Prediction model generation method, prediction model generation system, teacher data for generating prediction model, and learned prediction model

Abstract

To efficiently perform machine learning while suppressing increase in teacher data for machine learning corresponding to a shape of a space and an object therein different for each prediction target.SOLUTION: A prediction model generation method is a method in which one or more computers 10 execute processing of performing machine learning of a parameter using teacher data and generating a prediction model 20 having data in a first state acquired at least by data acquisition units 11, 12 as an input and data in a second state as an output. Teacher data includes data indicating at least one main flow 40 generated in a fluid in a space 2 by influence of external force and data indicating at least one local flow 41 generated when the main flow 40 hits at least one object 3 present in the space 2.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] FIELD OF THE DISCLOSURE Embodiments of the present invention relate to predictive model generation techniques. [Background technology]

[0002] In recent years, the application of machine learning has progressed in various academic fields. Machine learning is also being applied in many areas of the fluid field, such as shape optimization. Indoor ventilation, which has attracted attention for its potential to reduce the risk of infectious diseases, is also important for ensuring safety at workplaces involving harmful gases or dust from the perspective of spatial flow. However, numerical simulations using general-purpose fluid analysis software take a long time to make predictions and cannot be used in real time, so the application of machine learning is expected. There is also a known technology for evaluating ventilation using numerical simulations.

[0003] Conventional fluid simulations, for example, require the generation of computational grids and the numerical convergence of equations relating to momentum and mass, resulting in a large amount of calculation. When a machine learning prediction model is used, it is expected that predictions can be obtained in a short calculation time, but in order to enable such predictions, it is necessary to have the prediction model learn similar flow conditions as training data in advance. However, preparing the training data requires a large number of numerical simulations. Therefore, there is a demand for efficient machine learning while suppressing the increase in training data. In addition, there is a demand for the realization of a system that automatically generates training data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2018-169818 A [Non-patent literature]

[0005] [Non-Patent Document 1] “No.65 “Airflow Simulation””, [Retrieved July 10, 2023], Internet<URL:https: / / www.jfe-tec.co.jp / jfetec-news / 65 / 2p.html> Summary of the Invention [Problem to be solved by the invention]

[0006] The problem that the present invention aims to solve is to perform machine learning efficiently while suppressing an increase in training data for machine learning corresponding to the different shapes of spaces and the objects within them for each prediction target. [Means for solving the problem]

[0007] A predictive model generation method according to an embodiment of the present invention is a method in which one or more computers execute a process to input at least one of first state data or data simulating the first state indicating at least the shape of a space and the behavior of a fluid in the space, acquire teacher data which outputs at least one of second state data or data simulating the second state indicating the behavior of the fluid in the future relative to the first state, perform machine learning of parameters using the teacher data, and generate a predictive model using at least the first state data acquired by a data acquisition unit as input and the second state data as output, wherein the teacher data includes data indicating at least one main flow generated in the fluid in the space due to the influence of an external force, and data indicating at least one local flow generated when the main flow hits at least one object present in the space. Effect of the Invention

[0008] According to an embodiment of the present invention, machine learning can be performed efficiently while suppressing an increase in training data for machine learning corresponding to the different spatial shapes and objects within each prediction target. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a configuration diagram showing a predictive model generation system according to a first embodiment. [Diagram 2] FIG. 1 is a block diagram illustrating a predictive model generation system. [Diagram 3] FIG. 1 is a block diagram showing a prediction model. [Figure 4] FIG. 11 is a screen diagram showing a display mode in which the behavior of a fluid is visualized. [Diagram 5] FIG. 11 is a screen diagram showing a display mode in which the physical quantities of a fluid are visualized using isosurfaces. [Figure 6] FIG. 1 is an explanatory diagram showing an example of a mainstream in a space. [Figure 7] FIG. 13 is an explanatory diagram showing another example of mainstreams in space. [Figure 8] FIG. 13 is an explanatory diagram showing yet another example of mainstreams in space. [Figure 9] FIG. 13 is an explanatory diagram showing an example of a local flow. [Figure 10] FIG. 11 is an explanatory diagram showing another example of a local flow. [Figure 11] FIG. 13 is an explanatory diagram showing yet another example of a local flow. [Figure 12] 4 is a flowchart showing a teacher data generation process according to the first embodiment. [Figure 13] 10 is a flowchart showing a teacher data generation process according to the second embodiment. [Figure 14] 13 is a flowchart showing a teacher data generation process according to the third embodiment. [Figure 15] 13 is a flowchart showing a teacher data generation process according to the fourth embodiment. [Figure 16] 13 is a flowchart showing a teacher data generation process according to the fifth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] (First embodiment) Hereinafter, embodiments of a prediction model generation method, a prediction model generation system, training data for generating a prediction model, and a trained prediction model will be described in detail with reference to the drawings. First, a first embodiment will be described with reference to Figs. 1 to 12.

[0011] Reference numeral 1 in Fig. 1 denotes a prediction model generation system of the first embodiment. This prediction model generation system 1 has a function of generating a prediction model 20 (Fig. 3) and a function of predicting the behavior of a fluid based on the generated prediction model 20. In other words, the prediction model generation system 1 also serves as a flow behavior prediction system.

[0012] First, the function of the prediction model generation system 1 to predict the behavior of a fluid will be described.

[0013] The predictive model generation system 1 predicts and visualizes the behavior of a fluid inside a specified space 2 by machine learning using a neural network or the like. An example of the space 2 is the inside (room) of a room provided in a specified building. This space 2 may be a room provided in a vehicle, an aircraft, a ship, or the like. Examples of the room include a factory, a work site, a hospital examination room, a waiting room, an underground shopping mall, a train passenger compartment, and the like. Furthermore, the room may be a containment vessel of a nuclear power plant, or the like.

[0014] A fluid such as air exists inside the space 2. In addition, a certain object 3 (obstacle) that affects the behavior of the fluid exists inside the space 2. Examples of the object 3 include certain fixtures (furniture) such as chairs, desks, tables, and shelves. The object 3 may be a person in the room. In other words, the object 3 includes at least one of fixtures and a person. The shape of the room is composed of, for example, a floor surface 4, a wall surface 5, and a ceiling 6. Furthermore, the shape of the room includes the shape of at least one boundary portion 7 (open portion, intake and exhaust portion) through which air enters and exits, such as a window, a door, an intake port, and an exhaust port. This boundary portion 7 is a boundary where the fluid in the space 2 is affected by an external force. In addition, a part of the boundary portion 7 is provided with an exhaust fan 8 as a certain device for the purpose of intake and exhaust such as ventilation. This exhaust fan 8 is provided, for example, on the ceiling 6. The shape of the room and the shape of the object 3 existing in the room constitute the shape of the space 2. The behavior of the fluid changes according to the shape of the space 2.

[0015] Possible forms of these changes include, for example, the opening and closing of windows, the direction of the wind outside, starting or stopping of the exhaust fan 8, the strength or weakness of the airflow of the exhaust fan 8, changes in the position of fixtures or people, or an increase or decrease in the number of such fixtures or people.

[0016] The prediction model generation system 1 includes a computer 10 , a plurality of fluid data acquisition units 11 , and a spatial data acquisition unit 12 .

[0017] For example, the computer 10 is provided at a predetermined position near the room. The computer 10 may be provided inside the room or at a remote location connected via a network. A plurality of fluid data acquisition units 11 are provided near the floor 4, wall 5, ceiling 6, object 3, etc. of the room. The spatial data acquisition unit 12 is provided at a position overlooking the entire room, such as the upper part of the wall 5 of the room. A plurality of spatial data acquisition units 12 may be provided.

[0018] The fluid data acquiring unit 11 acquires data on the physical quantity of the fluid at an arbitrary position in the space 2. In this way, it is possible to predict the future behavior of the fluid from the physical quantity of the fluid at an actual position in the space 2.

[0019] The fluid data acquiring unit 11 is composed of a predetermined fluid sensor or the like. For example, the fluid data acquiring unit 11 measures at least one physical quantity of the speed, direction, temperature, pressure, components, etc. of the fluid present in the space 2. Each fluid data acquiring unit 11 is capable of communicating with the computer 10, and data indicating the measured physical quantity is transmitted to the computer 10. For example, time-series data indicating the temporal change in the physical quantity of the fluid at each point in the space 2 is acquired by each fluid data acquiring unit 11 and transmitted to the computer 10.

[0020] These time series data include information on the time and calendar (date, day of the week, etc.) when the physical quantities of the fluid were acquired, as well as information indicating the position of the fluid data acquisition unit 11 that acquired each physical quantity. In the following description, the data acquired by the fluid data acquisition unit 11 will be referred to as fluid data (FIG. 3).

[0021] The behavior of the fluid is at least one of the velocity and direction of the fluid. The fluid data includes the physical quantities (e.g., scalar, vector) of the velocity and direction of the fluid, and further includes other physical quantities (e.g., temperature, pressure, content of components). In this way, it is possible to predict the future physical quantities of the fluid in consideration of not only the velocity and direction of the fluid, but also other physical quantities of the fluid.

[0022] The spatial data acquisition unit 12 acquires data relating to the shape of the space 2. This data relating to the shape of the space 2 includes data on the shape of the object 3 that exists inside the space 2 and affects the behavior of the fluid. In this way, it is possible to predict the future behavior of the fluid by taking into consideration the influence of the object 3 existing inside the space 2 on the behavior of the fluid. This spatial data acquisition unit 12 senses the positions of the entrance and exit of the fluid in the space 2 (for example, the boundary portion 7), the object 3 existing in the space 2, etc., and acquires data indicating their current states. The spatial data acquisition unit 12 is capable of communicating with the computer 10, and the data acquired by the spatial data acquisition unit 12 is transmitted to the computer 10.

[0023] The space data acquisition unit 12 is composed of a predetermined camera or a shape measurement sensor. For example, the camera takes an image. This image contains information indicating the shape of the space 2 including the object 3. By taking an image with this camera, data on the shape of the space 2 including the surrounding object 3 is obtained from the computer 10.

[0024] The shape measurement sensor measures the three-dimensional shape (stereoscopic shape) of the space 2 including the object 3. For example, a depth sensor is used as the shape measurement sensor. Note that an infrared sensor or a laser sensor such as LiDAR may also be used as the shape measurement sensor.

[0025] For example, the shape measuring sensor can measure the distance from the computer 10 to the surrounding object 3 by projecting a laser onto the object 3 and receiving the reflected light with a light receiving element. The shape measuring sensor also uses a ToF (Time of Flight) method to convert the delay time between the projected pulse and the received pulse into distance, and measures the distance from the computer 10 to the surrounding object 3. By using this shape measuring sensor, three-dimensional point cloud data indicating the shape of the space 2 including the surrounding object 3 can be generated from the computer 10.

[0026] The shape of the space 2 includes information on the positions (coordinates) of the floor 4, wall 5, and ceiling 6 that constitute the space 2. The shape of the object 3 includes information on the position (coordinates) where the object 3 is placed in the space 2. The shape data of the object 3 also includes information on the time when the object 3 is present and calendar information (date, day of the week, etc.). In the following explanation, the data acquired by the spatial data acquisition unit 12 includes not only shape information, but also information indicating position and time. These are referred to as spatial data (FIG. 3).

[0027] Next, the configuration of the prediction model generation system 1 will be described with reference to the block diagram shown in Fig. 2. Note that the prediction model generation system 1 may include components other than those shown in Fig. 2, or some of the components shown in Fig. 2 may be omitted.

[0028] The computer 10 of the prediction model generation system 1 includes a display unit 13, a storage unit 14, an input unit 15, a communication unit 16, and a processing circuit 17.

[0029] The processing circuit 17 is a circuit equipped with, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a dedicated or general-purpose processor. This processor realizes various functions by executing various programs stored in the storage unit 14. The processing circuit 17 may also be configured with hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit). Various functions can also be realized by such hardware. The processing circuit 17 can also realize various functions by combining software processing by a processor and a program with hardware processing.

[0030] The processing circuit 17 has, for example, a function for generating a prediction model 20 (FIG. 3), a function for performing calculations, and a function for generating visualization information. These functions are realized by the CPU executing a program stored in a storage unit 14 such as a memory.

[0031] In other words, the predictive model generation system 1 is configured with a computer 10 having hardware resources such as a CPU, a GPU, a ROM (Read Only Memory), a RAM (Random Access Memory), a HDD (Hard Disk Drive), and an SSD (Solid State Drive), and the CPU executes various programs to realize information processing by software using the hardware resources. Furthermore, the predictive model generation method is realized by having the computer 10 execute the various programs.

[0032] Each component of the predictive model generation system 1 does not necessarily have to be provided in one computer 10. For example, one predictive model generation system 1 may be realized by a plurality of computers 10 connected to each other via a network. For example, a function for generating a predictive model 20 (FIG. 3), a function for performing calculations, and a function for generating visualization information may each be mounted on an individual computer 10. Also, the display unit 13 may be mounted on another computer 10 connected via a network. Note that the predictive model generation method and the fluid behavior prediction method may each be executed on an individual computer 10.

[0033] The display unit 13 displays predetermined information. The display unit 13 is, for example, a display that outputs a prediction result of the behavior of a fluid. The display unit 13 may be separate from the computer 10 or may be integrated with the computer 10.

[0034] Although a display is exemplified as the display unit 13 that displays an image, other modes may be used. For example, an image may be displayed using a head-mounted display or a projector. Furthermore, a printer that prints information on a paper medium may be used instead of a display. In other words, the display unit 13 may include a head-mounted display, a projector, or a printer.

[0035] The storage unit 14 stores various information necessary for predicting the behavior of a fluid and various information necessary for generating a prediction model 20 (FIG. 3). For example, the storage unit 14 stores spatial data, fluid data, and teacher data. Furthermore, the storage unit 14 stores the prediction model 20.

[0036] Predetermined information is input to the input unit 15 in response to an operation by a user operating the computer 10. The input unit 15 includes input devices such as a mouse, a keyboard, and a touch panel. That is, predetermined information is input to the computer 10 in response to an operation of these input devices. The user operating the computer 10 may be one person or multiple people.

[0037] The communication unit 16 communicates with the fluid data acquisition unit 11 and the spatial data acquisition unit 12. The communication unit 16, the fluid data acquisition unit 11, and the spatial data acquisition unit 12 may be connected by wireless communication or by wired communication. The communication unit 16 may communicate with another computer 10 connected via a predetermined network.

[0038] The processing circuitry 17 includes a prediction model 20 (FIG. 3) that has undergone machine learning. The processing circuitry 17 performs calculations to predict the future state of the fluid using this prediction model 20.

[0039] The processing circuitry 17 includes artificial intelligence (AI) that performs machine learning. The processing circuitry 17 also includes a deep learning unit that extracts a specific pattern from multiple patterns based on deep learning.

[0040] For prediction using the computer 10, a prediction technique based on learning of artificial intelligence is used. For example, a learning model generated by machine learning using a neural network, a learning model generated by other machine learning, a deep learning algorithm, a mathematical algorithm such as regression analysis, etc. In addition, forms of machine learning include forms such as clustering and deep learning.

[0041] The predictive model generation system 1 includes at least one computer 10 having artificial intelligence that performs machine learning. For example, the predictive model generation system 1 may be configured with one computer 10 having a neural network, or may be configured with multiple computers 10 having neural networks.

[0042] Here, the neural network is a mathematical model that expresses the characteristics of brain functions through a simulation by a computer 10. For example, this model shows that artificial neurons (nodes) that form a network through synaptic connections change the connection strength of the synapses through learning, and acquire problem-solving ability. Furthermore, the neural network acquires problem-solving ability through deep learning.

[0043] 3, the prediction model 20 includes an input layer 21, one or more intermediate layers 22, and an output layer 23. The input layer 21 receives input of first state data indicating at least the shape of the space 2 and the behavior of the fluid inside the space 2. The output layer 23 outputs second state data indicating at least the behavior of the fluid in the future relative to the first state.

[0044] The data of the first state includes spatial data acquired by the spatial data acquisition unit 12 and fluid data of the current state acquired by the fluid data acquisition unit 11. The spatial data includes information such as the shape of the room, the shape of the fixtures, the shape of a person, etc. The fluid data includes information such as the speed, direction, temperature, pressure, and components of the fluid, etc.

[0045] The data in the second state includes fluid data in a state further in the future than the first state. In the first embodiment, the prediction is made on the assumption that the spatial data will not change in the future, so the data in the second state does not include spatial data. However, the data in the second state may include future spatial data. For example, this future spatial data is identical to the data in the first state except for the time and calendar information. Here, the time and calendar information are information that is determined to change in the future.

[0046] The second state that is future than the first state may be one second, one minute, ten minutes, one hour, or one day after the first state. The time difference between the first state and the second state can be arbitrarily set by the user in advance.

[0047] The intermediate layer 22 has parameters machine-learned using training data in which at least one of data in a first state or data simulating the first state is input, and at least one of data in a second state or data simulating the second state is output.

[0048] This prediction model 20 causes the computer 10 to function in such a way that data in a first state is input to an input layer 21, calculations are performed in an intermediate layer 22, and data in a second state is output from an output layer 23.

[0049] For example, the neural network constituting the prediction model 20 is provided with an intermediate layer 22 having multiple layers. Each layer of the intermediate layer 22 is composed of multiple units. In addition, by having the multi-layered neural network learn in advance using teacher data, it is possible to automatically extract feature quantities in the pattern of changes in the state of a circuit or system. Note that the multi-layered neural network can be set to any number of intermediate layers, any number of units, any learning rate, any number of learning times, and any activation function on the user interface.

[0050] In addition, deep reinforcement learning may be used in the neural network, in which a reward function is set for each information item to be learned, and the information item with the highest value is extracted based on the reward function.

[0051] For example, a convolutional neural network (CNN) that has a proven track record in image recognition is used. In this CNN, the intermediate layer 22 is composed of a convolutional layer and a pooling layer. The convolutional layer obtains a feature map by applying filtering to nearby nodes in the previous layer. The pooling layer further reduces the feature map output from the convolutional layer to create a new feature map. At this time, by obtaining the maximum value of the pixels included in the region of interest in the feature map, it is possible to absorb slight deviations in the position of the feature amount.

[0052] The convolutional layer extracts local features from an image, and the pooling layer performs processing to consolidate local features. In these processes, the image is reduced while maintaining the features of the input image. In other words, the convolutional neural network (CNN) can significantly compress (abstract) the amount of information contained in an image. Then, the input image can be recognized and classified using the abstracted image image stored in the neural network.

[0053] In addition, there are various machine learning techniques such as autoencoder, LSTM (Long Short-Term Memory), SDF (Signed Distance Function), GAN (Generative Adversarial Network), RNN (Recurrent Neural Network), etc. These techniques may be applied to machine learning.

[0054] For example, CNN and LSTM are used for machine learning. For example, SDF is used for inputting spatial data to the prediction model 20. The prediction model 20 has, for example, both CNN that compresses spatial data and LSTM that performs time series prediction.

[0055] As the teacher data, data actually acquired by the spatial data acquiring unit 12 and the fluid data acquiring unit 11 can be used. Alternatively, data simulating these may be used as the teacher data. Note that data at a predetermined time is set as data of the first state, and data at a time in the future is set as data of the second state.

[0056] For example, when performing machine learning of the prediction model 20, first, a plurality of patterns of basic shapes and positions of the entrance and exit of the fluid in the space 2 (e.g., the boundary portion 7) and the object 3 existing in the space 2 are set. Then, for each pattern, data including the physical quantity of the fluid is obtained by numerical simulation or statistical data. Machine learning is performed using these as teacher data.

[0057] In other words, the parameters of the prediction model 20 are machine-learned using training data in which data obtained by simulating a plurality of states of the space 2 having an arbitrary basic shape is used as input and output. In this way, a large amount of training data can be obtained in advance by simulation based on the basic shape.

[0058] Next, when the behavior of the fluid is predicted, the data of the first state and the second state actually acquired by the spatial data acquisition unit 12 and the fluid data acquisition unit 11 are input to the prediction model 20. Time series data of physical quantities indicating the future behavior of the fluid at each point in the space 2 is predicted. The prediction result is displayed on the display unit 13. For example, first, the behavior of the fluid is displayed on the display unit 13 (FIG. 1) in accordance with the actual time. Next, the behavior of the fluid in the future than the actual time is displayed on the display unit 13. In this way, it is possible to display the prediction result in real time in accordance with the actual state of the fluid.

[0059] It takes time to perform a numerical simulation of the behavior of the fluid using a basic shape and to perform machine learning using the results of the simulation. However, by completing a prediction model 20 using machine learning in advance, it is possible to instantly predict the time series data of the behavior of the fluid in any space 2 and to grasp the time series data of the physical quantity of the fluid in the future. Therefore, an image of visualized information showing the state of the fluid can be displayed on the display unit 13 in real time.

[0060] The processing circuit 17 (FIG. 2) generates visualization information that visualizes the state of the fluid, based on the data of the second state output from the prediction model 20. This visualization information is displayed on the display unit 13. In this way, the user can see the visualized state of the fluid and can understand the ventilation status of the space 2.

[0061] Next, the form of an image of visualized information displayed on the screen 30 of the display unit 13 will be described with reference to FIG. 4 and FIG.

[0062] For example, as shown in Fig. 4, an image visualizing the behavior of a fluid displays a plan view of a specified space 2 as a room. An object 3 existing in this space 2 is also displayed. The fluid moving through the space 2 is displayed as a flow of arrows. The user can grasp the behavior of the fluid from the present to a specified future state.

[0063] Furthermore, the processing circuitry 17 (FIG. 2) calculates the physical quantities of the fluid at each position in the space 2, and generates visualization information showing these physical quantities as at least one of contours, vectors, and isosurfaces. For example, the processing circuitry 17 converts the physical quantities such as the velocity, direction, temperature, pressure, components, and age of air of the fluid predicted by the prediction model 20 into isosurfaces.

[0064] As shown in FIG. 5, in an image in which the physical quantities of a fluid are visualized by an isosurface, regions are displayed that are color-coded according to the physical quantities. For example, the age of air in the fluid is displayed by color coding. Also, the temperature distribution may be displayed by color coding. Furthermore, a contour bar 31 and the like are displayed. In this way, the user can grasp the state of the fluid by the isosurface. The user can grasp the real-time or future age of air.

[0065] The physical quantities of the fluid may be displayed not only as isosurfaces, but also as contours or vectors. Furthermore, these may be displayed in combination, or may be switched between in sequence. In this way, the user can grasp the physical quantities of the fluid from various perspectives.

[0066] Furthermore, the processing circuit 17 (FIG. 2) generates visualization information that highlights and displays areas of the space 2 where the value of a specific component contained in the fluid exceeds a threshold value among the physical quantities of the fluid at each position in the space 2. For example, components such as hydrogen or carbon dioxide are preset as targets to be highlighted and displayed.

[0067] If there is an area where the value of a component such as hydrogen or carbon dioxide exceeds a threshold, this area is highlighted and displayed in a manner different from the others. For example, the highlighted area is heated up (colored), surrounded by a line, or indicated by an arrow. In this way, the user can grasp the area of ​​space 2 where the value of a specific component exceeds the threshold.

[0068] In addition, the processing circuit 17 (Figure 2) can input the second state data output from the prediction model 20 (Figure 3) to the prediction model 20 as first state data, and further output the second state data indicating the future behavior of the fluid.

[0069] Continuing to predict future fluid behavior using machine learning will result in errors. For example, predicting the future one minute from now will produce less error than predicting the future 10 minutes from now. Therefore, in the second embodiment, the error can be reduced by repeatedly making predictions at intervals of a certain period.

[0070] Next, the function of the prediction model generation system 1 for generating the prediction model 20 (FIG. 3) will be described.

[0071] As training data, the results of a simulation, which is a numerical analysis, or actually measured fluid data and spatial data are often used. However, in either case, a large amount of training data is required. Therefore, the prediction model generation system 1 generates a large amount of training data.

[0072] In addition, the flow conditions in a space 2 such as a room differ for each space 2 to be predicted, in terms of the shape of the space 2, the items inside, the arrangement of people, the arrangement of entrances and exits, and the arrangement of windows, so it is difficult to prepare a general-purpose trained prediction model 20. In addition, in order to predict airflow using machine learning, it is necessary to prepare training data that includes the conditions to be predicted and have the prediction model 20 (FIG. 3) learn them. Since the conditions differ depending on the target, it is difficult to create a universal prediction model 20 in advance, and training data must be prepared for each prediction target. Therefore, when a user sets the conditions of the space 2 to be predicted, the prediction model generation system 1 generates training data necessary for machine learning, enabling machine learning that meets the required conditions.

[0073] In machine learning, training data can be prepared by setting the shapes and positions of people and objects that may be placed in the space 2. Here, a main flow 40 (FIGS. 6 to 8) which is a global fluid flow is separated from a local flow 41 (FIGS. 9 to 11) which is a secondary fluid flow (micro flow). By having these trained separately, training data can be automatically generated while preventing an increase in the amount of training data.

[0074] The flow of air fluid can be roughly divided into a main flow 40 and a local flow 41. An example of the main flow 40 is an external wind flowing from one of two open windows to the other, as shown in Figures 6 to 7. Also, as shown in Figure 8, there is a ventilation flow flowing from an open window to an exhaust fan 8.

[0075] An example of a local flow 41 is a Karman vortex generated downstream of a specific cylindrical object 3, as shown in Fig. 9. Also, there is a vortex generated downstream of a rectangular object 3 such as a fixture, as shown in Fig. 10. Also, as shown in Fig. 11, depending on the conditions, there are cases where a vortex is not generated downstream of the object 3, and a local flow 41 is generated along the surface of the object 3.

[0076] Unlike the main flow 40, the local flow 41 is determined by the conditions around the object 3, and so by learning the local flow 41 separately from the main flow 40, it is possible to expand the conditions of various flow patterns included in the learning range while suppressing an increase in the amount of required training data.

[0077] In other words, the training data includes data indicating at least one main flow 40 that occurs in the fluid in the space 2 due to the influence of an external force, and data indicating at least one local flow 41 that occurs when the main flow 40 hits at least one object 3 present in the space 2.

[0078] Furthermore, in at least some steps of the machine learning, the computer 10 separately performs a process of learning the mainstream 40 and a process of learning the local flows 41. In this way, it is possible to learn various fluid patterns by combining patterns of the mainstream 40 and the local flows 41. Therefore, it is possible to reduce the number of patterns of the mainstream 40 and the local flows 41, and as a result, it is possible to reduce the amount of training data required for machine learning.

[0079] Before performing machine learning of the prediction model 20, the user prepares training data. For example, data showing the behavior of a fluid in a flow path having one simple structure is created by the computer 10. Then, a plurality of data of various patterns in which the form of the simple structure and the flow path are changed is created. Note that data actually acquired by the spatial data acquisition unit 12 and the fluid data acquisition unit 11 may be used as training data.

[0080] Next, the teacher data generation process executed by the prediction model generation system 1 will be described with reference to the flowchart in Fig. 12. The above-mentioned drawings may be referenced. The following steps are at least a part of the processes included in the teacher data generation process, and other steps may be included in the teacher data generation process.

[0081] In the following description, an example is given in which the computer 10 constituting the predictive model generation system 1 performs various processes (including various processes such as judgment and setting), but other modes are also possible. For example, a user may perform some of the various processes, and the computer 10 may receive an input of the processing results and use them in its own processing. In other words, the subject of the predictive model generation method may be the computer 10 or the user.

[0082] First, in step S1, the processing circuitry 17 (FIG. 2) of the computer 10 sets the shape and size of the space 2 (FIG. 1) in which the behavior of a fluid is to be predicted.

[0083] In the next step S2, the processing circuit 17 sets the boundary portion 7 (FIG. 1) which is a boundary where the fluid in the space 2 is affected by an external force. For example, the processing circuit 17 sets the boundary portion 7 (open portion, intake / exhaust portion) such as a window and an exhaust fan 8.

[0084] Here, a pattern of the arrangement set of the boundary portion 7 is set. The arrangement set of the boundary portion 7 includes, for example, the position, type, and condition range of the boundary portion 7. The condition range is a range assumed for performing machine learning, and includes the range of operating conditions such as the window opening range, the range of external wind speed, and the range of air volume of the exhaust fan 8. The operating condition includes at least one operation of opening / closing the boundary portion 7 or intake / exhaust.

[0085] It is possible to set multiple patterns, rather than just one, for the layout set of the boundary portion 7. The conditions of the boundary portion 7 of a room or the like are the states of windows, doors, shutters, movable walls, etc., and can take various patterns, so by setting these as a set, it is possible to reduce the total number of combinations of the conditions of each boundary portion 7.

[0086] In the next step S3, the processing circuitry 17 sets the object 3 (FIG. 1) placed in the space 2. For example, the processing circuitry 17 sets the object 3 such as furniture or a person.

[0087] Here, a pattern is set for the arrangement set of the objects 3. The arrangement set of the objects 3 includes, for example, the basic shapes, positions, and types of the objects 3.

[0088] A plurality of patterns can be set, not just one, for the arrangement set of the object 3. The order of setting the boundary portion 7 and setting the object 3 does not matter.

[0089] In the next step S4, the processing circuitry 17 sets at least one representative case indicating a representative position of the boundary portion 7 and the object 3. For example, the processing circuitry 17 combines a layout set of the boundary portion 7 and a layout set of the object 3 to set as a representative case. It is possible to set a plurality of representative cases, rather than one type of representative case.

[0090] In the next step S5, the processing circuit 17 sets an operation condition indicating at least one operation of the boundary portion 7.

[0091] A numerical analysis (simulation) is performed for the set representative case, and at that time it is necessary to specify the operating conditions of the boundary part 7. A condition range is set corresponding to each boundary part 7, and it is desirable to cover that range with a small number of analyses. For example, an experimental design method is used, and the operating conditions of the boundary part 7 are set so that the condition range of each boundary part 7 included in the representative case is distributed among the analysis conditions.

[0092] In the next step S6, the processing circuitry 17 performs a numerical analysis of the behavior of the fluid in the representative case. The numerical analysis is, for example, an analysis by simulation.

[0093] In the next step S7, the processing circuitry 17 extracts the main flow 40 occurring in the space 2 from the results of the numerical analysis.

[0094] In the next step S8, the processing circuit 17 determines whether the extracted mainstream 40 satisfies the condition range assumed for machine learning. If the mainstream 40 satisfies the condition range (YES in step S8), the process proceeds to step S9. On the other hand, if the mainstream 40 does not satisfy the condition range (NO in step S8), the process proceeds to step S13. The condition range includes any threshold value previously determined by the user.

[0095] In step S9, which is reached if the answer is YES in step S8, the processing circuitry 17 extracts a local flow 41 occurring in the space 2 from the results of the numerical analysis.

[0096] In the next step S10, the processing circuitry 17 classifies the multiple local flows 41 extracted from the results of the numerical analysis according to the flow patterns. The flow patterns of the local flows 41 include, for example, Karman vortices (FIG. 9), vortices other than Karman vortices (FIG. 10), and flows along the surface of the object 3 (FIG. 11). Here, the processing circuitry 17 classifies the local flows 41 by, for example, clustering. Alternatively, the user may classify the local flows 41.

[0097] In the next step S11, the processing circuit 17 judges whether the classified local flow 41 satisfies the condition range assumed for machine learning. This judgment is performed for each classification of the local flow 41. Here, if the local flow 41 satisfies the condition range (YES in step S11), the process proceeds to step S12. On the other hand, if the local flow 41 does not satisfy the condition range (NO in step S11), the process proceeds to step S13.

[0098] In step S13, which is reached if step S8 is NO or step S11 is NO, the processing circuit 17 sets an analysis case different from the representative case. This analysis case includes at least one of the position of the boundary portion 7, the position of the object 3, and the operating condition. Then, returning to step S6, the processing circuit 17 performs a numerical analysis of the behavior of the fluid in the analysis case. In this way, when there is insufficient training data for machine learning, new training data can be obtained by performing additional numerical analysis of the analysis case.

[0099] For example, even if a window is fully opened with the intention of exhausting air, there may be cases where air is exhausted from another exhaust fan 8 and no air flows out from the window. In such cases, an analysis case is created in which the exhaust fan 8 is stopped, and an additional analysis is performed.

[0100] In step S11, the processing circuitry 17 determines whether or not the flow conditions of the local flows 41 corresponding to the classifications have been obtained for all types of objects 3. If there are insufficient objects 3, in step S13, the processing circuitry 17 creates an analysis case in which the missing objects 3 are replaced.

[0101] In step S12, which is reached when the answer is YES in step S11, the processing circuit 17 generates teacher data based on the extracted mainstream 40 and local flows 41. In this way, the teacher data for machine learning can be generated while preventing an increase in the amount of data. The processing circuit 17 also generates teacher data corresponding to the classified local flows 41. In this way, the local flows 41 are classified according to various flow modes, and the amount of data of the teacher data can be reduced by reducing the pattern of overlapping local flows 41. Then, the teacher data generation process is completed. The processing circuit 17 performs machine learning of the prediction model 20 using the generated teacher data.

[0102] Second embodiment Next, a second embodiment will be described with reference to Fig. 13. Note that the same components as those shown in the above-mentioned embodiment will be denoted by the same reference numerals and duplicated description will be omitted.

[0103] The teacher data generation process of the second embodiment differs from the teacher data generation process of the first embodiment (FIG. 12) only in step S5A, and the other steps are similar to those of the first embodiment.

[0104] In step S5A following step S5, the processing circuit 17 excludes from the numerical analysis, among the various items set, items that have already been learned by the prediction model 20 (FIG. 3). This exclusion process is performed before the numerical analysis is performed. Then, the processing circuit 17 proceeds to step S6 described above. Here, the various items include the shape and size of the space 2, the boundary portion 7, the object 3, the representative case, the analysis case, and the operating conditions.

[0105] In this way, the number of items to be subjected to the numerical analysis can be reduced, and the processing load related to the numerical analysis can be reduced. Note that the determination of whether or not an item has been learned by the prediction model 20 may be performed by the processing circuitry 17 or by the user.

[0106] The flow conditions in a space 2 such as a room differ for each prediction target in terms of the shape of the space 2, the placement of objects or people inside, the placement of entrances and windows, etc., so it is difficult to prepare a general-purpose trained prediction model 20. However, it is possible to prepare a prediction model 20 that has been trained in part in advance. When using such a prediction model 20, it is sufficient to create training data only for the untrained range. Therefore, in the second embodiment, the amount of data in the generated training data can be reduced by excluding items that have already been trained.

[0107] Third embodiment Next, a third embodiment will be described with reference to Fig. 14. Note that the same components as those shown in the above-described embodiments are denoted by the same reference numerals and will not be described again.

[0108] The teacher data generation process of the third embodiment differs from the teacher data generation process of the first embodiment (FIG. 12) only in step S10A, and the other steps are similar to those of the first embodiment.

[0109] In step S10A following step S10, the processing circuit 17 excludes the local flows 41 that have been learned by the prediction model 20 (FIG. 3) from among the local flows 41 extracted from the results of the numerical analysis from the targets for generating teacher data. Then, the processing circuit 17 proceeds to step S11 described above, and determines the degree of satisfaction for each classification of the local flows 41.

[0110] In this way, by using the trained prediction model 20 to classify the local flows 41 for which training data is to be generated, not only is the processing load for determining the degree of satisfaction for each classification of the local flows 41 reduced, but the amount of training data to be generated can also be reduced. Note that the determination of whether a local flow 41 is trained in the prediction model 20 may be made by the processing circuit 17 or by the user.

[0111] (Fourth embodiment) Next, a fourth embodiment will be described with reference to Fig. 15. Note that the same components as those shown in the above-mentioned embodiments are given the same reference numerals and the duplicated description will be omitted.

[0112] The teacher data generation process of the fourth embodiment differs from the teacher data generation process of the first embodiment (FIG. 12) only in step S4A, and the other steps are similar to those of the first embodiment.

[0113] In step S4A following step S3, the processing circuitry 17 sets an analysis case. This analysis case includes at least one of the position of the boundary portion 7, the position of the object 3, and the motion condition. Then, the processing circuitry 17 proceeds to the above-mentioned step S5, and performs processing using the analysis case instead of the above-mentioned representative case.

[0114] The computer 10 of the fourth embodiment has a function of automatically generating an analysis case. Therefore, the user can be saved from the trouble of generating an analysis case instead of a representative case. The computer 10 can set an analysis case in addition to or as a replacement for the representative case. The computer 10 can create an analysis case that covers the entire range and conditions from the data of the arrangement set of the boundary portion 7 or the arrangement set of the object 3, instead of the setting by the user. In this way, the work by the user can be reduced.

[0115] If the amount of numerical analysis is increased as a result of the generation of the analysis cases by the computer 10, the user may generate the analysis cases. In this way, the amount of numerical analysis can be reduced.

[0116] Fifth embodiment Next, a fifth embodiment will be described with reference to Fig. 16. Note that the same components as those shown in the above-mentioned embodiments are given the same reference numerals and the duplicated description will be omitted.

[0117] The teacher data generation process of the fifth embodiment differs from the teacher data generation process of the fourth embodiment (FIG. 15) only in step S4B, and the other steps are similar to those of the fourth embodiment.

[0118] In step S4B following step S4A, the processing circuitry 17 excludes from the generated analysis cases those analysis cases that have been trained by the prediction model 20 (FIG. 3) from the target of the numerical analysis. This exclusion process is performed before the numerical analysis. Then, the processing circuitry 17 proceeds to the above-mentioned step S5.

[0119] In this way, the trained prediction model 20 is used to classify the analysis cases for which training data is to be generated, thereby making it possible to reduce the amount of training data to be generated.

[0120] The present invention has been described above based on the first to fifth embodiments, but the configuration applied in any of the embodiments may be applied to other embodiments, and the configurations applied in each embodiment may be combined.

[0121] In the above-mentioned flowchart, the steps are executed in series, but the order of the steps is not necessarily fixed, and the order of some steps may be interchanged. Also, some steps may be executed in parallel with other steps.

[0122] The above-mentioned predictive model generation system 1 includes a control device, a storage device, an output device, an input device, and a communication interface. Here, the control device includes a highly integrated processor such as a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), or a dedicated chip. The storage device includes a read only memory (ROM), a random access memory (RAM), a hard disk drive (HDD), a solid state drive (SSD), and the like. The output device includes a display panel, a head mounted display, a projector, a printer, and the like. The input device includes a mouse, a keyboard, a touch panel, and the like. This predictive model generation system 1 can be realized by a hardware configuration using a normal computer 10.

[0123] The program or teacher data executed by the above-mentioned prediction model generation system 1 is provided by being pre-installed in a ROM or the like. Additionally or alternatively, the program or teacher data is provided by being stored in a non-transitory storage medium readable by the computer 10 as a file in an installable or executable format. Examples of the storage medium include CD-ROMs, CD-Rs, memory cards, DVDs, and flexible disks (FDs).

[0124] Furthermore, the programs or teacher data executed by this prediction model generation system 1 may be stored in a computer 10 connected to a network such as the Internet and provided by downloading via the network. Furthermore, this prediction model generation system 1 can also be configured by combining separate modules that independently fulfill the functions of the components and connect them to each other via a network or dedicated lines.

[0125] The aforementioned network includes predetermined communication lines such as the Internet, a Local Area Network (LAN), a Wide Area Network (WAN), and a mobile communication network.

[0126] In the above embodiment, a gas such as air is exemplified as the fluid, but other modes may be used. For example, the prediction model 20 may be used to predict the behavior of a liquid such as water.

[0127] In the above embodiment, the interior of a room is exemplified as the space 2 to be predicted, but other aspects may be used. For example, the space 2 to be predicted may be the interior of a specific manufacturing device, and the future state of a fluid therein may be predicted by the prediction model 20.

[0128] In the above embodiment, the exhaust fan 8 is exemplified as the predetermined device provided in the space 2 to be predicted, but other modes may be used. For example, the predetermined device may be a hydrogen removal device that is activated when the hydrogen concentration exceeds a threshold value, or a cooling device that is activated when the temperature of the fluid exceeds a threshold value.

[0129] In the above embodiment, the exhaust fan 8 is provided on the ceiling 6, but other configurations are also possible. For example, the exhaust fan 8 may be provided in the center of the space 2 (room). In that case, the exhaust fan 8 itself becomes the boundary part 7, and the fluid before passing through the exhaust fan 8 is not affected by an external force, and the fluid after passing through the exhaust fan 8 is affected by an external force.

[0130] According to at least one of the embodiments described above, the training data includes data indicating at least one mainstream flow 40 and data indicating at least one local flow 41. Machine learning can be performed efficiently while suppressing an increase in training data for machine learning corresponding to the shape of the space 2 that differs for each prediction target and the object 3 therein.

[0131] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, modifications, and combinations can be made without departing from the spirit of the invention. These embodiments or modifications thereof are within the scope of the invention and its equivalents as described in the claims, as well as within the scope and spirit of the invention. [Explanation of symbols]

[0132] 1...prediction model generation system, 2...space, 3...object, 4...floor, 5...wall, 6...ceiling, 7...boundary, 8...exhaust fan, 10...computer, 11...fluid data acquisition unit, 12...spatial data acquisition unit, 13...display unit, 14...memory unit, 15...input unit, 16...communication unit, 17...processing circuit, 20...prediction model, 21...input layer, 22...middle layer, 23...output layer, 30...screen, 31...contour, 40...main flow, 41...local flow.

Claims

1. A training data set is obtained by inputting at least one of data of a first state showing at least a shape of a space and a behavior of a fluid in the space or data simulating the first state, and outputting at least one of data of a second state showing a behavior of the fluid in a future from the first state or data simulating the second state; Performing machine learning of parameters using the training data, generating a prediction model using at least the first state data acquired by a data acquisition unit as an input and the second state data as an output; A method for performing a process on one or more computers, The teacher data is data indicative of at least one main flow induced in the fluid within the space by the effect of an external force; data indicative of at least one local flow resulting from the main flow impinging on at least one object present in the space; and Including, Methods for generating predictive models.

2. The computer separately performs a process of learning the mainstream and a process of learning the local flow in at least a part of the machine learning process. The method of generating a predictive model according to claim 1 .

3. At least one of the computer and one or more users who operate the computer A shape and a size of the space to be predicted as a behavior of the fluid are set; setting at least one boundary portion which is a boundary where the fluid in the space is affected by the external force; determining at least one of the objects disposed within the space; establishing at least one representative case that indicates a representative location of the boundary and the object; setting an operating condition indicative of at least one operation of the boundary portion; performing a numerical analysis of the behavior of the fluid in the representative case; extracting at least one of the main flow and the local flow occurring in the space from the result of the numerical analysis; generating the training data based on at least one of the extracted main flow and the extracted local flow; The method of generating a predictive model according to claim 1 .

4. At least one of the computer and the user classifying the local flows extracted from the results of the numerical analysis according to the state of the flows; generating the training data corresponding to the classified local flows; The method of generating a predictive model according to claim 3 .

5. At least one of the computer and the user determining whether or not at least one of the mainstream flow and the local flow extracted from the result of the numerical analysis satisfies a condition range assumed for performing the machine learning; If the condition range is not satisfied, at least one analysis case that indicates a position of the boundary portion and the object, which is different from the representative case, is set; performing the numerical analysis of the behavior of the fluid in the analysis case; extracting at least one of the main flow and the local flow occurring in the space from the result of the numerical analysis; generating the training data based on at least one of the extracted main flow and the extracted local flow; The method for generating a prediction model according to claim 3 or 4.

6. The computer has a function of automatically generating at least one analysis case indicating the boundary portion and the position of the object; At least one of the computer and the user sets the analysis case generated by the computer as an addition or replacement of the representative case. The method for generating a prediction model according to claim 3 or 4.

7. At least one of the computer and the user excluding the analysis cases that have already been trained by the prediction model from the targets of the numerical analysis before performing the numerical analysis; The method of generating a predictive model according to claim 6.

8. At least one of the computer and the user Prior to performing the numerical analysis, among the various items set, items that have been learned by the prediction model are excluded from the target of the numerical analysis. The method for generating a prediction model according to claim 3 or 4.

9. At least one of the computer and the user excluding the local flows that have been learned by the prediction model from among the local flows extracted from the result of the numerical analysis from targets for generating the training data; The method for generating a prediction model according to claim 3 or 4.

10. A training data set is obtained by inputting at least one of data of a first state showing at least a shape of a space and a behavior of a fluid in the space or data simulating the first state, and outputting at least one of data of a second state showing a behavior of the fluid in a future from the first state or data simulating the second state; Performing machine learning of parameters using the training data, generating a prediction model using at least the first state data acquired by a data acquisition unit as an input and the second state data as an output; One or more computers; The teacher data is data indicative of at least one main flow induced in the fluid within the space by the effect of an external force; data indicative of at least one local flow resulting from the main flow impinging on at least one object present in the space; and Including, Predictive model generation system.

11. The teaching data is an input of at least one of data of a first state showing at least a shape of a space and a behavior of a fluid in the space or data simulating the first state, and an output of at least one of data of a second state showing at least a behavior of the fluid in a future from the first state or data simulating the second state, data indicative of at least one main flow induced in the fluid within the space by the effect of an external force; data indicative of at least one local flow resulting from the main flow impinging on at least one object present in the space; and Including, One or more computers Perform machine learning of parameters, generating a prediction model using at least the first state data acquired by a data acquisition unit as an input and the second state data as an output; Used in processing, Training data for generating predictive models.

12. an input layer to which first state data indicating at least a shape of a space and a behavior of a fluid in the space is input; an output layer that outputs data of a second state that indicates a behavior of the fluid at least in the future relative to the first state; an intermediate layer in which parameters are machine-learned using teacher data that receives as input at least one of the first state data or data simulating the first state data and outputs at least one of the second state data or data simulating the second state data; Equipped with data indicative of at least one main flow induced in the fluid within the space by the effect of an external force; data indicative of at least one local flow resulting from the main flow impinging on at least one object present in the space; and The training data includes The first state data acquired by at least a data acquisition unit is input to the input layer, and an operation is performed in the intermediate layer, and the second state data is output from the output layer. to cause one or more computers to function in such a manner that A trained predictive model.

Images