third modified example, the trained model may be trained to estimate, from a plurality of calculated electromagnetic wave data corresponding to each of a plurality of arrival time widths, calculated electromagnetic wave data (second calculated electromagnetic wave data) corresponding to each of a plurality of arrival time widths later than the plurality of arrival time widths.

[0042] For example, when the time-series calculated electromagnetic wave data shown in FIG. 3 is used as training data, in the first modified example, the trained model may be trained so that calculated electromagnetic wave data at "time 2" can be estimated using calculated electromagnetic wave data at "time 0" and calculated electromagnetic wave data at "time 1" as input. In the second modified example, the trained model may be trained so that calculated electromagnetic wave data at "time 1" can be estimated using calculated electromagnetic wave data at "time 2" and calculated electromagnetic wave data at "time 3". In the third modified example, the trained model may be trained so that calculated electromagnetic wave data at "time 0" and calculated electromagnetic wave data at "time 1" as input can be estimated using calculated electromagnetic wave data at "time 2" and calculated electromagnetic wave data at "time 3". [Explanation of symbols]

[0043] 1...estimation device, 101...data acquisition unit, 102...learning unit, 103...model storage unit, 104...estimation unit, 105...calculation unit, DEC...decoder unit, ENC...encoder unit, L...layout, L1, L4...convolutional layer, L2...pooling layer, L3...upsampling layer.

Claims

1. A trained model for estimating electromagnetic waves generated by an electromagnetic wave source in space, learning is performed using, as training data, first distribution data determined by the position of the electromagnetic wave source in the space, second distribution data relating to electrical properties in the space, first calculated electromagnetic wave data at a predetermined time after the emission of the electromagnetic waves, which is simulation data for tracking electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, and second calculated electromagnetic wave data at a time after the predetermined time, which is simulation data for tracking electromagnetic waves emitted from the electromagnetic wave source for the predetermined time, estimating electromagnetic wave data after the time of emission from the electromagnetic wave source or after the arbitrary time based on the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at the arbitrary time; Trained model.

2. The trained model according to claim 1 , wherein the electromagnetic waves are radio wave intensity or electric field intensity.

3. the first calculated electromagnetic wave data includes calculated electromagnetic wave data at a plurality of times after the predetermined time; The trained model according to claim 1.

4. The second calculated electromagnetic wave data includes calculated electromagnetic wave data at a plurality of times. The trained model according to claim 1 or 3.

5. The trained model according to claim 1 is included, calculating final output electromagnetic wave data indicating an electromagnetic wave distribution accompanying continuous radiation of electromagnetic waves from the electromagnetic wave source by integrating the time-series electromagnetic wave data estimated by the trained model; Estimation device.

6. The trained model according to claim 1 is included, calculating final output electromagnetic wave data indicating an electromagnetic wave distribution accompanying continuous radiation of electromagnetic waves from the electromagnetic wave source by integrating the time-series electromagnetic wave data estimated by the trained model; Estimation system.

7. A learning method for learning a model that estimates electromagnetic waves generated by an electromagnetic wave source in space, comprising: using, as training data, first distribution data determined by the position of the electromagnetic wave source in the space, second distribution data relating to electrical properties in the space, first calculated electromagnetic wave data for a predetermined time after the emission of the electromagnetic waves, which is simulation data for tracking the electromagnetic waves emitted from the electromagnetic wave source for a predetermined time, and second calculated electromagnetic wave data for a time after the predetermined time, which is simulation data for tracking the electromagnetic waves emitted from the electromagnetic wave source for the predetermined time, learning a model based on the first distribution data, the second distribution data, and the electromagnetic wave data emitted from the electromagnetic wave source or the calculated electromagnetic wave data at an arbitrary time, so as to estimate the electromagnetic wave data after the emission time from the electromagnetic wave source or after the arbitrary time; How to learn.

Images