Analysis device, analysis method, analysis system, machine learning model creation method and program

Abstract

To improve prediction accuracy while reducing the cost of creating prediction data for physical phenomena. [Solution] The analysis device 1 includes an acquisition unit 131 that acquires analysis target spatial data at a first resolution and physical phenomenon data represented at a second resolution, and an analysis unit 132 that outputs high-resolution physical phenomenon data at a first resolution, which is output from a machine learning model that has been deeply trained using a plurality of training data based on learning spatial data that shows the positional relationships of one or more obstacles in a plurality of learning spaces including a space different from the analysis target space and learning physical phenomenon data corresponding to the learning spatial data.

Description

Technical Field

[0001] The present invention relates to an analysis device, an analysis method, an analysis system, a machine learning model creation method, and a program.

Background Art

[0002] Conventionally, a machine learning model has been used to analyze fluid phenomena, which are physical phenomena of gases or liquids. For example, a technique for predicting the microclimate of a city using a machine learning model is known (see, for example, Patent Document 1 and Non-Patent Document 1).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Non-Patent Documents

[0004]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] In conventional techniques, generating high-resolution predictive data for physical phenomena using low-resolution data representing those phenomena required creating a machine learning model by using high-resolution simulated data as training data. However, high-resolution simulations are time-consuming, resulting in high computational costs.

[0006] Furthermore, if the resolution required for prediction data does not match the resolution of the high-resolution data used in machine learning, the output data becomes overly sensitive to the input data of the machine learning model, leading to a decrease in prediction accuracy.

[0007] Therefore, the present invention has been made in view of these points, and aims to improve prediction accuracy while reducing the cost of creating prediction data for physical phenomena. [Means for solving the problem]

[0008] An analysis device according to a first aspect of the present invention includes: an acquisition unit that acquires analysis target space data of a first resolution showing the positional relationships of one or more obstacles included in an analysis target space where physical phenomena with different characteristic values ​​depending on position and time occur, and physical phenomenon data showing physical phenomena in the analysis target space, represented at a second resolution which is lower than the first resolution; a storage unit that stores parameters of a machine learning model that has been deeply trained using a plurality of training data based on a third resolution which is lower than the first resolution, learning space data showing the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the analysis target space, and a fourth resolution which is lower than the first resolution, learning physical phenomenon data corresponding to the learning space data, wherein the machine learning model outputs high-resolution physical phenomenon data of the first resolution when the analysis target space data and the physical phenomenon data are input; and an analysis unit that outputs high-resolution physical phenomenon data of the first resolution output from the machine learning model by inputting the analysis target space data and the physical phenomenon data to the machine learning model.

[0009] The analysis unit may output the high-resolution physical phenomenon data in association with the analysis target spatial data.

[0010] At least one of the learning spatial data sets corresponding to the plurality of training data sets may be data showing the positional relationship of one or more obstacles in a space different from the space under analysis, which includes a second obstacle whose similarity to the shape of at least one first obstacle included in the space under analysis is equal to or greater than a threshold. The threshold is, for example, a value set in advance as a value that can be considered to indicate that the tendencies of the obstacle shapes on physical phenomena are equivalent, or a value learned by the machine learning model through deep learning as a value that can be considered to indicate that the tendencies of the obstacle shapes on physical phenomena are equivalent.

[0011] The physical phenomenon data includes characteristic values ​​associated with coordinates in the analysis target space, the learning physical phenomenon data includes characteristic values ​​associated with coordinates in the learning space, and the machine learning model may generate obstacle-specific physical phenomenon data indicating the content of the physical phenomenon that occurs for each of the one or more obstacles indicated by the learning space data, based on the positions of one or more obstacles indicated by the learning space data and the characteristic values ​​associated with coordinates indicated by the learning physical phenomenon data, and perform deep learning using the training data based on the generated obstacle-specific physical phenomenon data.

[0012] The machine learning model may be a model that has undergone deep learning using training data based on a plurality of obstacle-specific physical phenomenon data, which are generated using attention corresponding to each of a plurality of regions containing the obstacle, based on the characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data, and which show a pattern of a plurality of characteristic values ​​corresponding to a plurality of coordinates in each of the plurality of regions where the influence of the obstacle is relatively high.

[0013] The machine learning model may generate the physical phenomenon data for each obstacle by interpolating the characteristic values ​​during the process of calculating the attention.

[0014] The machine learning model, when Q is a query indicating a pattern of characteristic values ​​based on physical phenomenon data corresponding to the region for which attention is to be calculated, V is a value including the physical phenomenon data, K is a key including the spatial data to be analyzed corresponding to the region, and N is the length of the spatial data to be analyzed corresponding to the region, The attention may be calculated using TIFF2026058009000002.tif2546.

[0015] The aforementioned physical phenomenon data and the aforementioned high-resolution physical phenomenon data may be data showing the distribution of the characteristic values ​​in the space under analysis.

[0016] The spatial data to be analyzed and the spatial data for learning may include data indicating the shape and properties of the one or more obstacles.

[0017] A second aspect of the present invention provides an analysis method comprising the steps of: acquiring analysis target space data of first resolution, which is performed by a computer and shows the positional relationships of one or more obstacles included in an analysis target space where physical phenomena with different characteristic values ​​depending on the position and time occur; and physical phenomenon data, which is represented at a second resolution lower than the first resolution and shows the physical phenomena in the analysis target space; and a machine learning model that has been deeply trained using a plurality of training data based on a third resolution lower than the first resolution, which is learning space data showing the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the analysis target space; and a fourth resolution lower than the first resolution, which is learning physical phenomenon data corresponding to the learning space data, wherein when the analysis target space data and the physical phenomenon data are input to the machine learning model, the machine learning model outputs high-resolution physical phenomenon data of first resolution, and the machine learning model outputs high-resolution physical phenomenon data of first resolution.

[0018] A third aspect of the present invention is a program that causes a computer to perform the following steps: acquire analysis target space data of first resolution showing the positional relationships of one or more obstacles included in an analysis target space in which physical phenomena with different characteristic values ​​depending on location and time occur, and physical phenomenon data showing physical phenomena in the analysis target space, represented at a second resolution which is lower than the first resolution; and deeply train a machine learning model that has been trained using a plurality of training data based on a third resolution which is lower than the first resolution, the learning space data showing the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the analysis target space, and a fourth resolution which is lower than the first resolution, the learning physical phenomenon data corresponding to the learning space data, wherein when the analysis target space data and the physical phenomenon data are input to the machine learning model, the program outputs the first resolution high-resolution physical phenomenon data output from the machine learning model by inputting the analysis target space data and the physical phenomenon data to the machine learning model.

[0019] A fourth aspect of the present invention provides an analysis system comprising an information terminal and an analysis device, wherein the information terminal transmits to the analysis device first resolution analysis space data showing the positional relationships of one or more obstacles included in the analysis space where physical phenomena with different characteristic values ​​occur depending on location and time, and physical phenomenon data showing the physical phenomena in the analysis space, represented at a second resolution which is lower than the first resolution, and the analysis device comprises an acquisition unit that acquires the analysis space data and the physical phenomenon data, and learning space data at a third resolution which is lower than the first resolution, showing the positional relationships of one or more obstacles in a plurality of learning spaces which include spaces different from the analysis space. The device comprises: a storage unit that stores parameters for a machine learning model that has been deeply trained using a plurality of training data based on a fourth resolution of learning physical phenomenon data having a lower resolution than the first resolution, and which outputs high-resolution physical phenomenon data of the first resolution when the spatial data to be analyzed and the physical phenomenon data are input to the machine learning model; and an analysis unit that outputs the high-resolution physical phenomenon data of the first resolution, output from the machine learning model by inputting the spatial data to be analyzed and the physical phenomenon data to the machine learning model, to the information terminal, and the information terminal displays the high-resolution physical phenomenon data.

[0020] A fifth corresponding machine learning model creation method of the present invention is a machine learning model creation method for outputting high-resolution physical phenomenon data of the first resolution, based on: first-resolution data of a target space for analysis showing the positional relationships of one or more obstacles included in a target space for analysis where physical phenomena with different characteristic values ​​depending on location and time occur; and physical phenomenon data, which includes the characteristic values ​​associated with coordinates in the target space for analysis, expressed at a second resolution lower than the first resolution, wherein the method comprises: third-resolution learning space data, which has a lower resolution than the first resolution, showing the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the target space for analysis; and fourth-resolution learning space data, which has a lower resolution than the first resolution. The method includes the steps of: acquiring physical phenomenon data, which includes characteristic values ​​associated with coordinates in the learning space corresponding to the learning space data; generating obstacle-specific physical phenomenon data by interpolating the characteristic values ​​in each of the multiple regions containing the obstacle, based on the pattern of characteristic values ​​shown by each of the attentions, which are attentions corresponding to each of the multiple regions where the influence of the obstacle is relatively high in each of the multiple regions; and creating the machine learning model by performing deep learning using the multiple obstacle-specific physical phenomenon data as training data.

[0021] The program according to the sixth aspect of the present invention causes a computer to create a machine learning model for outputting high-resolution physical phenomenon data of the first resolution based on analysis target space data of the first resolution indicating the positional relationship of one or more obstacles included in an analysis target space where physical phenomena with different characteristic values occur depending on the position and time, and physical phenomenon data including the characteristic values associated with the coordinates in the analysis target space represented by a second resolution lower than the first resolution. The computer is caused to acquire learning space data of a third resolution lower than the first resolution, which indicates the positional relationship of one or more obstacles in a plurality of learning spaces including a space different from the analysis target space, and learning physical phenomenon data of a fourth resolution lower than the first resolution, which includes the characteristic values associated with the coordinates in the learning space corresponding to the learning space data. The computer is caused to generate obstacle-specific physical phenomenon data by interpolating the characteristic values in the region based on the patterns of the characteristic values indicated by the attention corresponding to each of the plurality of regions including the obstacle, the attention indicating the patterns of the characteristic values corresponding to a plurality of coordinates with relatively high influence of the obstacle in each of the plurality of regions. The computer is caused to create the machine learning model by performing deep learning using the plurality of obstacle-specific physical phenomenon data as training data.

Effect of the Invention

[0022] According to the present invention, there is an effect that the prediction accuracy can be improved while suppressing the cost of creating prediction data of physical phenomena.

Brief Description of the Drawings

[0023] [Figure 1] It is a diagram for explaining the outline of the analysis system S. [Figure 2]This figure shows an example of an image in which high-resolution physical phenomenon data is displayed overlaid on an image of the space being analyzed. [Figure 3] This diagram shows examples of the data processing flow when creating a machine learning model M and when using the machine learning model M. [Figure 4] This flowchart shows the process for updating the parameters of a machine learning model M. [Figure 5] This is a diagram showing the configuration of the analysis device 1. [Figure 6] This figure shows an example illustrating the effectiveness of the analysis method according to this embodiment. [Modes for carrying out the invention]

[0024] [Overview of Analysis System S] Figure 1 is a diagram illustrating the overview of the analysis system S. The analysis system S is a system for analyzing physical phenomena whose characteristic values ​​differ depending on location and time. Examples of physical phenomena whose characteristic values ​​differ depending on location and time include heat conduction, fluid movement, and electromagnetic wave radiation, and the characteristic values ​​are values ​​such as temperature, humidity, wind speed, and electromagnetic wave intensity.

[0025] The analysis system S comprises an analysis device 1 and an information terminal 2. The analysis device 1 is a computer for analyzing physical phenomena. The analysis device 1 may be a personal computer or a cloud server. The analysis device 1 receives input from the information terminal 2 in the form of analysis target space data indicating the positional relationships of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​depending on location and time occur, and physical phenomenon data with a lower resolution than the analysis target space data. The analysis device 1 outputs physical phenomenon data with a higher resolution than the input physical phenomenon data by using a machine learning model M created based on the training spatial data.

[0026] Physical phenomenon data is data that includes characteristic values ​​associated with coordinates in the space under analysis. Physical phenomenon data and high-resolution physical phenomenon data are, for example, data that show the distribution of characteristic values ​​in the space under analysis. As an example, physical phenomenon data and high-resolution physical phenomenon data are heatmap data in which the color or density differs depending on the characteristic value of each of the multiple locations in the space under analysis.

[0027] The spatial data to be analyzed and the spatial data for training include data indicating the shape and properties of one or more obstacles (e.g., buildings) present in the space. The properties of the obstacles are indicated by the type of material that makes up the obstacle, physical properties such as the coefficient of friction of the material surface, or whether or not the obstacles generate heat.

[0028] Resolution corresponds to the minimum distance over which the value of spatial and physical phenomena data changes. In other words, resolution corresponds to the distance between adjacent pixels in spatial and physical phenomena data. A higher resolution means a smaller distance, and a lower resolution means a larger distance.

[0029] Traditionally, creating such machine learning models required deep learning using a large amount of training data that associated spatial data showing the location and shape of buildings with low-resolution and high-resolution training data on physical phenomena. Since creating high-resolution training data on physical phenomena is not easy, creating machine learning models traditionally required a great deal of time and effort.

[0030] The machine learning model M according to this embodiment is created without preparing high-resolution training data of physical phenomena. The machine learning model M operates by a processor executing a program. The processor may be the processor of the analysis device 1, or it may be the processor of another computer. In the following description, it will be assumed that the machine learning model M operates by a processor of the analysis device 1 executing a program.

[0031] The machine learning model M generates obstacle-specific physical phenomenon data that describes the physical phenomena that occur for each of the one or more obstacles indicated in the training spatial data, based on the positions of one or more obstacles indicated in the training spatial data and the characteristic values ​​associated with the coordinates indicated in the training physical phenomenon data. This model is then deeply learned using training data based on the generated obstacle-specific physical phenomenon data.

[0032] The machine learning model M is a model that has undergone deep learning using training data based on one or more obstacle-specific physical phenomenon data generated using attention corresponding to each of several regions containing obstacles, based on characteristic values ​​associated with coordinates shown in the training physical phenomenon data. Attention is data that represents a pattern of physical phenomenon data calculated using patterns of multiple characteristic values ​​corresponding to multiple coordinates where the influence of obstacles is relatively high in each of several regions. The machine learning model M generates obstacle-specific geographic phenomenon data by interpolating characteristic values ​​in the process of calculating attention. Specifically, the machine learning model M calculates attention by interpolating characteristic values ​​in the pattern of characteristic values ​​by applying transformer technology, and generates obstacle-specific physical phenomenon data by calculating that attention.

[0033] The machine learning model M, when Q is a query that shows a pattern of characteristic values ​​based on physical phenomenon data corresponding to the region for which attention is to be calculated, V is a value that includes the physical phenomenon data, K is a key that includes the analysis target spatial data corresponding to the region containing the obstacle, and N is the length of the analysis target spatial data corresponding to that region, Attention A is calculated using TIFF2026058009000003.tif2546. Query Q and Value V may contain data other than physical phenomenon data. Key K may contain data other than the analysis target space data.

[0034] Here, key K corresponds to an instruction such as "create a flow around a building," (K TV) corresponds to the amplitude of the pattern of characteristic values ​​around obstacles in space. The machine learning model M is relatively large (K T Select the pattern Q of characteristic values ​​corresponding to V). The degree of influence (K) T The selection of characteristic value pattern Q based on V) is a process that interpolates characteristic values ​​between multiple positions corresponding to the selected multiple characteristic values.

[0035] Furthermore, query Q may include analysis target spatial data and physical phenomenon data corresponding to the region for which attention is to be calculated. Also, value V and key K may include analysis target spatial data and physical phenomenon data.

[0036] Here, among the multiple training spatial data corresponding to multiple training data, at least one training spatial data is data showing the positional relationship of one or more obstacles in a space different from the space under analysis. The space different from the space under analysis is, for example, a space containing a second obstacle whose similarity to the shape of at least one first obstacle included in the space under analysis is greater than or equal to a threshold. The threshold is a value that is pre-set as a value that can be considered to indicate that the tendencies of the obstacle shapes on physical phenomena are equivalent, or a value learned by the machine learning model M through deep learning.

[0037] "A second obstacle whose similarity to the shape of the first obstacle is above a threshold" means that the first and second obstacles have shapes that have an equivalent effect on the physical phenomenon. An example of a second obstacle whose similarity to the shape of the first obstacle is above a threshold is an obstacle that includes a part with the same shape as the first obstacle, or an obstacle that includes a part with a similar shape to the first obstacle.

[0038] The parameters of the machine learning model M are updated by a computer running a program, for example, using the following procedure. This computer may be analysis device 1 or another computer. Here, we illustrate the case where analysis device 1 creates the machine learning model M.

[0039] First, the analysis device 1 calculates attention to each of the multiple regions represented by the low-resolution training spatial data, based on the low-resolution training physical phenomenon data, thereby identifying a characteristic value pattern composed of multiple characteristic values ​​that have a relatively high correlation with each of the multiple regions. The analysis device 1 generates characteristic values ​​that interpolate the multiple characteristic values ​​included in the identified characteristic value pattern through attention calculation, thereby creating training physical phenomenon data with a higher resolution than the training physical phenomenon data. The analysis device 1 can efficiently create a machine learning model M by performing deep learning using the created high-resolution training physical phenomenon data.

[0040] Information terminal 2 is a terminal used by user U, who uses the analysis results of physical phenomena, and is, for example, a personal computer, tablet, or smartphone. Analysis device 1 and information terminal 2 can send and receive various types of data via a network.

[0041] When information terminal 2 receives an operation from user U to specify the area to be analyzed for which physical phenomenon data will be created, it transmits the spatial data to be analyzed corresponding to that area to analysis device 1. The spatial data to be analyzed is three-dimensional data showing the location and shape of buildings in the space for which physical phenomenon data will be created. As described above, the spatial data to be analyzed may further indicate the properties of the buildings. Information terminal 2 may transmit data indicating the area to be analyzed to analysis device 1 in lieu of, or together with, the spatial data to be analyzed.

[0042] Furthermore, the information terminal 2 transmits low-resolution physical phenomenon data, which shows the characteristic values ​​of the physical phenomena in the space under analysis, to the analysis device 1. For example, the information terminal 2 transmits low-resolution physical phenomenon data obtained from an external server that provides low-resolution physical phenomenon data to the analysis device 1. If the analysis device 1 can obtain low-resolution physical phenomenon data corresponding to a specific area, the information terminal 2 may not transmit the low-resolution physical phenomenon data to the analysis device 1, but instead transmit data indicating the area under analysis to the analysis device 1.

[0043] Analysis device 1 acquires data of the target space and low-resolution physical phenomenon data, and inputs this data into machine learning model M. Analysis device 1 acquires high-resolution physical phenomenon data corresponding to the target space, which is created by machine learning model M based on the data of the target space and low-resolution physical phenomenon data, from machine learning model M. Analysis device 1 transmits the acquired high-resolution physical phenomenon data to information terminal 2.

[0044] Information terminal 2 displays high-resolution physical phenomenon data received from analysis device 1. For example, information terminal 2 displays the high-resolution physical phenomenon data overlaid on an image of the analysis target space showing buildings included in the analysis target space specified by user U.

[0045] Figure 2 shows an example of an image in which high-resolution physical phenomenon data is displayed overlaid on an image of the space under analysis. The high-resolution physical phenomenon data shown in Figure 2 indicates the temperature distribution in the space under analysis. In Figure 2, the temperature of each of the multiple locations within the space under analysis is represented by the intensity of the color, but the temperature may also be represented by differences in color.

[0046] Figure 3 shows examples of data processing flows when creating a machine learning model M and when using a machine learning model M. Figure 3(a) shows an example of data processing flows when creating a machine learning model M. Figure 3(b) shows an example of data processing flows when using a machine learning model M.

[0047] In the example shown in Figure 3(a), spatial data with a resolution of 20m and physical phenomenon data with a resolution of 100m corresponding to the space represented by the spatial data are input to the machine learning model M. 20m resolution is a resolution in which the presence or absence of a building is represented by a quadrilateral pixel corresponding to a 20m square area. The machine learning model M is created by performing deep learning using training data based on high-resolution physical phenomenon data generated by interpolating characteristic values ​​within the input 100m resolution physical phenomenon data.

[0048] The training data is based on obstacle-specific physical phenomenon data, which describes the physical phenomena that occur for each of the one or more obstacles in the space represented by the learning spatial data. When the machine learning model M performs unsupervised learning, the obstacle-specific physical phenomenon data output by the machine learning model M may be used as training data. When the machine learning model M performs supervised learning, the training data is a pair of low-resolution physical phenomenon data and obstacle-specific physical phenomenon data (ground truth data).

[0049] In the example shown in Figure 3(b), analysis target spatial data with a resolution of 5m and physical phenomenon data with a resolution of 100m corresponding to the analysis target space indicated by the analysis target spatial data are input to the analysis device 1. The analysis device 1 inputs the analysis target spatial data with a resolution of 5m and the physical phenomenon data with a resolution of 100m into the machine learning model M and acquires the physical phenomenon data with a resolution of 5m.

[0050] In Figure 3(b), an example is shown where the resolution of the spatial data to be analyzed is 5m, but the resolution of the spatial data to be analyzed is arbitrary. As described above, the machine learning model M performs deep learning using high-resolution physical phenomenon data, which is obtained by interpolating characteristic values ​​through the calculation of attention calculated for each region containing a building. Therefore, it can output physical phenomenon data with the same resolution as the input spatial data to be analyzed. As a result, the analysis device 1 outputs physical phenomenon data with the same resolution as the input spatial data to be analyzed.

[0051] [How to update the parameters of a machine learning model M] Figure 4 is a flowchart showing the process for updating the parameters of the machine learning model M. The process shown in Figure 4 may be performed by any training computer, for example, but the following explanation will illustrate the case where the analysis device 1 performs the process shown in Figure 4.

[0052] The analysis device 1 creates a machine learning model M for outputting high-resolution physical phenomenon data at first resolution, based on first-resolution analysis space data showing the positional relationships of one or more obstacles contained in the analysis space where physical phenomena with different characteristic values ​​depending on position and time occur, and physical phenomenon data including characteristic values ​​associated with coordinates in the analysis space, represented at a second resolution which is lower than the first resolution.

[0053] To create a machine learning model M, the analysis device 1 first acquires training spatial data at a third resolution, which is lower than the first resolution (S11). The training spatial data at the third resolution is training spatial data that shows the positional relationships of one or more obstacles in multiple training spaces, including spaces different from the space to be analyzed. The analysis device 1 also acquires training physical phenomenon data at a fourth resolution, which is lower than the first resolution (S12). The training physical phenomenon data at the fourth resolution is training physical phenomenon data that includes characteristic values ​​associated with coordinates in the training space corresponding to the training spatial data. The third and fourth resolutions may be the same or different.

[0054] Next, the analysis device 1 selects multiple regions containing obstacles (S13). The shape of the multiple regions is arbitrary and may be a square, a rectangle, or a circle.

[0055] Next, the analysis device 1 calculates attention corresponding to each of the multiple regions containing the obstacle, based on characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data (S14). Attention is calculated based on a pattern of multiple characteristic values ​​corresponding to multiple coordinates in each of the multiple regions where the influence of the obstacle is relatively high.

[0056] Next, the analysis device 1 interpolates characteristic values ​​in the region by calculating attention and generates physical phenomenon data for each obstacle (S15). The analysis device 1 updates the parameters of the machine learning model M by performing deep learning using training data based on the one or more physical phenomenon data for each obstacle that it has generated (S16).

[0057] If the analysis device 1 determines that there are other areas to which obstacle-specific physical phenomenon data can be generated (YES in S17), it repeats the processes from S13 to S16. If the analysis device 1 determines that there are no other areas to which obstacle-specific physical phenomenon data can be generated (NO in S17), it determines whether there is other training data (learning spatial data and learning physical phenomenon data) targeting different seasons, times of day, or locations (S18).

[0058] If the analysis device 1 determines that there is other training data (YES in S18), it repeats the processes from S11 to S17. If the analysis device 1 determines that there is no other training data (NO in S18), it terminates the process of updating the parameters of the machine learning model M.

[0059] Furthermore, the first resolution does not need to be constant across all regions of the spatial data under analysis; it may be distributed within a resolution range higher than the second resolution. Similarly, the second resolution does not need to be constant across all regions of the corresponding spatial data of physical phenomena; it may be distributed within a resolution range lower than the first resolution. The third resolution does not need to be constant across all regions of the training spatial data; it may be distributed within a resolution range lower than the first resolution. The fourth resolution does not need to be constant across all regions of the corresponding spatial data of physical phenomena for training; it may be distributed within a resolution range lower than the first resolution.

[0060] [Configuration of Analysis Device 1] The configuration and operation of the analysis device 1 will be described in detail below. Figure 5 shows an example of the configuration of the analysis device 1. The analysis device 1 has a communication unit 11, a storage unit 12, and a control unit 13. The control unit 13 has an acquisition unit 131 and an analysis unit 132.

[0061] The communication unit 11 has a communication interface for sending and receiving data to and from the information terminal 2. The communication unit 11 receives, for example, data of the space to be analyzed and low-resolution physical phenomenon data from the information terminal 2. The communication unit 11 also transmits high-resolution physical phenomenon data generated by the analysis unit 132 to the information terminal 2.

[0062] The memory unit 12 includes storage media such as ROM (Read Only Memory), RAM (Random Access Memory), and SSD (Solid State Drive). The memory unit 12 stores the program executed by the control unit 13. The memory unit 12 also stores the machine learning model parameters 121, which are parameters of the machine learning model M.

[0063] The machine learning model M, upon receiving the data of the space to be analyzed and the data of physical phenomena, outputs high-resolution physical phenomenon data at a first resolution. The machine learning model M is a machine learning model that has undergone deep learning using training spatial data at a third resolution lower than the first resolution, and training physical phenomenon data at a fourth resolution lower than the first resolution. The training spatial data is data that shows the positional relationships of one or more obstacles in multiple training spaces, including spaces different from the space to be analyzed. The training physical phenomenon data is physical phenomenon data corresponding to the training spatial data, and is data that shows characteristic values ​​of physical phenomena for each position (coordinate) in the space indicated by the training spatial data.

[0064] The control unit 13 is, for example, a CPU (Central Processing Unit). The control unit 13 functions as an acquisition unit 131 and an analysis unit 132 by executing a program stored in the storage unit 12.

[0065] The acquisition unit 131 acquires data of the analysis target space at a first resolution, which shows the positional relationship of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​occur depending on the position and time, and physical phenomenon data that shows the physical phenomena in the analysis target space, which is represented at a second resolution, which is lower than the first resolution. The acquisition unit 131 notifies the analysis unit 132 of the acquired data. The acquisition unit 131 may also store the acquired data in the storage unit 12.

[0066] The analysis unit 132 generates and outputs first-resolution physical phenomenon data by analyzing the second-resolution physical phenomenon data input from the acquisition unit 131. Specifically, the analysis unit 132 outputs first-resolution high-resolution physical phenomenon data output from the machine learning model M by inputting the analysis target spatial data and physical phenomenon data into the machine learning model M.

[0067] The analysis unit 132 outputs high-resolution physical phenomenon data in association with the analysis target spatial data. For example, as shown in Figure 2, the analysis unit 132 outputs the high-resolution physical phenomenon data to be displayed on the information terminal 2 overlaid on the analysis target spatial data. The analysis unit 132 may also output the high-resolution physical phenomenon data to be displayed on the information terminal 2 alongside the analysis target spatial data. The analysis unit 132 may also output the high-resolution physical phenomenon data to a printing device.

[0068] [Examples] Figure 6 shows an example illustrating the effectiveness of the analysis method according to this embodiment. Figure 6(a) is physical phenomenon data showing measured values ​​of the temperature distribution in a certain analysis target space. Figure 6(b) is the data output by a conventional machine learning model described in Non-Patent Literature 1 when analysis target space data and physical phenomenon data with different resolutions than those used during training are input to the model. In this case, the data output by the machine learning model contains a lot of noise, and the physical phenomenon data output by the machine learning model differs significantly from the physical phenomenon data showing measured values ​​as shown in Figure 6(a).

[0069] Figure 6(c) shows the high-resolution physical phenomenon data output by the machine learning model M according to this embodiment when the analysis target spatial data and physical phenomenon data with different resolutions than those used during training are input to the machine learning model M. The high-resolution physical phenomenon data shown in Figure 6(c) has almost no noise like that seen in Figure 6(b), and can be seen to be close to the measured values ​​shown in Figure 6(a).

[0070] Thus, when using the analysis method according to this embodiment, regardless of the resolution of the spatial data and physical phenomenon data input to the machine learning model M, physical phenomenon data that approximates the measured physical phenomenon data is output. Therefore, if the physical phenomenon data output by the machine learning model does not contain noise like that seen in Figure 6(b) when spatial data and physical phenomenon data of various resolutions are input to the machine learning model, it is highly probable that the analysis method according to this embodiment has been implemented.

[0071] Although the present invention has been described above using embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes are possible within the scope of its gist. For example, all or part of the apparatus can be configured by functionally or physically distributing and integrating in any unit. Furthermore, new embodiments resulting from any combination of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiments resulting from the combinations are combined with the effects of the original embodiments. [Explanation of symbols]

[0072] 1 Analysis device 2. Information terminals 11 Communications Department 12 Storage section 13 Control Unit 131 Acquisition Department 132 Analysis Department

Claims

1. An acquisition unit acquires: first resolution data of the analysis target space showing the positional relationship of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​depending on the position and time occur; and physical phenomenon data representing the physical phenomena in the analysis target space, expressed at a second resolution which is lower than the first resolution. A machine learning model that has been deeply trained using a plurality of training data based on: training spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of training spaces including a space different from the space to be analyzed; and training physical phenomenon data of a fourth resolution lower than the first resolution, which corresponds to the training physical phenomenon data; wherein the machine learning model stores parameters of the machine learning model that outputs high-resolution physical phenomenon data of the first resolution when the space to be analyzed and the physical phenomenon data are input. An analysis unit that outputs high-resolution physical phenomenon data of the first resolution, which is output from the machine learning model by inputting the spatial data to be analyzed and the physical phenomenon data into the machine learning model, An analytical device having the following features.

2. The analysis unit outputs the high-resolution physical phenomenon data in association with the analysis target spatial data. The analysis apparatus according to claim 1.

3. At least one of the learning spatial data sets corresponding to the plurality of training data sets is data showing the positional relationship of one or more obstacles in a space different from the space to be analyzed, which includes a second obstacle whose similarity to the shape of at least one first obstacle included in the space to be analyzed is equal to or greater than a threshold. The analysis apparatus according to claim 1.

4. The threshold is a value that is set in advance as a value that can be considered to represent an equivalent tendency for the shape of an obstacle to affect a physical phenomenon, or a value that the machine learning model has learned through deep learning as a value that can be considered to represent an equivalent tendency for the shape of an obstacle to affect a physical phenomenon. The analysis apparatus according to claim 3.

5. The physical phenomenon data includes the characteristic values ​​associated with coordinates in the space under analysis, The learning physical phenomenon data includes the characteristic values ​​associated with coordinates in the learning space, The machine learning model generates obstacle-specific physical phenomenon data indicating the content of the physical phenomenon that occurs for each of the one or more obstacles indicated by the learning spatial data, based on the positions of one or more obstacles indicated by the learning spatial data and the characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data, and is a model that has undergone deep learning using the training data based on the generated obstacle-specific physical phenomenon data. The analysis apparatus according to claim 1.

6. The machine learning model is a model that has undergone deep learning using training data based on a plurality of obstacle-specific physical phenomenon data, which are generated using attention corresponding to each of a plurality of regions containing the obstacle, based on the characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data, and which show patterns of a plurality of characteristic values ​​corresponding to a plurality of coordinates in each of the plurality of regions where the influence of the obstacle is relatively high. The analysis apparatus according to claim 5.

7. The machine learning model generates the physical phenomenon data for each obstacle by interpolating the characteristic values ​​in the process of calculating the attention. The analysis apparatus according to claim 6.

8. The machine learning model, when Q is a query indicating a pattern of characteristic values ​​based on physical phenomenon data corresponding to the region for which attention is to be calculated, V is a value including the physical phenomenon data, K is a key including the spatial data to be analyzed corresponding to the region, and N is the length of the spatial data to be analyzed corresponding to the region, The attention is calculated by the following: The analysis apparatus according to claim 6.

9. The aforementioned physical phenomenon data and the aforementioned high-resolution physical phenomenon data are data showing the distribution of the characteristic values ​​in the space under analysis. The analysis apparatus according to claim 1.

10. The spatial data to be analyzed and the spatial data to be learned include data indicating the shape and properties of the one or more obstacles. The analysis apparatus according to claim 1.

11. A computer executes The steps include obtaining: first-resolution data of the analysis target space showing the positional relationship of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​occur depending on location and time; and physical phenomenon data showing the physical phenomena in the analysis target space, represented at a second resolution which is lower than the first resolution; A machine learning model that has been deeply trained using a plurality of training data based on: training spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of training spaces including a space different from the space to be analyzed; and training physical phenomenon data of a fourth resolution lower than the first resolution, which corresponds to the training physical phenomenon data; wherein when the space to be analyzed and the physical phenomenon data are input, the machine learning model outputs high-resolution physical phenomenon data of the first resolution, and the machine learning model outputs high-resolution physical phenomenon data of the first resolution by inputting the space to be analyzed and the physical phenomenon data into the machine learning model, and the machine learning model outputs high-resolution physical phenomenon data of the first resolution. Tail analysis method.

12. On the computer, The steps include obtaining: first-resolution data of the analysis target space showing the positional relationship of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​occur depending on location and time; and physical phenomenon data showing the physical phenomena in the analysis target space, represented at a second resolution which is lower than the first resolution; A machine learning model that has been deeply trained using a plurality of training data based on: training spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of training spaces including a space different from the space to be analyzed; and training physical phenomenon data of a fourth resolution lower than the first resolution, which corresponds to the training physical phenomenon data; wherein when the space to be analyzed and the physical phenomenon data are input, the machine learning model outputs high-resolution physical phenomenon data of the first resolution, and the machine learning model outputs high-resolution physical phenomenon data of the first resolution by inputting the space to be analyzed and the physical phenomenon data into the machine learning model, and the machine learning model outputs high-resolution physical phenomenon data of the first resolution. A program to execute.

13. Equipped with an information terminal and an analysis device, The aforementioned information terminal is The analysis device is provided with first-resolution data of the analysis target space, which shows the positional relationship of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​depending on location and time occur, and physical phenomenon data representing the physical phenomena in the analysis target space, which is expressed at a second resolution lower than the first resolution. The aforementioned analysis device is An acquisition unit that acquires the aforementioned spatial data to be analyzed and the aforementioned physical phenomenon data, A machine learning model that has been deeply trained using a plurality of training data based on: training spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of training spaces including a space different from the space to be analyzed; and training physical phenomenon data of a fourth resolution lower than the first resolution, which corresponds to the training physical phenomenon data; wherein the machine learning model stores parameters of the machine learning model that outputs high-resolution physical phenomenon data of the first resolution when the space to be analyzed and the physical phenomenon data are input. An analysis unit that outputs the high-resolution physical phenomenon data of the first resolution, output from the machine learning model by inputting the spatial data to be analyzed and the physical phenomenon data into the machine learning model, to the information terminal, It has, The aforementioned information terminal is an analysis system that displays the high-resolution physical phenomenon data.

14. A method for creating a machine learning model to output high-resolution physical phenomenon data at a first resolution, based on: first-resolution data of a target space for analysis that shows the positional relationships of one or more obstacles contained in a target space for analysis where physical phenomena with different characteristic values ​​depending on location and time occur; and physical phenomenon data that includes the characteristic values ​​associated with coordinates in the target space for analysis, represented at a second resolution which is lower than the first resolution. The steps include: acquiring learning spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the space to be analyzed; and acquiring learning physical phenomenon data of a fourth resolution lower than the first resolution, which includes characteristic values ​​associated with coordinates in the learning space corresponding to the learning spatial data; The steps include generating obstacle-specific physical phenomenon data by interpolating the characteristic values ​​in each of the multiple regions containing the obstacle, based on the characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data, and each of the attentions indicating a pattern of multiple characteristic values ​​corresponding to multiple coordinates where the influence of the obstacle is relatively high in each of the multiple regions, and The steps include creating the machine learning model by performing deep learning using multiple sets of obstacle-specific physical phenomenon data as training data, A method for creating a machine learning model that has [specific features / abilities].

15. A program for causing a computer to create a machine learning model for outputting high-resolution physical phenomenon data at a first resolution, based on first-resolution data of an analysis target space that shows the positional relationships of one or more obstacles contained in the analysis target space where physical phenomena with different characteristic values ​​depending on location and time occur, and physical phenomenon data that includes the characteristic values ​​associated with coordinates in the analysis target space, represented at a second resolution which is lower than the first resolution. To the aforementioned computer, The steps include: acquiring learning spatial data of a third resolution lower than the first resolution, which shows the positional relationships of one or more obstacles in a plurality of learning spaces including spaces different from the space to be analyzed; and acquiring learning physical phenomenon data of a fourth resolution lower than the first resolution, which includes characteristic values ​​associated with coordinates in the learning space corresponding to the learning spatial data; The steps include generating obstacle-specific physical phenomenon data by interpolating the characteristic values ​​in each of the multiple regions containing the obstacle, based on the characteristic values ​​associated with the coordinates indicated by the learning physical phenomenon data, and each of the attentions indicating a pattern of multiple characteristic values ​​corresponding to multiple coordinates where the influence of the obstacle is relatively high in each of the multiple regions, and The steps include creating the machine learning model by performing deep learning using multiple sets of obstacle-specific physical phenomenon data as training data, A program to execute.

Images