Network learning system, environmental distribution map estimation system, and environmental distribution map estimation method

Abstract

Reducing the number of sensors used for estimating the environmental distribution map and shortening the time required for estimating the environmental distribution map. 【Solution means】 A control unit is provided. The control unit acquires a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, and acquires second environmental information, which is environmental information measured by a second sensor group among the first sensor groups. A network learning system that causes a GAN (Generative Adversarial Network) to learn such that a second environmental distribution map output by a generator approaches the first environmental distribution map by using the first environmental distribution map as an input to the discriminator of the GAN and the second environmental information as an input to the generator of the GAN.

Description

Technical Field

[0001] The present disclosure relates to a network learning system, an environmental distribution map estimation system, and an environmental distribution map estimation method.

Background Art

[0002] Non-Patent Document 1 describes the results of examining the differences in the indoor environment between radiant air conditioning and convective air conditioning, and the effects of the differences in the indoor environment on skin surface temperature, psychological responses, and intellectual productivity.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the method of arranging a large number of sensors and measuring environmental information with each sensor (hard sensing), problems such as physical constraints in sensor arrangement and high sensor costs occur. In the method of estimating an environmental distribution map without arranging sensors (CFD analysis), it takes time for the estimation, and there is a possibility that the estimation result is greatly deviated from the actual situation.

[0005] An object of the present disclosure is to reduce the number of sensors used for estimating an environmental distribution map and shorten the time required for estimating the environmental distribution map.

Means for Solving the Problems

[0006] The network learning system of the present disclosure includes a control unit. The control unit acquires a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, acquires second environmental information, which is environmental information measured by a second sensor group among the first sensor group, and uses a GAN (Generative Adversarial Network) with the first environmental distribution map as the input of the discriminator of the GAN and the second environmental information as the input of the generator of the GAN to train the generator so that the second environmental distribution map output by the generator approaches the first environmental distribution map. According to this network learning system, it is possible to reduce the number of sensors used for estimating the environmental distribution map and shorten the time required for estimating the environmental distribution map.

[0007] The GAN may be a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), and the control unit may train the cDCGAN by using the condition regarding the second environmental distribution map as a further input of the generator of the cDCGAN. In this way, it is possible to reduce the deviation between the estimated environmental distribution map and the actual environmental distribution map and improve the quality of the estimated environmental distribution map.

[0008] The environmental information may include any one of temperature, radiant temperature, humidity, wind speed, and carbon dioxide concentration. In this way, it is possible to reduce the number of sensors used for estimating the distribution map of any one of temperature, radiant temperature, humidity, wind speed, and carbon dioxide concentration and shorten the time required for estimating such a distribution map.

[0009] In addition, the environmental distribution map estimation system of the present disclosure includes a control unit, and the control unit is a GAN (Generative Adversarial Network). The first environmental distribution map generated from the first environmental information, which is the environmental information measured by the first sensor group, is used as the input to the discriminator of the GAN, and the second environmental information, which is the environmental information measured by the second sensor group among the first sensor group, is used as the input to the generator of the GAN. Thus, information on the generator of the GAN, which is learned so that the second environmental distribution map output by the generator approaches the first environmental distribution map, is obtained. By using the third environmental information, which is the environmental information measured by the second sensor group, as the input to the learned generator of the GAN, a third environmental distribution map output by the generator is obtained as the estimated environmental distribution map. This is an environmental distribution map estimation system. According to this environmental distribution map estimation system, it is possible to reduce the number of sensors used for estimating the environmental distribution map and shorten the time required for estimating the environmental distribution map.

[0010] The GAN is a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), and the control unit may obtain information on the learned generator of the cDCGAN by using the condition regarding the second environmental distribution map as an additional input to the generator of the cDCGAN. In this way, it is possible to reduce the deviation between the estimated environmental distribution map and the actual environmental distribution map and improve the quality of the estimated environmental distribution map.

[0011] Furthermore, in the environmental distribution map estimation method of the present disclosure, a control unit of a computer acquires a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, acquires second environmental information, which is environmental information measured by a second sensor group among the first sensor group, and uses a GAN (Generative Adversarial Network) with the first environmental distribution map as an input to a discriminator of the GAN and the second environmental information as an input to a generator of the GAN, so as to train the generator such that a second environmental distribution map output by the generator approaches the first environmental distribution map. Then, by using third environmental information, which is environmental information measured by the second sensor group, as an input to the generator of the trained GAN, a third environmental distribution map output by the generator is acquired as an estimated environmental distribution map. According to this environmental distribution map estimation method, it is possible to reduce the number of sensors used for estimating the environmental distribution map and shorten the time required for estimating the environmental distribution map.

[0012] The GAN may be a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), and the control unit may train the cDCGAN by using a condition related to the second environmental distribution map as a further input to the generator of the cDCGAN. In this way, it is possible to reduce the deviation between the estimated environmental distribution map and the actual environmental distribution map and improve the quality of the estimated environmental distribution map.

Brief Description of the Drawings

[0013]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

[0014] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

[0015] [Overall Configuration of Distribution Map Generation System] FIG. 1 is a diagram showing an example of the overall configuration of a distribution map generation system 10 according to the present embodiment. The distribution map generation system 10 is a system that generates a thermal environment distribution map in a target space 100. The thermal environment distribution map is a distribution map of thermal environment information in the target space 100. Here, the thermal environment information includes any one of temperature, radiant temperature, humidity, wind speed, and carbon dioxide concentration. Incidentally, the thermal environment information and the thermal environment distribution map may be more generally regarded as environment information and an environment distribution map, but will be described below as the thermal environment information and the thermal environment distribution map. As shown in the drawing, the distribution map generation system 10 includes an air conditioner 200 and a control device 300. The distribution map generation system 10 further includes sensors 400a to 400t, an input / output device 500, and an information processing device 600.

[0016] The target space 100 is a space that is the target of the thermal environment distribution map generated by the distribution map generation system 10. The target space 100 is a space surrounded by a wall 110, a door 120, a floor 130, and a ceiling (not shown). And a window 140 is provided in the wall 110. Also, indoor units 210a to 210c of the air conditioner 200 described later and the input / output device 500 described later are installed in the target space 100. Further, although not shown, lighting, a personal computer, a monitor, a printer, a desk, a chair, a sofa, a table, a rug, etc. are also arranged in the target space 100.

[0017] The air conditioner 200 is a device that conditions the air in the target space 100. The air conditioner 200 includes indoor units 210a to 210c, an outdoor unit 220, and a pipe 230. The indoor units 210a to 210c are installed in the target space 100 and perform heat exchange between the refrigerant that has passed through the pipe 230 and the air in the target space 100, thereby absorbing heat from the air in the target space 100 or discharging heat into the target space 100. The outdoor unit 220 is installed outside the target space 100 and performs heat exchange between the refrigerant that has passed through the pipe 230 and the air outside the target space 100, thereby discharging heat outside the target space 100 or absorbing heat from the air outside the target space 100. The pipe 230 is a pipe that connects the indoor units 210a to 210c and the outdoor unit 220, and the refrigerant passes through its interior. In the figure, the indoor units 210a to 210c are shown, but when there is no need to distinguish them, they may also be referred to as the indoor unit 210. Although three indoor units 210 are shown in the figure, one, two, or four or more indoor units 210 may be provided.

[0018] The control device 300 is a device that controls the air conditioner 200 to operate based on the set conditions.

[0019] Sensors 400a to 400t are installed at each position within the target space 100 and measure the thermal environment information of each position. Sensors 400a to 400t are used to collect learning data. Sensors 400a to 400t constitute the first sensor group. Among sensors 400a to 400t, sensors 400a to 400e are sensors for constant installation for estimation. Sensors 400a to 400e constitute the second sensor group. And among sensors 400a to 400t, the remaining sensors 400f to 400t are recovered after the learning data is collected. Incidentally, in the figure, sensors 400a to 400t are shown, but when there is no need to distinguish them, they may also be referred to as sensor 400. Here, 20 sensors 400 are shown as sensors 400 for collecting learning data, but this is merely an example. As sensors 400 for collecting learning data, fewer than 20 or more than 20 sensors 400 may be provided. Also, here, 5 sensors 400 are shown as sensors 400 for constant installation, but this is also merely an example. As sensors 400 for constant installation, fewer than 5 or more than 5 sensors 400 may be provided. Furthermore, here, the sensors 400 for constant installation are provided on the wall 110, but this is also merely an example. The sensors 400 for constant installation may be provided at any location indoors.

[0020] The input / output device 500 is arranged, for example, within the target space 100 and is used to give an instruction to the information processing device 600. Also, the input / output device 500 outputs the thermal environment distribution map generated by the information processing device 600. The input / output device 500 may be, for example, a touch panel type operation display device.

[0021] The information processing device 600 acquires a thermal environment distribution map generated from the thermal environment information measured by the sensors 400a to 400t. The information processing device 600 acquires the thermal environment information measured by the sensors 400a to 400e. Then, based on these data, the information processing device 600 trains a cDCGAN (conditional convolutional adversarial network). Thereby, the information processing device 600 creates a real-time thermal environment distribution map of the entire room using the thermal environment information measured by a small number of sensors 400 arranged indoors.

[0022] In such a distribution map generation system 10, in the present embodiment, the user performs the following operations. First, the user installs the sensors 400a to 400t for a certain period and measures the thermal environment information with the sensors 400a to 400t. Then, the user causes the information processing device 600 to generate a thermal environment distribution map from the thermal environment information measured by the sensors 400a to 400t. After that, the user leaves the sensors 400a to 400e for permanent installation and collects the other sensors 400f to 400t. Second, the user provides the information processing device 600 with the thermal environment distribution map generated from the thermal environment information measured by the sensors 400a to 400t. The user also provides the information processing device 600 with the thermal environment information measured by the sensors 400a to 400e. Thereby, the user causes the information processing device 600 to train the cDCGAN using the provided thermal environment distribution map and thermal environment information. Third, the user measures the thermal environment information with the sensors 400a to 400e for permanent installation. Then, the user inputs this thermal environment information into the cDCGAN to obtain and output a thermal environment distribution map in real time.

[0023] [Hardware Configuration of Information Processing Device] FIG. 2 is a diagram showing a hardware configuration example of the information processing apparatus 600 in the present embodiment. As shown in the figure, the information processing apparatus 600 includes a processor 601 which is an arithmetic means. The information processing apparatus 600 further includes a RAM (Random Access Memory) 602, a ROM (Read Only Memory) 603, and a storage device 604 which are storage means. The RAM 602 is a main storage device (main memory), and is used as a working memory when the processor 601 performs arithmetic processing. Programs and data such as preset setting values are held in the ROM 603, and the processor 601 reads programs and data directly from the ROM 603 and executes processing. The storage device 604 is a storage means for programs and data. A program is stored in the storage device 604, and the processor 601 reads the program stored in the storage device 604 into the main storage device and executes it. Further, the results of processing by the processor 601 are stored and saved in the storage device 604. As the storage device 604, for example, a magnetic disk device, an SSD (Solid State Drive), or the like is used.

[0024] [Functional Configuration of Information Processing Apparatus] FIG. 3 is a block diagram showing a functional configuration example of the information processing apparatus 600 in the present embodiment. As shown in the figure, the information processing apparatus 600 includes a first information acquisition unit 610, a first distribution map acquisition unit 620, and a second information acquisition unit 630. The information processing apparatus 600 further includes a cDCGAN storage unit 640, a cDCGAN learning unit 650, and a generator acquisition unit 660. The information processing apparatus 600 further includes a third information acquisition unit 670 and a third distribution map acquisition unit 680.

[0025] The first information acquisition unit 610 acquires the thermal environment information (hereinafter referred to as "first thermal environment information") measured by the sensors 400a to 400t constituting the first sensor group for learning. Specifically, the first information acquisition unit 610 may acquire the first thermal environment information from the sensors 400a to 400e among the sensors 400a to 400t via a communication line. This is because the sensors 400a to 400e are also used for estimating the thermal environment distribution map, and it is preferable to be able to acquire the first thermal environment information in real time at that time. Here, the communication line may be wired or wireless. Also, the first information acquisition unit 610 may directly acquire the first thermal environment information from the sensors 400f to 400t among the sensors 400a to 400t. For example, the first information acquisition unit 610 may acquire the first thermal environment information by directly connecting the sensors 400f to 400t to the information processing device 600. Since the sensors 400f to 400t are collected after the collection of learning data, there is no need to construct communication facilities or perform communication settings. However, the first information acquisition unit 610 may also acquire the first thermal environment information from the sensors 400f to 400t via a communication line.

[0026] The first thermal environment information is an example of the first environmental information that is the environmental information measured by the first sensor group.

[0027] The first distribution map acquisition unit 620 acquires the first thermal environment distribution map generated from the first thermal environment information acquired by the first information acquisition unit 610. The first thermal environment distribution map may be generated by performing linear interpolation or non-linear interpolation for each point where the first thermal environment information has not been acquired. As the linear interpolation, for example, the two-dimensional bilinear method is used. The first distribution map acquisition unit 620 may also acquire the first thermal environment distribution map by generating the first thermal environment distribution map from the first thermal environment information.

[0028] The first thermal environment distribution map is an example of the first environmental distribution map generated from the first environmental information. The process of the first distribution map acquisition unit 620 is an example of the process of acquiring the first environmental distribution map.

[0029] The second information acquisition unit 630 acquires thermal environment information (hereinafter referred to as "second thermal environment information") measured by sensors 400a to 400e constituting the second sensor group for learning. Specifically, the second information acquisition unit 630 acquires the second thermal environment information by extracting the second thermal environment information from the first thermal environment information acquired by the first information acquisition unit 610.

[0030] The second thermal environment information is an example of second environment information that is environment information measured by the second sensor group among the first sensor groups. The process of the second information acquisition unit 630 is an example of the process of acquiring the second environment information.

[0031] The cDCGAN storage unit 640 stores the cDCGAN. Note that, as will be described later, instead of the cDCGAN storage unit 640, a cGAN storage unit or a GAN storage unit may be provided. However, hereinafter, it will be described assuming that the cDCGAN storage unit 640 is provided. FIG. 4 shows the configuration of the cDCGAN 700. The cDCGAN 700 includes a generator 730 and a discriminator 770.

[0032] When the noise 710 and the label 720 are input, the generator 730 outputs a generated image 740. The generator 730 is composed of a transposed convolutional layer, a batch normalization layer, and a ReLU activation function. The transposed convolutional layer converts the noise 710 into a high-resolution image. The batch normalization layer normalizes the output of each layer to improve the stability of learning. The ReLU activation function introduces non-linearity and helps in learning more complex relationships. Finally, the tanh function outputs a colored image in the range from "-1" to "1". Also, a conditional vector indicating the conditions regarding the generated image 740 is given to the generator 730 as the label 720. The conditional vector may be, for example, the average temperature of each point of the thermal environment distribution map output as the generated image 740.

[0033] When the discriminator 770 receives the generated image 740, the real image 750, and the label 760 as inputs, it outputs a discrimination result 780 indicating whether the generated image 740 is real or fake. The discriminator 770 is composed of a convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The convolutional layer extracts features from the input image. The batch normalization layer performs normalization. The LeakyReLU activation function alleviates the problem of vanishing gradients. The discriminator 770 finally outputs a probability as the discrimination result 780 through a sigmoid function. Also, a conditional vector is given to the discriminator 770 as the label 720. With such a configuration, the discriminator 770 determines the authenticity of the generated image 740 generated by the generator 730.

[0034] Returning again to FIG. 3, the description of the functional configuration example of the information processing apparatus 600 will be continued. The cDCGAN learning unit 650 inputs the second thermal environment information acquired by the second information acquisition unit 630 into the generator 730 as noise 710. A conditional vector is also input to the generator 730 as the label 720. Then, the generator 730 outputs a second thermal environment distribution map as the generated image 740. Also, the cDCGAN learning unit 650 inputs the first thermal environment distribution map acquired by the first distribution map acquisition unit 620 into the discriminator 770 as the real image 750. The generated image 740 and the label 760 are also input to the discriminator 770. Then, the discriminator 770 outputs a discrimination result 780 indicating whether the generated image 740 is real or fake. Thereby, the cDCGAN learning unit 650 trains the cDCGAN 700 so that the generated image 740 output by the generator 730 approaches the real image 750.

[0035] The second thermal environment distribution map output by the generator 730 as the generated image 740 is an example of the second environment distribution map output by the generator. The process of the cDCGAN learning unit 650 is an example of a process of training the GAN so that the second environment distribution map approaches the first environment distribution map by using the first environment distribution map as the input to the discriminator of the GAN and the second environment information as the input to the generator of the GAN. In addition, the conditional vector input to the generator 730 as the label 720 is an example of the conditions regarding the second environmental distribution diagram. The process of the cDCGAN learning unit 650 is an example of a process of training the cDCGAN by using the conditions regarding the second environmental distribution diagram as an additional input to the generator.

[0036] The generator acquisition unit 660 acquires information regarding the generator 730 from the cDCGAN 700 trained by the cDCGAN learning unit 650.

[0037] The process of the generator acquisition unit 660 is an example of a process of acquiring information of the generator of the GAN trained such that the second environmental distribution diagram approaches the first environmental distribution diagram by using the first environmental distribution diagram as the input to the discriminator of the GAN and the second environmental information as the input to the generator of the GAN.

[0038] The third information acquisition unit 670 acquires thermo-environmental information (hereinafter referred to as "third thermo-environmental information") measured by the sensors 400a to 400e constituting the second sensor group for estimation. Specifically, the third information acquisition unit 670 may acquire the third thermo-environmental information from the sensors 400a to 400e via a communication line. Here, the communication line may be wired or wireless.

[0039] The third thermo-environmental information is an example of the third environmental information which is environmental information measured by the second sensor group.

[0040] The third distribution diagram acquisition unit 680 inputs the third thermo-environmental information acquired by the third information acquisition unit 670 to the generator 730 as the noise 710. A conditional vector is also input to the generator 730 as the label 720. Then, the generator 730 outputs a thermo-environmental distribution diagram (hereinafter referred to as "third thermo-environmental distribution diagram") as the generated image 740. The third distribution diagram acquisition unit 680 acquires the third thermo-environmental distribution diagram output by the generator 730 as the estimated thermo-environmental distribution diagram.

[0041] The third thermal environment distribution map output by the generator 730 as the generated image 740 is an example of the third environment distribution map output by the generator. The process of the third distribution map acquisition unit 680 is an example of a process of acquiring the third environment distribution map as an estimated environment distribution map by using the third environment information as the input of the learned GAN generator.

[0042] In addition, the information processing apparatus 600 may have a configuration for performing matching between the estimated thermal environment distribution map and the individual's thermal preference. Such a configuration includes, although not shown in the drawings, for example, a control information output unit that outputs control information for matching to the control apparatus 300. In this case, the control information output unit first receives the target thermal environment information of the individual's area in the room. Next, the control information output unit determines whether the thermal environment information of the individual's area in the estimated thermal environment distribution map deviates from the target thermal environment information of that area. Next, if it is determined that those thermal environment information deviate, the control information output unit outputs control information for reducing the deviation to the control apparatus 300. For example, when the air conditioner 200 is performing heating operation, assume that the temperature of the individual's area in the estimated thermal environment distribution map is lower than the target temperature of that area. Then, the control information output unit outputs control information to the control apparatus 300 to direct the air flow direction of the air conditioner 200 toward that area and increase the air volume of the air conditioner 200. Thereby, the possibility of obtaining a comfortable and energy-saving air conditioning system for all people in the room is increased.

[0043] [Operation of Information Processing Apparatus] FIG. 5 is a flowchart showing an operation example of the information processing apparatus 600 in the present embodiment.

[0044] As shown in the drawing, in the information processing apparatus 600, first, the first information acquisition unit 610 acquires the first thermal environment information (step 801). Here, as described above, the first thermal environment information is the thermal environment information measured by the sensors 400a to 400t constituting the first sensor group for learning.

[0045] Next, in the information processing apparatus 600, the first distribution map acquisition unit 620 acquires the first thermal environment distribution map (step 802). For example, the first distribution map acquisition unit 620 may acquire the first thermal environment distribution map by generating the first thermal environment distribution map from the first thermal environment information acquired in step 801.

[0046] Next, in the information processing apparatus 600, the second information acquisition unit 630 acquires the second thermal environment information (step 803). Here, as described above, the second thermal environment information is the thermal environment information measured by the sensors 400a to 400e constituting the second sensor group for learning. For example, the second information acquisition unit 630 may acquire the second thermal environment information by extracting the second thermal environment information from the first thermal environment information acquired in step 801.

[0047] Next, in the information processing apparatus 600, the cDCGAN learning unit 650 causes the cDCGAN 700 to be learned (step 804). Specifically, the cDCGAN learning unit 650 inputs the first thermal environment distribution map acquired in step 802 into the discriminator 770 as the real image 750. The cDCGAN learning unit 650 inputs the second thermal environment information acquired in step 803 into the generator 730. The cDCGAN learning unit 650 makes the second thermal environment distribution map output as the generated image 740 by the generator 730 approach the first thermal environment distribution map input as the real image 750.

[0048] Thereafter, in the information processing apparatus 600, the generator acquisition unit 660 acquires information regarding the generator 730 from the cDCGAN 700 learned in step 804 (step 805).

[0049] Next, in the information processing apparatus 600, the third information acquisition unit 670 acquires the third thermal environment information (step 806). Here, as described above, the third thermal environment information is the thermal environment information measured by the sensors 400a to 400e constituting the second sensor group for estimation.

[0050] Next, in the information processing apparatus 600, the third distribution map acquisition unit 680 acquires the third thermal environment distribution map (step 807). Specifically, the third distribution map acquisition unit 680 inputs the third thermal environment information acquired in step 806 to the generator 730. The third distribution map acquisition unit 680 acquires the third thermal environment distribution map output as the generated image 740 by the generator 730 as the estimated thermal environment distribution map.

[0051] [Modification Example] In the above, the thermal environment distribution map is estimated using the cDCGAN 700, but it is not limited to this.

[0052] For example, it may be possible to estimate the thermal environment distribution map using cGAN (conditional adversarial generative network). The cDCGAN 700 employs convolutional neural networks (CNNs) for both the generator 730 and the discriminator 770. As a result, higher-quality images can be generated compared to cGAN. However, if such high-quality images are not a concern, cGAN can also be used.

[0053] Alternatively, it may be possible to estimate the thermal environment distribution map using GAN (adversarial generative network). cGAN improves the quality of the generated image by providing a conditional vector to both the generator and the discriminator. For example, cGAN outputs an image closer to the actual thermal environment distribution map as the generated image by providing the indoor average temperature as the conditional vector. However, if such an image close to the actual situation is not a concern, GAN can also be used.

[0054] Also, in the above description, the information processing apparatus 600 performs both the learning of the cDCGAN 700 and the estimation using the cDCGAN 700, but this is not restrictive. The first information processing apparatus may perform the learning of the cDCGAN 700, and the second information processing apparatus may perform the estimation using the cDCGAN 700. In this case, the first information acquisition unit 610, the first distribution map acquisition unit 620, the second information acquisition unit 630, and the cDCGAN learning unit 650 will be provided in the first information processing apparatus. The generator acquisition unit 660, the third information acquisition unit 670, and the third distribution map acquisition unit 680 will be provided in the second information processing apparatus. The cDCGAN storage unit 640 may be provided in the first information processing apparatus, may be provided in the second information processing apparatus, or may be provided in an apparatus different from both of these.

[0055] [Processor] In the case of this embodiment, each process is executed by an arbitrary computer. An arbitrary computer may be realized as a processor as hardware, a program as software, or a combination thereof. An arbitrary computer may be a general-purpose computer, a computer for a specific purpose, a workstation, or any other system capable of executing each process.

[0056] The processor executes various processes in cooperation with a program. The processor can function as each unit in this embodiment. The execution order of the processes by the processor is not limited to the order described in this embodiment and can be changed as necessary. The processor can be configured by one or more pieces of hardware. The type of hardware that constitutes the processor is not limited to a specific type. The processor may be, for example, a CPU (central processing unit) or an MPU (micro processing unit). The processor may be, for example, a programmable logic device such as an FPGA (field programmable gate array). The processor may be, for example, a dedicated circuit for executing specific processing such as an ASIC (application specific integrated circuit). The processor may be, for example, a GPU (graphic processing unit) or hardware such as an NPU (neural processing unit).

[0057] The processor is not limited to a combination of a plurality of pieces of hardware of the same type, but can also be configured by a combination of a plurality of pieces of hardware of different types. When a plurality of pieces of hardware execute one or more processes of a certain processor, the plurality of pieces of hardware may exist in physically separate devices from each other. Alternatively, in this case, the plurality of pieces of hardware may exist in the same device. The hardware is constituted by an electrical circuit (circuitry) that combines circuit elements such as semiconductor elements. In any of the embodiments, the execution order of each process by the processor is not limited to the order described in each embodiment, and can be changed as necessary. The processor is an example of a control unit.

[0058] The program may be software such as microcode in addition to firmware. The program may be, for example, a program module group. Each function constituting the program module group may be realized by a processor that executes each function. The program in each embodiment may be program code or a plurality of code segments stored in one or more non-transitory computer-readable media. Here, the non-transitory computer-readable media may be, for example, semiconductor memory, magnetic or optical storage media, or other storage. The program may be stored separately and divided among a plurality of non-transitory computer-readable media existing in devices physically separated from each other. The program code or a plurality of code segments may be represented by any combination of procedures, functions, subprograms, routines, subroutines, modules, software packages, classes, instructions, data structures, program statements. The program code or a plurality of code segments may be connected to other code segments or hardware circuits by transmitting and receiving information, data, arguments, parameters, or the content of memory.

Explanation of Signs

[0059] 10…Distribution map generation system, 100…Target space, 200…Air conditioner, 300…Control device, 400a~400t…Sensor, 500…Input / output device, 600…Information processing device, 610…First information acquisition unit, 620…First distribution map acquisition unit, 630…Second information acquisition unit, 640…cDCGAN storage unit, 650…cDCGAN learning unit, 660…Generator acquisition unit, 670…Third information acquisition unit, 680…Third distribution map acquisition unit

Claims

1. Comprising a control unit, The control unit, Obtains a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, Obtains second environmental information, which is environmental information measured by a second sensor group among the first sensor group, By using a GAN (Generative Adversarial Network) with the first environmental distribution map as the input to the discriminator of the GAN and the second environmental information as the input to the generator of the GAN, the second environmental distribution map output by the generator is learned to approach the first environmental distribution map, A network learning system.

2. The GAN is a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), The control unit learns the cDCGAN by using a condition regarding the second environmental distribution map as an additional input to the generator of the cDCGAN. The network learning system according to claim 1.

3. The environmental information includes any one of temperature, radiant temperature, humidity, wind speed, and carbon dioxide concentration. The network learning system according to claim 1 or claim 2.

4. Comprising a control unit, The control unit, Is a GAN (Generative Adversarial Network). By using a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, as the input to the discriminator of the GAN and second environmental information, which is environmental information measured by a second sensor group among the first sensor group, as the input to the generator of the GAN, information of the generator of the GAN is obtained, where the second environmental distribution map output by the generator is learned to approach the first environmental distribution map, By using third environmental information, which is environmental information measured by the second sensor group, as the input to the generator of the learned GAN, a third environmental distribution map output by the generator is obtained as an estimated environmental distribution map, An environmental distribution map estimation system.

5. The GAN is a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), The control unit obtains information of the generator of the learned cDCGAN by using a condition regarding the second environmental distribution map as an additional input to the generator of the cDCGAN. The environmental distribution map estimation system according to claim 4.

6. The control unit of a computer, Obtains a first environmental distribution map generated from first environmental information, which is environmental information measured by a first sensor group, Obtain second environmental information, which is environmental information measured by a second sensor group among the first sensor groups, By using a GAN (Generative Adversarial Network), with the first environmental distribution map as the input to the discriminator of the GAN and the second environmental information as the input to the generator of the GAN, train the generator so that the second environmental distribution map output by the generator approaches the first environmental distribution map, Use the third environmental information, which is environmental information measured by the second sensor group, as the input to the generator of the trained GAN, and obtain, as an estimated environmental distribution map, the third environmental distribution map output by the generator, An environmental distribution map estimation method.

7. The GAN is a cDCGAN (Conditional Deep Convolutional Generative Adversarial Network), The control unit trains the cDCGAN by using a condition regarding the second environmental distribution map as a further input to the generator of the cDCGAN. The environmental distribution map estimation method according to claim 6.

Images