Data processing system

Abstract

In the present invention, acquired is data on a target image representing an image obtained by imaging, from above the water surface, a light source in a lighting-emitting state disposed in the water in a water treatment tank. The target image includes a portion representing the light source and a surrounding portion surrounding the light source. The data on the target image is input to a trained machine learning model, thereby acquiring a water quality estimation result. The trained machine learning model has been trained using data on a plurality of training images representing images obtained by imaging the light source disposed in the water with different water qualities. Data representing the estimation result is transmitted to an external device.

Description

Data Processing System

[0001] The present specification relates to a technique for acquiring information related to the water quality of a wastewater treatment device.

[0002] Various wastewater treatment devices are used to treat wastewater. Furthermore, various processes are performed to manage the wastewater treatment devices. For example, Patent Document 1 discloses a technology for detecting abnormalities in the water treatment state based on the color of the water surface obtained from color image data of the water to be treated. Patent Document 1 further discloses the following technology for detecting abnormalities not manifested in the color of the water surface. The imaging device and a first data processing device are connected via a communication line so that data can be transmitted bidirectionally. The imaging device captures images of events occurring in the water to be treated, which is the target of treatment in a septic tank, and events occurring in the septic tank. The first data processing device uses the image data obtained by the imaging to detect abnormalities in the water treatment device. For example, the area of ​​foamy matter appearing on the water surface is detected, and it is determined whether the area of ​​the foamy matter is within a normal range.

[0003] Japanese Patent Application Laid-Open No. 2021-023852

[0004] Information related to water quality is used to manage wastewater treatment equipment, but there is room for improvement in obtaining information related to water quality.

[0005] This specification discloses a technique for acquiring information related to the water quality of a wastewater treatment device.

[0006] The techniques disclosed in this specification can be implemented in the following application examples.

[0007] [Application Example 1] A data processing system comprising: a target image acquisition unit configured to acquire data of a target image representing an image taken from above the water surface of a light source that is placed underwater in a water treatment tank included in a wastewater treatment treatment device, the target image including a portion representing the light source and a peripheral portion surrounding the light source; an estimation unit configured to acquire an estimation result of the water quality of the water in the water treatment tank by inputting the data of the target image into a trained machine learning model, the trained machine learning model being trained to use data of a plurality of training images representing images taken with a light source placed underwater at different water qualities, and output an estimation result indicating a water quality that corresponds to the data of the training images input to the machine learning model; and a transmission unit configured to transmit data representing the estimation result to an external device.

[0008] The appearance of a lit light source placed underwater in a water treatment tank can change with changes in water quality. That is, the appearance of the light source represented by the target image changes with changes in water quality. Therefore, water quality can be estimated by using the target image. Furthermore, the trained machine learning model used to obtain the water quality estimation result is trained to use data from multiple training images taken with different water qualities to output an estimation result indicating the water quality associated with the training image. Therefore, an appropriate water quality estimation result can be obtained. Furthermore, data representing the water quality estimation result is transmitted to an external device, so that the water quality estimation result can be used via the external device.

[0009] [Application Example 2] A data processing system according to Application Example 1, wherein the target image acquisition unit acquires data of the target image by using a specific portion of an image of the light source taken from above the water surface, the specific portion including a portion representing the light source and a surrounding portion surrounding the light source.

[0010] The appearance of the intensity and spread of light from a lit underwater light source can vary depending on the water quality. The specific portion can represent the intensity and spread of light from the light source. When data of the target image is obtained by using the specific portion, the accuracy of the water quality estimation using the data of the target image is improved.

[0011] Application Example 3 In the data processing system according to Application Example 2, the target image acquisition unit selects, as the specific portion, a portion of the captured image that includes an area where a predetermined number or more of pixels that are brighter than a threshold value are consecutive.

[0012] This configuration allows the specific portion to be appropriately selected.

[0013] Application Example 4 The data processing system according to any one of Application Examples 1 to 3, wherein the water treatment tank is a disinfection tank configured to disinfect water.

[0014] This configuration reduces the possibility of contamination adhering to the light source, thereby reducing the possibility of a decrease in the accuracy of estimating water quality.

[0015] Application Example 5 The data processing system according to any one of Application Examples 1 to 4, wherein the water quality includes an index indicating water turbidity.

[0016] The appearance of the light source represented by the target image changes with changes in the turbidity of the water, so by using the target image data, an index indicating the turbidity of the water can be appropriately estimated.

[0017] [Application Example 6] A data processing system according to any one of Application Examples 1 to 5, wherein the trained machine learning model has been trained using a plurality of sets of training image data and water quality measurement results obtained by a specific process including a plurality of set processes, the set processes including photographing a light source that is turned on and placed underwater in a first water treatment tank included in a training wastewater treatment device from above the water surface and measuring the water quality of the water in the first water treatment tank, the training image data representing the photographed image of the light source, and the plurality of set processes including a plurality of specific set processes that are performed at a plurality of times when the water quality of the water in the first water treatment tank differs from each other.

[0018] According to this configuration, a machine learning model is trained to estimate various variations in water quality.

[0019] [Application Example 7] A data processing system according to any one of Application Examples 1 to 6, wherein the light source is detachably attached to the wastewater treatment device, the light source is positioned lower than a predetermined standard water level in the water treatment tank and is visible when observing the water treatment tank from above looking downward, and the light source includes an upward-facing light-emitting diode.

[0020] According to this configuration, it is easy to acquire data of the target image that includes the portion representing the light source and the surrounding portion surrounding the light source.

[0021] Application Example 8 The data processing system according to any one of Application Examples 1 to 7, wherein the wastewater treatment device is a distributed wastewater treatment device that is installed at a wastewater generation site and treats the wastewater.

[0022] According to this configuration, it is possible to obtain an estimated result of the water quality of a wastewater treatment device installed at the site where wastewater is generated.

[0023] [Application Example 9] The data processing system according to Application Example 8, wherein the wastewater treatment device includes a body having a manhole that houses the water treatment tank and a lid that covers the manhole, and the target image represents the image taken inside the body with the manhole covered by the lid.

[0024] This configuration reduces variations in the brightness of the environment at the time of photographing, improving the accuracy of water quality estimation using data from the target image.

[0025] [Application Example 10] A treatment system comprising: a captured image acquisition unit that acquires captured image data from a digital camera, the digital camera being configured to generate captured image data representing a light source by capturing an image of the light source from above the water surface when the light source is on and placed underwater in a water treatment tank included in a wastewater treatment device; a target image acquisition unit that is configured to acquire data of a target image including a portion representing the light source and a peripheral portion surrounding the light source by using the captured image data; an estimation unit that is configured to acquire an estimation result of the water quality of the water in the water treatment tank by inputting the target image data into a trained machine learning model, the trained machine learning model being trained to use data of a plurality of training images representing images of a light source placed underwater captured with different water qualities, and output an estimation result indicating a water quality that corresponds to the data of the training images input to the machine learning model; and a display processing unit that performs processing to display the estimation result on a display device of a terminal device.

[0026] According to this configuration, the user can recognize the estimated water quality result by observing the display device of the terminal device.

[0027] Application Example 11 A method for training a machine learning model, comprising: performing a plurality of set processes, the set processes including photographing a light source that is turned on and placed underwater in a first water treatment tank included in a training wastewater treatment device from above the water surface and measuring the water quality of the water in the first water treatment tank, the plurality of set processes including specific plurality of set processes that are performed at a plurality of times when the water quality of the water in the first water treatment tank differs from one another; obtaining a plurality of training image data including a portion representing the light source and a peripheral portion surrounding the light source by using data of a plurality of photographed images of the light source obtained by the plurality of set processes; and adjusting parameters of the machine learning model, using a plurality of sets of data of the plurality of training images and the water quality measurement results associated with the data of each training image, to output an estimation result indicating a water quality associated with the data of the training image input to the machine learning model.

[0028] According to this configuration, the parameters of the machine learning model are appropriately adjusted.

[0029] [Application Example 12] A wastewater treatment device comprising: one or more water treatment tanks including a specific water treatment tank; and a light source detachably attached to the wastewater treatment device, wherein the light source is positioned lower than a predetermined standard water level in the specific water treatment tank and is visible when observing the specific water treatment tank from above looking downward, and the light source includes an upward-facing light-emitting diode.

[0030] According to this configuration, it is possible to easily obtain a photographed image that represents the light source and is suitable for estimating water quality.

[0031] [Application Example 13] An image acquisition system comprising: a captured image acquisition unit that acquires data of a captured image, which is an image captured from above the water surface of a light source that is on and placed underwater in a water treatment tank included in a wastewater treatment device; and a target image acquisition unit that acquires data of a target image that includes a portion that represents the light source and a surrounding area surrounding the light source by using a specific portion of the captured image that is a part that includes a portion that represents the light source and a surrounding area that surrounds the light source.

[0032] The intensity and spread of light from a lit underwater light source can appear differently depending on the water quality. The specific portion can represent the intensity and spread of light from the light source. By using the specific portion, data of the target image suitable for estimating water quality can be obtained.

[0033] Application Example 14 In the image acquisition system according to Application Example 13, the target image acquisition unit selects, as the specific portion, a portion of the captured image that includes an area where a predetermined number or more of pixels that are brighter than a threshold are consecutive.

[0034] This configuration allows the specific portion to be appropriately selected.

[0035] The technology disclosed in this specification can be realized in various forms, such as a data processing method, a data processing device, a data processing system, a computer program for realizing the functions of those methods, devices, or systems, a recording medium on which that computer program is recorded (e.g., a non-transitory recording medium), etc.

[0036] 1 is a diagram showing a system according to an embodiment. (A) and (B) are diagrams showing an example of a wastewater treatment device 800. (A)-(D) are diagrams showing an example of a light source module 900. (A) is a block diagram showing an example of a machine learning model ML. (B) is a flowchart showing an example of a training process. (A)-(D) are diagrams showing an overview of a captured image. (B) is a flowchart showing an example of a pre-processing process. (C) is a flowchart showing an example of a process for extracting a specific portion. (A)-(F) are diagrams showing examples of images processed by the extraction process. (C) is a flowchart showing an example of a recording process. (D) is a diagram showing an example of data processed in the recording process. (E) is a diagram showing an example of a database DT. (F) is a sequence diagram showing an example of a display process. (A) and (B) are diagrams showing an example of a management page. (E) is a flowchart showing a second embodiment of the recording process. (E) is a flowchart showing a third embodiment of the recording process. (A) is a diagram showing an example of an additional light source. (B) is a diagram showing an example of a captured image. (C) is a diagram showing an example of a database. (E) is a diagram showing an example of a management page.

[0037] A. First Embodiment: A1. System Configuration: Fig. 1 is a diagram showing a system according to one embodiment. This system 1000 includes a first processing device 100A, a second processing device 100B, a terminal device 200, and a management server 300.

[0038] The first treatment device 100A is disposed at the installation site of the first wastewater treatment device 800A. A digital camera CAMa and a light source 910a are attached to the first wastewater treatment device 800A. The first treatment device 100A controls the digital camera CAMa and the light source 910a. As will be described in detail later, the light source 910a is disposed underwater in a water treatment tank included in the first wastewater treatment device 800A. The digital camera CAMa photographs the light source 910a when it is turned on from above the water surface. The first treatment device 100A transmits image data of the light source 910a photographed by the digital camera CAMa to the management server 300.

[0039] The second treatment device 100B is disposed at the installation site of the second wastewater treatment device 800B. A digital camera CAMb and a light source 910b are attached to the second wastewater treatment device 800B. The second treatment device 100B controls the digital camera CAMb and the light source 910b. Like the light source 910a and digital camera CAMa, the light source 910b is disposed underwater, and the digital camera CAMb photographs the light source 910b while it is turned on from above the water surface. The second treatment device 100B transmits image data of the light source 910b photographed by the digital camera CAMb to the management server 300.

[0040] In this embodiment, the wastewater treatment devices 800A and 800B are so-called distributed wastewater treatment devices. The wastewater treatment devices 800A and 800B are installed at the site where wastewater is generated (for example, near a facility such as a home or a store) and treat the wastewater. The second wastewater treatment device 800B is installed at a location different from the installation location of the first wastewater treatment device 800A.

[0041] The management server 300 estimates the water quality of the wastewater treatment devices 800A, 800B using images acquired through the treatment devices 100A, 100B, and records the estimation results. The terminal device 200 communicates with the management server 300 to display the estimation results.

[0042] The treatment devices 100A, 100B and the management server 300 are provided by a service provider that provides management services for the wastewater treatment devices 800A, 800B. The service provider may be the manufacturer of the wastewater treatment devices 800A, 800B or the management provider of the wastewater treatment devices 800A, 800B.

[0043] In this embodiment, the terminal device 200 and the management server 300 are connected to the Internet IT. The terminal device 200 can communicate with the management server 300 via the Internet IT. The processing devices 100A and 100B and the management server 300 are connected to a cellular network CN. The processing devices 100A and 100B can communicate with the management server 300 via the cellular network CN.

[0044] In this embodiment, the treatment devices 100A and 100B have the same hardware configuration. Hereinafter, when the treatment devices 100A and 100B are not distinguished from each other, the alphabetical letters at the end of the reference numerals will be omitted and the devices will be simply referred to as treatment device 100. Similarly, the wastewater treatment devices 800A and 800B have the same hardware configuration, the digital cameras CAMa and CAMb have the same hardware configuration, and the light sources 910a and 910b have the same hardware configuration. The arrangement of the digital camera CAMb and the light source 910b in the second wastewater treatment device 800B is the same as the arrangement of the digital camera CAMa and the light source 910a in the first wastewater treatment device 800A. Hereinafter, when the digital cameras CAMa and CAMb are not distinguished from each other, they will be referred to as the digital camera CAM. When the light sources 910a and 910b are not distinguished from each other, they will be referred to as the light source 910.

[0045] The processing device 100 includes a processor 110, a storage device 115, a display unit 140, an operation unit 150, a device interface 160, and a communication interface 180. These elements are connected to one another via a bus (not shown). The storage device 115 includes a volatile storage device 120 and a non-volatile storage device 130. The processing device 100 may be, for example, a small personal computer. Alternatively, the processing device 100 may be a so-called IoT gateway.

[0046] The processor 110 is a device configured to process data, such as a central processing unit (CPU) or a system on a chip (SoC). The volatile storage device 120 is, for example, a dynamic random access memory (DRAM), and the non-volatile storage device 130 is, for example, a flash memory.

[0047] The display unit 140 is a device configured to visually display information, such as a light-emitting diode (LED), a seven-segment display, a liquid crystal display, an organic EL display, etc. The operation unit 150 is a device configured to receive operations by a user, such as a button, a lever, or a touch panel overlaid on the display unit 140.

[0048] The device interface 160 is an interface for connecting external devices such as a digital camera CAM and a light source 910 (for example, a Universal Serial Bus (USB), a General Purpose Interface Bus (GPIB), etc.).

[0049] The communication interface 180 is an interface for communicating with other devices. In this embodiment, the communication interface 180 has an interface for a cellular network (e.g., Universal Mobile Telecommunications System (UMTS), Long Term Evolution (LTE), IMT-Advanced, IMT-2020, etc.). The communication interface 180 is connected to the cellular network CN.

[0050] The non-volatile storage device 130 stores data of a program P11. In accordance with the program P11, the processor 110 controls the digital camera CAM and the light source 910 connected to the device interface 160, and periodically (for example, once a month) transmits image data representing the light source 910 photographed by the digital camera CAM to the management server 300 (details will be described later).

[0051] The nonvolatile memory device 130 of the first treatment device 100A stores data of a first device identifier IDa. The nonvolatile memory device 130 of the second treatment device 100B stores data of a second device identifier IDb. The device identifiers IDa and IDb are identifiers that identify the individual treatment devices 100A and 100B (i.e., the individual wastewater treatment devices 800A and 800B).

[0052] The data for the program P11 and the device identifier are stored in the non-volatile storage device 130 by the service provider when the processing device 100 is shipped. The processing device 100 may obtain the data for the program P11 from a server (not shown) via a network, or from a portable storage device (e.g., a USB flash drive) (not shown) connected to the processing device 100. The same applies to the data for the device identifier.

[0053] The management server 300 has a processor 310, a storage device 315, and a communication interface 380. These elements are connected to each other via a bus (not shown). The storage device 315 includes a volatile storage device 320 and a non-volatile storage device 330.

[0054] The processor 310 is a device configured to process data, such as a CPU or an SoC. The volatile storage device 320 is, for example, a DRAM, and the non-volatile storage device 330 is, for example, a flash memory.

[0055] The communication interface 380 is an interface for communicating with other devices. In this embodiment, the communication interface 380 has an interface for an IEEE802.3 wired LAN. The communication interface 380 is connected to the Internet IT via a relay device (e.g., a router) not shown. The communication interface 380 is also connected to a cellular network CN via a relay device (e.g., a router) not shown.

[0056] The non-volatile storage device 330 stores data for the programs P31, P32, and P33, the database DT, and the machine learning model ML. The first program P31 is a program for training the machine learning model ML. The second program P32 is a program for recording water quality estimation results in the database DT. The third program P33 is a program for providing information from the database DT to an external device such as the terminal device 200. In this embodiment, the third program P33 causes the processor 310 to realize the function of a web server that provides web pages. The programs P31, P32, and P33 are uploaded to the management server 300 by the service provider. Details of the programs P31, P32, and P33, the database DT, and the machine learning model ML will be described later.

[0057] The terminal device 200 is a terminal device of the manager of the wastewater treatment devices 800A and 800B, and is, for example, a personal computer, a tablet computer, or a smartphone.

[0058] The terminal device 200 has a processor 210, a storage device 215, a display unit 240, an operation unit 250, and a communication interface 280. These elements are connected to each other via a bus (not shown). The storage device 215 includes a volatile storage device 220 and a non-volatile storage device 230.

[0059] The processor 210 is a device configured to process data, such as a CPU or an SoC. The volatile storage device 220 is, for example, a DRAM, and the non-volatile storage device 230 is, for example, a flash memory.

[0060] The display unit 240 is a device configured to display images, such as a liquid crystal display or an organic EL display. The operation unit 250 is a device configured to be operated by a user, such as a button, a lever, or a touch panel overlaid on the display unit 240. The display unit 240 and the operation unit 250 may form a so-called touch screen. The user can input various requests and instructions to the terminal device 200 by operating the operation unit 250. The display unit 240 may display operation elements (e.g., buttons, sliders, etc.), and the displayed elements may be operated through operation of the operation unit 250.

[0061] The communication interface 280 is an interface for communicating with other devices (for example, it includes one or more interfaces from among an IEEE802.3 wired LAN, an IEEE802.11 wireless LAN, and a cellular network interface). In this embodiment, the communication interface 280 is connected to the Internet (IT) via a network (not shown) (for example, a local area network, a cellular network, etc.).

[0062] The non-volatile storage device 230 stores data for a program P21. In this embodiment, the program P21 is a web browser program. The administrator can access web pages provided by the management server 300 through the web browser.

[0063] A2. Configuration of wastewater treatment device: Figures 2(A) and 2(B) are diagrams showing an example of a wastewater treatment device 800. Figure 2(A) shows the schematic configuration of the wastewater treatment device 800 as seen from the side, and Figure 2(B) shows the schematic configuration of the wastewater treatment device 800 as seen from below. In each figure, the first direction X indicates the longitudinal direction (horizontal direction) of the wastewater treatment device 800. The second direction Y is the horizontal direction perpendicular to the first direction X. The third direction Z indicates the vertically upward direction. Hereinafter, the third direction Z will also be referred to as the +Z direction, and the direction opposite to the third direction Z will also be referred to as the -Z direction. Similarly, for other directions, the same direction and opposite direction are expressed by a positive sign or a negative sign before the sign.

[0064] The wastewater treatment device 800 has a body 801 that forms the outer surface of the wastewater treatment device 800. The body 801 is provided with an inlet 804 and an outlet 805. The body 801 houses an impurity removal tank 810, an anaerobic filter bed tank 820, a contact filter bed tank 830, a treated water tank 840, and a disinfection tank 850. These water treatment tanks 810, 820, 830, 840, and 850 are arranged in this order in the +X direction. When viewed downward as shown in FIG. 2(B), the contact filter bed tank 830 is a U-shaped water treatment tank having portions located in the +Y direction, +X direction, and -Y direction of the treated water tank 840. The disinfection tank 850 is arranged above the treated water tank 840.

[0065] Wastewater that flows into wastewater treatment device 800 through inlet 804 is treated sequentially in water treatment tanks 810, 820, 830, 840, and 850, and then flows out of wastewater treatment device 800 through outlet 805. Wastewater treatment device 800 performs a purification process using an oxygen-containing gas (here, air) supplied by a blower (not shown).

[0066] Wastewater from the inlet 804 flows into the impurity removal tank 810. The impurity removal tank 810 has an inlet baffle 812 that separates impurities from the water. After the impurities have been separated, the water flows into the anaerobic filter bed tank 820 through an opening 814 in the partition wall 802.

[0067] Filter media 822 for anaerobic microorganisms to adhere to is provided within anaerobic filter bed tank 820. Water flowing into anaerobic filter bed tank 820 is anaerobically treated by the anaerobic microorganisms. The treated water in anaerobic filter bed tank 820 flows into contact filter bed tank 830 through opening 824 provided in partition wall 803.

[0068] The contact filter bed tank 830 has an air diffuser 834, an aerobic filter medium 833 placed on the air diffuser 834, and a contact material 832 placed on the aerobic filter medium 833. Air is supplied to the air diffuser 834 by a blower (not shown). The air diffuser 834 supplies oxygen-containing bubbles to the water in the contact filter bed tank 830. Aerobic microorganisms adhere to the contact material 832 and the aerobic filter medium 833. The water in the contact filter bed tank 830 is aerobically treated by the aerobic microorganisms. The water treated in the contact filter bed tank 830 flows into the treated water tank 840 through an opening 836 at the bottom of the contact filter bed tank 830.

[0069] Water temporarily remains in the treatment water tank 840. A circulation air lift pump 880 is provided in the treatment water tank 840. An intake port 882 of the circulation air lift pump 880 is located at the bottom of the treatment water tank 840. The circulation air lift pump 880 uses air from a blower (not shown) to transfer water containing solids that have settled at the bottom of the treatment water tank 840 to the impurity removal tank 810.

[0070] The disinfection tank 850 has a discharge air lift pump 870. An inlet 872 (FIG. 2A) of the discharge air lift pump 870 is located at the same height as the standard water level LWL in the treatment water tank 840. The discharge air lift pump 870 uses air from a blower (not shown) to gradually transfer water near the water surface WL in the treatment water tank 840 (i.e., water from which solids have been separated) to the disinfection tank 850. An outlet 874 (FIG. 2B) of the discharge air lift pump 870 is located in the upstream portion of the disinfection tank 850. When a large amount of water temporarily flows into the wastewater treatment device 800 (e.g., during peak inflow), the water levels in the water treatment tanks 810, 820, 830, and 840 upstream of the discharge air lift pump 870 may temporarily rise above the standard water level LWL.

[0071] The disinfection tank 850 has a chemical cylinder 854 filled with a disinfectant (e.g., a solid chlorine agent) and a storage chamber 858 that temporarily stores disinfected water. Water transferred to the disinfection tank 850 by the discharge air lift pump 870 is disinfected by contact with the disinfectant. The disinfected water flows into the storage chamber 858. The water temporarily stagnates in the storage chamber 858. The water in the storage chamber 858 flows out of the wastewater treatment device 800 through the outlet 805.

[0072] The body 801 further has manholes 891, 892, and 893. The wastewater treatment device 800 has covers 891a, 892a, and 893a that cover the manholes 891, 892, and 893. The first manhole 891 is located above the impurity removal tank 810, the second manhole 892 is located above the anaerobic filter bed tank 820, and the third manhole 893 is located above the water treatment tanks 830, 840, and 850.

[0073] A light source module 900 is disposed in the disinfection tank 850. A digital camera CAM is attached to the underside of a lid 893a that covers the third manhole 893.

[0074] A3. Configuration of the light source module 900: Figures 3(A) to 3(D) are diagrams showing examples of the light source module 900. Figure 3(A) shows the light source module 900 as seen from the side. Figure 3(B) shows the light source module 900 as seen facing downward (i.e., in the -Z direction). The light source module 900 has a light source 910, a support 920 that supports the light source 910, and a cap 930.

[0075] The light source 910 has a pipe 914 that extends approximately horizontally, a light emitting element 912 arranged in the pipe 914, and a connecting portion 916. The connecting portion 916 connects the light emitting element 912 to the pipe 914, thereby fixing the light emitting element 912 to the pipe 914. The support portion 920 has a pipe 921 that extends in the −Z direction and a corner pipe 923 connected to the lower end of the pipe 921. The corner pipe 923 connects the pipe 914 of the light source 910 to the pipe 921 of the support portion 920. The opposite end of the pipe 914 of the light source 910 is closed by a cap 930.

[0076] In this embodiment, the light-emitting element 912 is a light-emitting diode. The light-emitting diode has a lens. The diameter of the lens may be, for example, 3 mm. The color of light emitted by the light-emitting diode may be, for example, white. The pipe 914 of the light source 910 is translucent, allowing light from the light-emitting element 912 to pass through the pipe 914. The light-emitting element 912 faces in the +Z direction. In other words, the light source 910 is configured to emit light that includes a component directed upward (i.e., toward the water surface). The pipe 914 and connection portion 916 of the light source 910, the pipes 921 and 923 of the support portion 920, and the cap 930 are formed of, for example, resin (polyvinyl chloride, acrylic resin, etc.).

[0077] Figures 3(C) and 3(D) show the light source module 900 attached to the wastewater treatment device 800. Figure 3(C) shows a part of the body 801, the disinfection tank 850, and the light source module 900 as seen from the side. Figure 3(D) shows the disinfection tank 850 and the light source module 900 as seen from below.

[0078] As shown in FIG. 2A, a pipe holder 809 is fixed to a portion of the body 801 near the third manhole 893. As shown in FIG. 3D, the pipe holder 809 is a C-shaped holder. The pipe 921 of the support part 920 is detachably attached to the pipe holder 809. As shown in FIG. 3C, the light source 910 is disposed at a position lower than the standard water level WL5 in the storage chamber 858 of the disinfection tank 850. In this manner, the light source 910 is disposed under the water in the disinfection tank 850. The standard water level WL5 indicates the standard water level in the disinfection tank 850 (here, the storage chamber 858). The standard water level WL5 indicates a stable water level when no water flows into the wastewater treatment device 800. The height of the standard water level WL5 is, for example, the same as the height of the lower end of the outlet 805.

[0079] A power cable 9c for supplying power is connected to the light-emitting element 912 of the light source 910. The power cable 9c passes through the inside of the support portion 920 and reaches the outside of the support portion 920 from the upper end 921u of the pipe 921. The power cable 9c is then connected to the device interface 160 of the treatment device 100. A control cable Cc is also connected to the digital camera CAM. The control cable Cc is connected to the device interface 160 of the treatment device 100. Although not shown, the treatment device 100, like the digital camera CAM, is disposed inside the wastewater treatment device 800. For example, the treatment device 100 is attached to the underside of a lid 893a that covers a third manhole 893 ( FIG. 2(A) ). Alternatively, the treatment device 100 may be disposed outside the wastewater treatment device 800. In this case, a through-hole through which the cables 9c and Cc pass is provided in the wastewater treatment device 800. The through-hole may be formed in the lid 893a or the body 801. Alternatively, the cables 9c and Cc may be passed through the gap between the lid 893a and the body 801.

[0080] 3(D), the light source 910 is disposed in a position where it can be seen when observing the disinfection tank 850 from above the disinfection tank 850 looking downward. The digital camera CAM can photograph the underwater light source 910 from above the water surface WS5 of the disinfection tank 850.

[0081] A4. Training Process: FIG. 4 is a block diagram showing an example of the machine learning model ML. In this embodiment, the machine learning model ML is a classifier that uses a convolutional neural network. The machine learning model ML classifies an image IMi representing a photographed light source 910 into one of multiple classes CL representing different ranges of water quality. Hereinafter, it is assumed that transparency T is estimated as water quality. The multiple classes CL may include, for example, ten classes CL1-CL10 corresponding to ten ranges obtained by equally dividing the range of transparency T from zero to 100.

[0082] The machine learning model ML includes a pre-processing unit MLc and a post-processing unit MLf following the pre-processing unit MLc. In this embodiment, each of the processing units MLc and MLf includes multiple layers connected in series. The pre-processing unit MLc includes one or more convolutional layers and one or more pooling layers (e.g., max pooling or average pooling). The pre-processing unit MLc may include, for example, a set of a convolutional layer followed by a pooling layer. The total number of sets may be one or may be any number greater than one (e.g., three). The convolutional layer performs so-called convolution processing on input data (an image or a feature map output from a previous layer) using a filter (also called a kernel) associated with the convolutional layer. This generates a feature map representing local features. The pooling layer performs so-called downsampling on the feature map output from the previous layer. This generates a feature map with reduced spatial resolution.

[0083] The post-processing unit MLf has one or more fully connected layers. The post-processing unit MLf generates certainty data OD by linearly transforming the feature map output from the pre-processing unit MLc. The certainty data OD represents the certainty CS of each of the multiple classes CL. In the example of FIG. 4, the certainty data OD represents ten certainty factors CS1-CS10 of the ten classes CL1-CL10. Of the multiple certainty factors CS of the multiple classes CL, the class CL associated with the highest certainty factor CS represents the estimated water quality of the image IMi.

[0084] 5 is a flowchart showing an example of a training process for the machine learning model ML. In steps S110 to S135, an operator prepares training data. The training data represents a training set that is a set of a photographed image representing the light source 910 and a measured transparency. The training data represents a plurality of training sets that show different transparency levels.

[0085] In this example, an operator prepares training data using an experimental wastewater treatment device 800 (FIGS. 2A and 2B). The operator introduces wastewater (e.g., domestic wastewater or artificial raw water simulating domestic wastewater) into the wastewater treatment device 800 and proceeds with water treatment using the wastewater treatment device 800. A digital camera CAM and a light source module 900 are attached to the wastewater treatment device 800.

[0086] In S110, the worker turns on the light source 910 and causes the digital camera CAM to capture the light source 910 while it is on. After capturing the image, the light source 910 is turned off. FIGS. 6A-6D are diagrams showing an overview of the captured images. Captured images IM11-IM14 are rectangular images with two sides parallel to a first direction D1 and two sides parallel to a second direction D2 perpendicular to the first direction D1. Although not shown, the data of captured images IM11-IM14 is bitmap data representing the color values ​​of each of a plurality of pixels arranged in a matrix along the first direction D1 and the second direction D2. The color values ​​are represented by the respective gradation values ​​(e.g., values ​​greater than or equal to zero and less than or equal to 255) of red (R), green (G), and blue (B). The width W indicates the number of pixels in the first direction D1, and the height H indicates the number of pixels in the second direction D2.

[0087] In this embodiment, the orientation of the digital camera CAM is adjusted in advance so that the light source 910 is positioned inside the edges of the captured images IM11-IM14. The digital camera CAM's shooting settings (e.g., shutter speed, ISO sensitivity, aperture, angle of view, etc.) are set to predetermined specific settings. The specific settings are experimentally determined in advance so that images suitable for estimating water quality using the machine learning model ML are obtained. The light-emitting element 912 of the light source 910 is experimentally selected in advance to have a brightness suitable for estimating water quality using the machine learning model ML. Furthermore, in this embodiment, the shooting range of the digital camera CAM includes the disinfection tank 850, as well as a portion of the contact filter bed tank 830 and a portion of the treated water tank 840. However, portions far from the light source 910 may be excluded from the shooting range. Photography is performed with all manholes 891-893 of the wastewater treatment device 800 closed with their lids 891a-893a. At the time of photographing, there is no light source other than the light source 910 that is turned on inside the wastewater treatment device 800. In other words, photographing is performed in a state where no light other than the light from the light source 910 is irradiated into the interior of the wastewater treatment device 800. Therefore, the interior of the wastewater treatment device 800 is dark. In the photographed images IM11-IM14, the portion representing the light source 910 is displayed brightly. Other portions (e.g., portions representing the walls of the water treatment tanks 830, 840, 850, etc.) are displayed in a dark color (e.g., black).

[0088] The appearance of the light source 910 in the captured image changes depending on the water quality. The captured images IM11-IM14 in Figures 6(A)-6(D) are associated with transparency levels T1-T4, respectively. The order of transparency level T is T1 to T4.

[0089] When the transparency T is relatively high, the portion representing the light source 910 in the captured image is brighter than when the transparency T is relatively low. For example, the portion representing the light source 910 in the captured image IM11 is brighter than the portion representing the light source 910 in the captured image IM14. Furthermore, the brighter the portion representing the light source 910, the brighter the area surrounding the light source 910 may be. Furthermore, when the transparency T is relatively low, light may be more likely to scatter in water than when the transparency T is relatively high. In other words, when the transparency T is relatively low, the distribution range of light centered on the light source 910 may be wider than when the transparency T is relatively high. When the distribution range of light is wide, the area surrounding the light source 910 is brighter (however, the surrounding area is usually darker than the area representing the light source 910).

[0090] In this way, the brightness of the area representing the light source 910, the brightness of the area surrounding the light source 910, and the distribution of light centered on the light source 910 change depending on the water quality. The captured image representing the light source 910 and its surrounding area represents various characteristics that change depending on the water quality. The machine learning model ML can estimate the water quality using the image representing the various characteristics that change depending on the water quality.

[0091] In S112 (FIG. 5), the worker acquires the captured image data from the digital camera CAM. For example, the worker connects the control cable Cc connected to the digital camera CAM to a work terminal (not shown, for example, a personal computer) to store the captured image data in the terminal device.

[0092] In S115, the operator samples water from the disinfection tank 850 (here, the water in the reservoir 858) and measures the transparency T of the sampled water. The measurement of the transparency T is performed in accordance with the method defined by the Japanese Industrial Standard "JIS K 0102." Alternatively, the transparency T may be measured using a sensor. For example, a digital transparency meter may be used that includes a sensor that is placed in water and outputs a signal correlated with the transparency, and a conversion device that uses the signal obtained from the sensor to generate digital data representing the transparency according to a conversion formula suitable for the above standard.

[0093] In S120, the operator uses a computer to perform preprocessing of the captured images to generate training image data. Various computers may be used to generate the training image data and to train the machine learning model ML, which will be described later. In this embodiment, the management server 300 ( FIG. 1 ) is used. The first program P31 is a program for the training process.

[0094] 7 is a flowchart showing an example of pre-processing. An operator uses a work terminal (not shown) to upload data of the captured image and data representing the measured transparency T to the management server 300. Then, the administrator uses the work terminal to input an instruction to start the training data generation process to the management server 300. The processor 310 of the management server 300 executes the pre-processing in accordance with the first program P31.

[0095] In S210, the processor 310 extracts a specific portion representing the light source 910 and its surrounding area from the captured image (this type of processing is also called cropping or trimming). FIG. 8 is a flowchart showing an example of the specific portion extraction processing. FIGS. 9(A) to 9(F) are diagrams showing examples of images processed by the extraction processing. FIG. 9(A) shows the image to be processed. Here, it is assumed that the captured image IM11 described in FIG. 6(A) is processed.

[0096] In S310 (FIG. 8), the processor 310 performs grayscale conversion on the original image (here, the captured image IM11). As the correspondence between the color gradation values ​​and the grayscale gradation values, a known relationship can be adopted (for example, the correspondence between the RGB values ​​in the RGB color space and the luminance value Y in the YCbCr color space).

[0097] In S315, the processor 310 binarizes the grayscale image. The binarization method may be any of various methods for distinguishing between bright and dark regions. In this embodiment, so-called Otsu's binarization is employed. Otsu's binarization determines a threshold value by analyzing the image. Alternatively, a predetermined threshold value may be used for binarization. FIG. 9B shows an example of a binary image obtained by grayscale conversion and binarization of the captured image IM11. The binary image IM11b shows a bright region Ab and a dark region Ad. In the figure, the bright region Ab is not hatched, and the dark region Ad is hatched. The bright region Ab includes a portion representing the light source 910.

[0098] In S320 (FIG. 8), the processor 310 initializes the parameters. In this embodiment, the parameters are initialized as follows: Bright pixel counter Nb=0 Bright pixel flag FL=F (false) First position range RG1=OF1 to W-OF1 Second position range RG2=OF2 to H-OF2 The processor 310 uses these parameters Nb, FL, RG1, and RG2 to search for neighboring pixels, which are pixels near the light source 910, from the binary image IM11b (FIG. 9B).

[0099] FIG. 9C shows the relationship between the binary image IM11b, the position ranges RG1 and RG2, and the offsets OF1 and OF2. For ease of explanation, the hatching indicating the dark area Ad has been omitted from the figure. In this embodiment, the light source 910 is located inside the edge of the binary image IM11b (i.e., the captured image IM11). The processor 310 searches for nearby pixels in the target area SA of the binary image IM11b, excluding the edge. The first position range RG1 indicates the range of pixel positions in the first direction D1 of the target area SA, and the second position range RG2 indicates the range of pixel positions in the second direction D2 of the target area SA. Here, the origin of the pixel positions is the position of pixel O in the upper left corner of the binary image IM11b. The first position range RG1 is the remainder of the range of pixel positions in the first direction D1 of the binary image IM11b, excluding the offsets OF1 at both ends. The second position range RG2 is the range of pixel positions in the binary image IM11b in the second direction D2, excluding the offsets OF2 at both ends. The offsets OF1 and OF2 are experimentally determined in advance so that the light source 910 is included in the target area SA.

[0100] After S320 (FIG. 8), the processor 310 executes a first loop process L1 between start L1S and end L1E. The first loop process L1 is repeated while changing the pixel position P1 of the target pixel Pt in the first direction D1 by one within a first position range RG1 from OF1 to W-OF1. The first loop process L1 includes a second loop process L2 between start L2S and end L2E. The second loop process L2 is repeated while changing the pixel position P2 of the target pixel Pt in the second direction D2 by one within a second position range RG2 from OF2 to H-OF2. The second loop process L2 includes S235-S360.

[0101] The scan line SL in FIG. 9C represents the movement of the pixel of interest Pt through the loop processes L1 and L2. As shown by one scan line SL, the second loop process L2 is repeated while moving the pixel of interest Pt from the top end (the end in the -D2 direction) of the target area SA to the bottom end (the end in the +D2 direction). The first loop process L1 is repeated while moving the pixel of interest Pt (and thus the scan line SL) from the left end (the end in the -D1 direction) of the target area SA to the right end (the end in the +D1 direction). While moving the pixel of interest Pt in this manner, the processor 310 executes steps S235-S360 ( FIG. 8 ) for the pixel of interest Pt. Below, an overview of each of steps S235-S360 will be described, followed by a description of specific situations with reference to the figures.

[0102] In S325, the processor 310 determines whether the pixel of interest Pt is a bright pixel, which is a pixel included in the bright region Ab ( FIG. 9C ). If the pixel of interest Pt is a bright pixel (S325: Yes), the processor 310 increments the bright pixel counter Nb by 1 in S330 and proceeds to S335. If the pixel of interest Pt is not a bright pixel (S325: No), the processor 310 skips S330 and proceeds to S335.

[0103] In S335, the processor 310 determines whether the bright pixel counter Nb is 1. If Nb=1 (S335: Yes), the processor 310 sets the bright pixel flag FL to TR (true) in S340 and proceeds to S345. If the bright pixel counter Nb is not 1 (S335: No), the processor 310 skips S340 and proceeds to S345.

[0104] In S345, the processor 310 determines whether the following conditions are met: FL = TR, and the pixel of interest Pt is a dark pixel that is a pixel included in the dark region Ad. If the conditions are met (S345: Yes), the processor 310 initializes the bright pixel counter Nb to zero and the bright pixel flag FL to F in S350. The processor 310 then proceeds to S360. If the conditions are not met (S345: No), the processor 310 skips S350 and proceeds to S360.

[0105] In S360, the processor 310 determines whether the bright pixel counter Nb is equal to a predetermined threshold NbTh, which is experimentally determined in advance so that the bright pixel counter Nb is equal to or greater than the threshold NbTh when the target pixel Pt is located within the bright region Ab.

[0106] If the bright pixel counter Nb is different from the threshold value NbTh (S360: No), the processor 310 moves the pixel position P2 of the target pixel Pt to the next position and repeats the second loop process L2. If the pixel position P2 is at the bottom edge of the target area SA, the processor 310 returns the pixel position P2 to the top edge of the target area SA and moves the pixel position P1 to the next position. The processor 310 then executes S235-S360 for the new target pixel Pt. As will be described later, if Nb = NbTh (S360: Yes), the processor 310 ends the loop processes L1 and L2 and executes S365. Therefore, if the bright pixel counter Nb is different from the threshold value NbTh (S360: No), the bright pixel counter Nb is less than the threshold value NbTh.

[0107] 9C, when the pixel of interest Pt moves within the dark region Ad, the determination result in S325 is No, and the bright pixel counter Nb is maintained at zero. Therefore, the determination result in S335 is No, and the bright pixel flag FL is maintained at F. Therefore, the determination result in S345 is No, and the determination result in S360 is No. Then, such loop processing L1 and L2 is repeated until the pixel of interest Pt becomes a bright pixel.

[0108] The scan line SLi in FIG. 9C is a scan line that passes through the bright region Ab. The first pixel Pt1 indicates a bright pixel when the pixel of interest Pt switches from a dark pixel to a bright pixel on the scan line SLi. If the first pixel Pt1 is the pixel of interest Pt, the determination result in S325 is Yes, and the bright pixel counter Nb is updated from zero to one in S330. Therefore, the determination result in S335 is Yes, and the bright pixel flag FL is set to TR in S340. If the pixel of interest Pt is located within the bright region Ab, the determination result in S345 is No. If the bright pixel counter Nb is less than the threshold value NbTh (S360: No), as described above, the pixel of interest Pt is moved and loop processes L1 and L2 are repeated. When the pixel of interest Pt moves within the bright region Ab, the bright pixel counter Nb is incremented by one.

[0109] The second pixel Pt2 in FIG. 9C represents a dark pixel when the pixel of interest Pt switches from a bright pixel to a dark pixel on the scan line SLi. Here, the pixel of interest Pt moves to the second pixel Pt2 while the bright pixel counter Nb remains less than the threshold value NbTh. This movement of the pixel of interest Pt is likely to occur, for example, when the scan line SLi passes through the edge of the bright region Ab, as shown in FIG. 9C. Because the pixel of interest Pt is a dark pixel, the determination result in S325 is No, and the increment of the bright pixel counter Nb is stopped. Because the determination result in S345 is Yes, the bright pixel counter Nb and the bright pixel flag FL are initialized in S350. Because the determination result in S360 is No, the processor 310 repeats loop processing L1 and L2 while moving the pixel of interest Pt.

[0110] The scan line SLj in FIG. 9D is the scan line used in processing after the scan line SLi (FIG. 9C) is processed. The third pixel Pt3 indicates a bright pixel when the pixel of interest Pt switches from a dark pixel to a bright pixel on the scan line SLj. When the third pixel Pt3 is the pixel of interest Pt, processing proceeds in the same manner as when the first pixel Pt1 in FIG. 9C is the first pixel Pt1. The bright pixel counter Nb is updated from zero to one (S325: Yes, S330). The bright pixel flag FL is set to TR (S335: Yes, S340). When the pixel of interest Pt moves within the bright region Ab, the bright pixel counter Nb is incremented by one.

[0111] When the fourth pixel Pt4 in the bright region Ab (FIG. 9D) is the pixel of interest Pt, the bright pixel counter Nb updated in S330 is assumed to be equal to the threshold value NbTh. Such a large bright pixel counter Nb can occur when the scan line SLj passes near the center of the bright region Ab, as shown in FIG. 9D. Because the bright pixel counter Nb is equal to the threshold value NbTh, the determination result in S360 (FIG. 8) is Yes. In this case, the pixel of interest Pt is a neighboring pixel, that is, a pixel near the light source 910. In S365, the processor 310 determines a partial region based on the pixel of interest Pt and extracts a partial image, which is an image of the partial region, from the captured image IM11.

[0112] FIG. 9E shows an example of a partial region when the fourth pixel Pt4 is the pixel of interest Pt. In this embodiment, the processor 310 determines the partial region Ac as a rectangular region of a predetermined size that is located at a predetermined position relative to the pixel of interest Pt. The predetermined position of the partial region Ac relative to the pixel of interest Pt and the predetermined size of the partial region Ac are experimentally determined in advance so that the partial region Ac includes both the portion of the captured image representing the light source 910 and its surrounding portion (i.e., the portion where light from the light source 910 may be distributed). The region Ar in the figure is a region that is NbTh pixels or less away from the pixel of interest Pt. In this embodiment, the sizes of the partial region Ac in the first direction D1 and the second direction D2 are sufficiently large so that the partial region Ac includes the region Ar. The center position of the partial region Ac may be the same as the pixel of interest Pt. Alternatively, the center position of the partial region Ac may be shifted in the first direction D1 and the second direction D2 from the pixel of interest Pt.

[0113] 9(F) shows an example of a specific portion extracted from the captured image IM11. The specific portion IMc is a portion corresponding to the partial area Ac of the captured image IM11. As shown in the figure, the specific portion IMc includes a portion representing the light source 910 and a portion 910s surrounding the light source 910. After S365 (FIG. 8), the processor 310 ends the processing of FIG. 8, i.e., S210 of FIG. 7.

[0114] If the loop processes L1 and L2 end without finding a neighboring pixel in the target area SA, the processor 310 determines in S370 that a specific portion has not been detected, and ends the process of Fig. 8. In this case, the processor 310 may end the process of Fig. 7, and therefore the process of Fig. 5.

[0115] After S210 (FIG. 7), in S220-S240, the processor 310 converts the data format of the specific portion IMc acquired in S210 into a format suitable for input to the machine learning model ML. In this embodiment, in S220, the processor 310 generates a grayscale image by performing grayscale conversion on the specific portion IMc. The grayscale conversion method is the same as the method in S310 (FIG. 8). In S230, the processor 310 generates a resolution-converted image by converting the resolution of the grayscale image to a resolution suitable for input to the machine learning model ML. A known method (e.g., interpolation) can be used as the resolution conversion method. In S240, the processor 310 normalizes the gradation values ​​of the resolution-converted image to generate a normalized image. In this embodiment, the gradation values ​​before normalization are expressed in 256 gradations from zero to 255. Processor 310 converts the gradation value into a real value greater than or equal to 0 and less than or equal to 1 by multiplying the gradation value by a constant (1 / 255). This completes the process in Fig. 7, i.e., the process in S120 in Fig. 5. Processor 310 uses the normalized image as a training image.

[0116] In S125, the processor 310 stores the training set data, which is a set of the data on the measurement results of water quality (here, transparency T) and the data on the training images, in the storage device 315 (e.g., the non-volatile storage device 330). The processor 310 then ends the training data generation process. Hereinafter, the entire process of photographing (S110) and measuring the water quality (S115) will be referred to as set process PR1.

[0117] In S130, the operator determines whether the generation of training data is complete. The completion condition may be various conditions indicating that training data capable of properly training the machine learning model ML has been generated. As described with reference to FIG. 4 , in this embodiment, the machine learning model ML classifies an image representing a light source 910 into one of multiple classes CL1-CL10. The transparency T of each training set is associated with one of the multiple classes CL1-CL10. The completion condition may be, for example, that for each of the multiple classes CL1-CL10, the total number of training sets associated with the class is equal to or greater than a predetermined number threshold. The number threshold is determined in advance experimentally so as to enable proper training of the machine learning model ML.

[0118] If the completion condition is not met (S130: No), in S135, the operator waits for a predetermined time to elapse (for example, at least five minutes but not more than one day). Then, after the predetermined time has elapsed, the processes of S110-S125 (including the set process PR1) are executed again. The water quality of each water treatment tank 810-850 of the wastewater treatment device 800 (FIGS. 2(A) and 2(B)) changes variously over time. Therefore, by repeating S110-S135, multiple training sets showing different transparency levels T are obtained. Hereinafter, the entire process of S110-S135 that is repeated is referred to as the specific process PR2. The specific process PR2 includes multiple executions of the set process PR1.

[0119] The operator may change the water quality of the wastewater treatment device 800 by adjusting the water quality (e.g., organic matter concentration) of the wastewater flowing into the wastewater treatment device 800. This can reduce the number of days required to prepare training data.

[0120] If the completion condition is met (S130: Yes), the administrator uses a work terminal (not shown) to input an instruction to start training processing for the machine learning model ML to the management server 300. The processor 310 of the management server 300 executes the training processing in accordance with the first program P31. The training processing is performed using multiple training sets of training images and transparency T. The processor 310 generates confidence data OD by executing calculations of the machine learning model ML using the training images. The processor 310 adjusts multiple calculation parameters used in the calculations of each layer of the machine learning model ML (e.g., multiple weights and multiple biases of multiple filters in a convolutional layer, multiple weights and multiple biases of a fully connected layer, etc.) so that the confidence CS of the class CL that includes the transparency T associated with the training images is maximized.

[0121] The training process includes steps S140 to S170. In step S140, the processor 310 initializes a plurality of calculation parameters of the machine learning model ML. For example, each calculation parameter is set to a random value.

[0122] At S145, the processor 310 acquires V (V is a predetermined integer equal to or greater than 1) target training sets that are part of the plurality of training sets stored in the non-volatile storage device 330. The V target training sets are selected from unused training sets among the plurality of training sets. Alternatively, the V training sets may be randomly selected from the plurality of training sets.

[0123] At S150, the processor 310 generates V pieces of confidence data OD by performing the operation of the machine learning model ML using V training images of the V target training sets.

[0124] In S155, the processor 310 calculates an error value indicating the difference between the certainty data OD and the target data associated with the training image used to generate the certainty data OD. The target data indicates the target value (i.e., the correct answer) of the certainty data OD. Specifically, the certainty CS of the class CL including the transparency T associated with the training image is 1, and the certainty CS of the other classes CL is zero. The error value is calculated based on a predetermined loss function (such an error value is also referred to as a loss value). In this embodiment, the so-called cross entropy is used as the loss function. The cross entropy can be used as an index value indicating the amount of deviation between two probability distributions (e.g., the certainty data OD and the target data). Note that the loss function may be various other functions (e.g., the sum of squares error).

[0125] In S155, the processor 310 calculates V error values. The processor 310 uses the V error values ​​to adjust a plurality of calculation parameters of the machine learning model ML. For example, the processor 310 adjusts the calculation parameters so that an index value calculated using the V error values ​​becomes smaller (the index value may be various values ​​that have a correlation with the magnitude of the difference, such as an average value, a maximum value, a median value, or a sum value). As an algorithm for adjusting the calculation parameters, for example, an algorithm using backpropagation and gradient descent may be adopted.

[0126] In S165, the processor 310 determines whether training is complete. Various conditions may be used to determine whether training is complete. In this embodiment, the processor 310 executes calculations on the machine learning model ML using a plurality of evaluation training sets, which are different from the training sets used for training, among the plurality of training sets of training data. As a result, the processor 310 generates a plurality of pieces of evaluation confidence data OD. The processor 310 then determines that training is complete when a condition indicating that a plurality of error values ​​obtained from the plurality of evaluation confidence data OD are small is satisfied (for example, the average error value is equal to or less than a predetermined reference value).

[0127] If the training is not complete (S165: No), the processor 310 proceeds to S145 and continues the training. If the training is complete (S165: Yes), in S170, the processor 310 stores data representing the trained machine learning model ML in the storage device 315 (e.g., the non-volatile storage device 330). Then, the processor 310 ends the processing of FIG. 5.

[0128] The trained machine learning model ML can predict the appropriate class CL (i.e., the range of transparency T) for images of the light source 910 associated with various transparency T.

[0129] A5. Recording Process: Figure 10 is a flowchart showing an example of the recording process. The processing device 100 (Figure 1) periodically transmits data of captured images showing the light source 910 captured by the digital camera CAM to the management server 300. The management server 300 records the range of transparency T estimated using the captured image data and the trained machine learning model ML in the database DT.

[0130] The recording process includes S410 and S420. S410 is executed by the processing device 100. The processor 110 of the processing device 100 executes S410 in accordance with the program P11. S420 is executed by the management server 300. The processor 310 of the management server 300 executes S420 in accordance with the second program P32.

[0131] S410 includes S510, S520, and S523. In S510, the processor 110 of the treatment device 100 turns on the light source 910 and sends a shooting instruction to the digital camera CAM. The digital camera CAM takes a photo in accordance with the instruction and generates data of the captured image. As with the shooting in S110 (FIG. 5), the shooting settings of the digital camera CAM are set to specific settings when shooting in S510. Furthermore, the shooting is performed with all manholes 891-893 of the wastewater treatment device 800 closed with their lids 891a-893a. After shooting, the processor 110 turns off the light source 910.

[0132] 11 is a diagram showing an example of data processed in the recording process. Captured image IMs shows an example of the captured image generated in S510. As with captured images IM11-IM14 described in FIGS. 6A-6D, the portion representing the light source 910 is displayed bright, and other portions are displayed dark.

[0133] In S520 (FIG. 10), the processor 110 acquires the captured image data from the digital camera CAM. In S523, the processor 110 transmits the captured image data and the device identifier (e.g., the first device identifier IDa) of the processing device 100 to the management server 300.

[0134] S420 includes S526, S530, S540, and S550. In S526, the processor 310 of the management server 300 receives data from the processing device 100. In S530, the processor 310 generates data of a target image by performing preprocessing on the captured image. The method of generating the target image is the same as the method of generating the training image described in S120 of FIG. 5. The processor 310 generates a preprocessed image by performing preprocessing (FIG. 7) using the captured image. The processor 310 uses the preprocessed image as the target image.

[0135] 11 shows a partial area Acs in the captured image IMs. The partial area Acs corresponds to the partial area Ac in FIG. 9E. The target image IMt is generated using a specific portion IMcs of the captured image IMs that corresponds to the partial area Acs. The target image IMt includes a portion representing the light source 910 and a portion 910ss surrounding the light source 910.

[0136] In S540 (FIG. 10), the processor 310 generates confidence data ODt (FIG. 11) representing the estimation result by inputting data of the target image IMt into the trained machine learning model ML. The confidence data ODt is generated by executing the operation of the machine learning model ML using the target image IMt.

[0137] In S550, the processor 310 records the estimated transparency ET (FIG. 11) represented by the certainty data ODt in the database DT. FIG. 12 is a diagram showing an example of the database DT. The database DT shows the correspondence between the device identifier ID, the date YMD, and the estimated transparency ET.

[0138] The device identifier ID is the device identifier of the processing device 100. In the example of Fig. 12, data of the first device identifier IDa and data of the second device identifier IDb are recorded in the database DT.

[0139] The date YMD represents the year, month, and day of photography. In this embodiment, the date on which the data of the captured image IMs was received in S526 (FIG. 10) is used as the date YMD. Alternatively, the processor 110 of the processing device 100 may acquire the date of photography in S510 (FIG. 10). In S523, the processor 110 may transmit data on the date of photography to the management server 300 along with the data of the captured image. The data on the date of photography may be included in the data of the captured image (e.g., Exif). The processor 310 of the management server 300 may record the date of photography represented by the data received from the processing device 100 as the date YMD in the database DT.

[0140] The estimated transparency ET indicates the range of the transparency T represented by the certainty data ODt generated in S540 (FIG. 10). As described above, the range associated with the class CL associated with the highest certainty CS is used as the estimated transparency ET.

[0141] In S550 (FIG. 10), the processor 310 records one new row of the record in the database DT, and the recording process then ends.

[0142] A6. Display Processing: FIG. 13 is a sequence diagram illustrating an example of the display processing. The administrator of the wastewater treatment devices 800A, 800B (FIG. 1) can refer to the estimated transparency ET stored in the database DT (FIG. 12) of the management server 300 by operating the terminal device 200. The processor 310 of the management server 300 executes processing as a web server that provides web pages in accordance with the third program P33. Various steps of the management server 300 in FIG. 13 are executed by the processor 310 functioning as a web server. Hereinafter, the processor 310 will be referred to as the server processor 310. The administrator also launches a web browser by operating the operation unit 250 of the terminal device 200. The processor 210 of the terminal device 200 executes processing as a web browser in accordance with the program P21. Various steps of the terminal device 200 in FIG. 13 are executed by the processor 210 functioning as a browser. Hereinafter, the processor 210 will be referred to as the terminal processor 210.

[0143] In S610, the administrator operates the operation unit 250 of the terminal device 200 to input an instruction to access the uniform resource locator (URL) of the management page, which is a web page for management of the management server 300. The terminal processor 210 accesses the web page in accordance with the instruction. Note that the server processor 310 may perform user authentication before accessing the management page. If the user authentication fails, the server processor 310 may prohibit access to the management page.

[0144] In S620, the server processor 310 transmits the management page data to the terminal device 200. In S630, the terminal processor 210 displays the management page on the display unit 240. FIG. 14(A) is a diagram showing an example of the management page. In this embodiment, the management page WP displays an input field F1 for inputting a device identifier, a send button B1, and a water quality display area MA. In S630 (FIG. 13), the water quality display area MA is blank.

[0145] In S640, the administrator operates the operation unit 250 to input into the input field F1 the apparatus identifier of the treatment device 100 associated with the wastewater treatment device to be referenced (e.g., IDa or IDb). After this, the administrator operates the send button B1. In response to the operation of the send button B1, the processor 110 transmits a request for apparatus information to the management server 300. This request includes data of the input identifier IDr, which is the apparatus identifier input into the input field F1.

[0146] In S650, in response to the request in S640, the server processor 310 references the database DT ( FIG. 12 ) and obtains the date YMD and estimated transparency ET associated with the input identifier IDr. In S660, the server processor 310 transmits device information data representing the date YMD and estimated transparency ET to the terminal device 200. In this embodiment, the device information data represents a graph to be displayed in the water quality display area MA ( FIG. 14(A) ). In S670, the terminal processor 210 displays the device information on the display unit 240. FIG. 14(B) shows an example of a management page WP representing device information. The terminal processor 210 displays the device information in the water quality display area MA of the management page WP. In the example of FIG. 14(B), the horizontal axis represents the date YMD, and the vertical axis represents the estimated transparency ET. The administrator can understand the status of the wastewater treatment device by observing the water quality display area MA. This completes the display process.

[0147] The database DT (FIG. 12) may further record a site name associated with the equipment identifier ID. The wastewater treatment equipment to be referenced may be specified using the site name instead of the equipment identifier.

[0148] As described above, in this embodiment, the processor 310 of the management server 300 executes the following processing in accordance with the programs P32 and P33. In steps S526-S530 (FIG. 10), the processor 310 acquires data for the target image IMt (FIG. 11). The target image IMt represents a captured image of the light source 910 (FIGS. 3(C) and 3(D)) that is turned on. The light source 910 is placed underwater in the water treatment tank 850 included in the wastewater treatment device 800. The target image IMt is an image captured from above the water surface WS5. As described in FIG. 11, the target image IMt includes a portion representing the light source 910 and a surrounding portion 910ss that surrounds the light source 910.

[0149] In S540 ( FIG. 10 ), the processor 310 inputs data of the target image IMt into the trained machine learning model ML to obtain an estimation result (in this embodiment, an estimated transparency ET) of the water quality of the water in the water treatment tank 850. As described in FIG. 5 , data of a plurality of training images representing images captured of a light source 910 placed underwater is used to train the machine learning model ML. The plurality of training images includes a plurality of training images captured with different water qualities. The machine learning model ML is trained to output an estimation result indicating a water quality associated with the training image data input to the machine learning model ML.

[0150] In S660 (FIG. 13), the processor 310 transmits device information data to the terminal device 200. The device information data representing the graph of the estimated transparency ET (FIG. 14(B)) is an example of data representing an estimation result. The terminal device 200 is an example of an external device that is the destination of the data representing the estimation result.

[0151] As described above, the appearance of the light source 910 placed underwater in the water treatment tank 850 may change with changes in water quality. That is, the appearance of the light source 910 represented by the target image IMt changes with changes in water quality. Therefore, the processor 310 can estimate water quality by using the target image IMt. Furthermore, the trained machine learning model ML used to obtain the water quality estimation result is trained to output an estimation result indicating the water quality associated with the training image using data from multiple training images representing images taken with different water qualities. Therefore, the processor 310 can obtain an appropriate water quality estimation result. The processor 310 then transmits data representing the water quality estimation result to an external device (in this embodiment, the terminal device 200). Therefore, the user can use the water quality estimation result via the external device. The management server 300, which performs such processing, is an example of a data processing system.

[0152] Let us assume that the light source 910 is omitted and a photographed image of the water surface in the disinfection tank 850 is used. While such a photographed image can show the appearance of the water surface, it may be difficult to show differences in water quality. In this embodiment, the use of the underwater light source 910 allows the water quality to be appropriately estimated.

[0153] Furthermore, in this embodiment, as described in S530 (FIG. 10), the processor 310 acquires data of the target image IMt by performing preprocessing on the captured image IMs (FIG. 11) of the light source 910, which was captured from above the water surface WS5 (FIG. 3C). As described in FIGS. 7 and 11, in the preprocessing, the processor 310 acquires data of the target image IMt by using a specific portion IMcs, which is a portion of the captured image IMs that includes a portion representing the light source 910 and a surrounding portion 910ss surrounding the light source 910. As described above, the appearance of the intensity and spread of light from the light source 910 when it is turned on underwater may change depending on the water quality. The specific portion IMcs can represent the intensity and spread of light from the light source 910. When data of the target image IMt is acquired by using the specific portion IMcs, the accuracy of the water quality estimation using the data of the target image IMt is improved.

[0154] Furthermore, in this embodiment, as described with reference to FIGS. 8 and 9E, when the bright pixel counter Nb is equal to the threshold value NbTh (S360: Yes), the processor 310 determines the partial region Acs ( FIG. 11 ), i.e., the specific portion IMcs, based on the pixel of interest Pt. The method for determining the partial region Ac is the same as the method for determining the partial region Ac in FIG. 9E. As described with reference to FIG. 9E, the partial region Ac includes an area Ar that is NbTh pixels or less away from the pixel of interest Pt. That is, the partial region Ac includes NbTh bright pixels that are consecutive in the second direction D2. The partial region Ac in FIG. 11 also includes NbTh bright pixels that are consecutive in the second direction D2. In this way, the processor 310 selects, as the specific portion IMcs, a portion of the captured image IMs ( FIG. 11 ) that includes an area where bright pixels equal to or greater than the threshold value NbTh are consecutive. As described in S310 and S315 ( FIG. 8 ), bright pixels are pixels that are brighter than the threshold. With this configuration, the processor 310 can appropriately select a specific portion IMcs from the captured image IMs, which is a portion that includes the portion representing the light source 910 and the surrounding portion 910ss surrounding the light source 910. Note that, as described in S315, in this embodiment, the brightness threshold is determined by analyzing the image. Alternatively, the threshold may be experimentally determined in advance so as to obtain an appropriate specific portion.

[0155] In this embodiment, the water treatment tank 850 is a disinfection tank configured to disinfect water. In the water of the disinfection tank 850, the possibility that contaminants such as biofilms will adhere to the light source 910 is reduced. In other words, the possibility that the appearance of the light source 910 represented by the target image IMt will change due to contaminants is reduced. As a result, the possibility of a decrease in the accuracy of the water quality estimation is reduced.

[0156] In this embodiment, the estimated water quality also includes transparency T. Transparency T is an index indicating the turbidity of the water. The appearance of the light source 910 represented by the target image IMt changes with changes in the turbidity of the water. Therefore, the processor 310 can appropriately estimate the index indicating the turbidity of the water by using the target image IMt.

[0157] Furthermore, in this embodiment, as described in FIG. 5 , the machine learning model ML is trained using multiple training sets obtained by a specific process PR2 including multiple set processes PR1. The training set data is a set of data on the measurement results of water quality (here, transparency T) and training image data. The set process PR1 includes S110 and S115. In S110, a digital camera CAM photographs an illuminated light source 910 disposed underwater in a water treatment tank 850 included in a training wastewater treatment device 800 ( FIGS. 3(C) and 3(D) ) from above the water surface WS5. In S115, the water quality of the water in the water treatment tank 850 is measured. As described in FIGS. 7 and 9(F) , the training image data represents a photographed image of the light source 910. As described in S135 ( FIG. 5 ), the multiple set processes PR1 include multiple specific set processes PR1 performed at multiple times when the water quality of the water treatment tank 850 differs from one another.

[0158] According to this configuration, the machine learning model ML is trained to be able to estimate various changes in water quality, and therefore the machine learning model ML can appropriately estimate water quality that exhibits various values.

[0159] Furthermore, in this embodiment, as described with reference to FIGS. 3A-3D, the light source 910 is detachably attached to the wastewater treatment device 800. The light source 910 is positioned lower than the predetermined standard water level WL5 of the water treatment tank 850 and is positioned so that it is visible when observing the water treatment tank 850 from above toward below. The light source 910 includes an upward-facing light-emitting diode (light-emitting element 912). This configuration makes it easy to acquire data on a target image that includes a portion representing the light source 910 and the surrounding area surrounding the light source 910. Furthermore, because the light source 910 can be removed from the wastewater treatment device 800, cleaning the light source 910 (e.g., removing dirt adhering to the light source 910) is easy.

[0160] In this embodiment, the wastewater treatment devices 800A and 800B are distributed wastewater treatment devices installed at the wastewater generation site and treat the wastewater. The processor 310 can acquire water quality estimates for the wastewater treatment devices 800A and 800B installed at the wastewater generation site. Furthermore, such wastewater treatment devices 800A and 800B can be installed in various locations, including locations far from the base of operations (e.g., the office location) of the administrator of the wastewater treatment devices 800A and 800B. Furthermore, multiple wastewater treatment devices 800A and 800B can be installed in locations far from each other. Traveling to the installation site of such wastewater treatment devices to measure water quality is a significant burden for the administrator. In this embodiment, the administrator can remotely access the water quality estimates via an external device such as the terminal device 200 without traveling to the installation site of the wastewater treatment device. Therefore, the administrator's burden for managing the wastewater treatment device is significantly reduced.

[0161] Moreover, in this embodiment, the wastewater treatment device 800 (FIG. 2(A)) includes a body 801 having manholes 891-893, and lids 891a-893a that cover the manholes 891-893. The body 801 houses water treatment tanks 810-850. As explained in S510 (FIG. 10), the target image IMt represents an image taken inside the body 801 with the manholes 891-893 covered by the lids 891a-893a. With this configuration, the variation in environmental brightness between multiple images is reduced. In other words, the variation in environmental brightness between multiple target images is reduced. Therefore, the accuracy of water quality estimation using data from the target image is improved.

[0162] In this embodiment, the system 1000 includes a treatment device 100 and a management server 300. The processor 110 of the treatment device 100 executes the following process according to program P11. The processor 310 of the management server 300 executes the following process according to programs P32 and P33. In S520 (FIG. 10), the processor 110 acquires data of a captured image IMs (FIG. 11) from the digital camera CAM. The digital camera CAM is configured to generate data of the captured image IMs representing the light source 910 by capturing an image of the light source 910 (FIGS. 3(C) and 3(D)) while it is turned on. The light source 910 is disposed underwater in a water treatment tank 850 included in the wastewater treatment device 800. The digital camera CAM captures an image of the light source 910 from above the water surface WS5.

[0163] In S526-S530 (Figure 10), the processor 310 of the management server 300 uses data from the captured image IMs (Figure 11) to obtain data for the target image IMt, which includes a portion representing the light source 910 and a surrounding portion 910ss surrounding the light source 910.

[0164] In S540 ( FIG. 10 ), the processor 310 inputs data of the target image IMt into the trained machine learning model ML to obtain an estimation result (in this embodiment, an estimated transparency ET) of the water quality of the water in the water treatment tank 850. As described in FIG. 5 and other drawings, the machine learning model ML has been trained to use multiple training images taken with different water qualities to output an estimation result indicating the water quality associated with the training image.

[0165] In S660 (FIG. 13), the processor 310 transmits device information data to the terminal device 200. As described above, the device information data represents the graph to be displayed in the water quality display area MA (FIG. 14(A)). The graph of estimated transparency ET represented by the device information (FIG. 14(B)) is an example of an estimation result. The transmission of such device information data is an example of processing for displaying the estimation result on the display unit 240 of the terminal device 200.

[0166] According to the above configuration, the user can recognize the estimated water quality result by observing the display unit 240 of the terminal device 200.

[0167] In this embodiment, the processor 310 of the management server 300 executes the following process in accordance with the first program P31. In the training process of FIG. 5 , the processor 310 executes a set process PR1 multiple times. The set process PR1 includes S110 and S115. In S110, a digital camera CAM photographs a lit light source 910 disposed underwater in a water treatment tank 850 included in a training wastewater treatment device 800 ( FIGS. 3(C) and 3(D) ) from above the water surface WS5. In S115, the water quality of the water in the water treatment tank 850 is measured. As described in S135 ( FIG. 5 ), the multiple set processes PR1 include specific multiple set processes PR1 executed at multiple times when the water quality of the water in the water treatment tank 850 differs from one another.

[0168] In the repeated step S120, the processor 310 acquires data for a plurality of training images by using data for a plurality of captured images of the light source 910 (for example, data for the captured image IM11 ( FIG. 9(A) )) obtained by multiple sets of the set process PR1. As described in FIGS. 7 and 9(F) , the plurality of training images include a portion representing the light source 910 and a surrounding portion 910s surrounding the light source 910.

[0169] In S150-S155, the processor 310 uses multiple training sets to adjust the parameters of the machine learning model ML so as to output an estimation result indicating water quality associated with the data of training images input to the machine learning model ML. The training set is a set of data of multiple training images and water quality measurement results associated with the data of each training image.

[0170] According to the above configuration, the processor 310 can appropriately adjust the parameters of the machine learning model ML.

[0171] In this embodiment, the management server 300 estimates the water quality of each of multiple wastewater treatment devices (e.g., the first wastewater treatment device 800A and the second wastewater treatment device 800B) using a common machine learning model ML. Therefore, the burden of preparing the machine learning model ML is reduced compared to when a machine learning model ML is prepared for each wastewater treatment device.

[0172] B. Second Example: FIG. 15 is a flowchart showing a second example of the recording process. The difference from the recording process of FIG. 10 is that the processes of S530 and S540 are executed by the processing device 100 instead of the management server 300 ( FIG. 1 ). In this example, S410b executed by the processing device 100 includes S530 and S540 in addition to S510 and S520. Furthermore, S543b is executed instead of S523. The program P11 of the processing device 100 is modified to execute this process. S420b executed by the management server 300 includes S550. Furthermore, S546b is executed instead of S526. S530 and S540 are omitted from S420b. The program P32 of the management server 300 is modified to execute this process.

[0173] In this embodiment, the processing device 100 calculates the estimated transparency ET using the machine learning model ML. Data of the trained machine learning model ML is stored in the non-volatile storage device 130 by the service provider when shipping the processing device 100. Alternatively, the processing device 100 may obtain the data of the machine learning model ML from a server (not shown) via a network or from a portable storage device (e.g., a USB flash drive) (not shown) connected to the processing device 100.

[0174] The processes of S510, S520, S530, and S540 are the same as the processes of S510, S520, S530, and S540 in Fig. 10. The processor 110 of the processing device 100 acquires data of the captured image IMs (Fig. 11) (S510, S520), performs preprocessing of the captured image IMs to generate data of the target image IMt (S530), and acquires data of the estimated transparency ET using the target image IMt and the machine learning model ML (S540). In S543b, the processor 110 transmits the data of the estimated transparency ET, the data of the date YMD, and the data of the identifier of the processing device 100 to the management server 300.

[0175] In S546b, the processor 310 of the management server 300 receives data from the processing device 100. The process of S550 is the same as the process of S550 in Fig. 10. In S550, the processor 310 records a new record representing the identifier of the processing device 100, the date YMD, and the estimated transparency ET in the database DT. Then, the recording process ends.

[0176] The display process executed in this embodiment is the same as the display process in Fig. 13. The administrator can operate the terminal device 200 to view the management page WP (Fig. 14(B)) showing the estimated transparency ET.

[0177] As described above, in this embodiment, the processor 110 of the processing device 100 executes the following steps in accordance with the program P11: acquire data of the target image IMt (S530 ( FIG. 15 )); acquire an estimated transparency ET by inputting the data of the target image IMt into the trained machine learning model ML (S540); and transmit data representing the estimation result to an external device (the management server 300 in this embodiment) (S543b). The processor 110 of the processing device 100 can acquire an appropriate estimated transparency ET, similar to the management server 300 in the embodiment of FIG. 10. The processor 110 also transmits data representing the water quality estimation result to the external device (the management server 300 in this embodiment). Therefore, a user can use the water quality estimation result via the external device. For example, an administrator can use the water quality estimation result by accessing the management server 300 using the terminal device 200. The processing device 100 that executes such processing is an example of a data processing system. The external device to which the data representing the estimation result is sent is not limited to the management server 300, and may be various devices different from the processing device 100. For example, the processing device 100 may send the data representing the estimation result to a terminal device of the administrator (e.g., the terminal device 200).

[0178] In addition, in this embodiment, the training process (FIG. 5) and the display process (FIG. 13) are the same as those in the first embodiment. In addition, the recording process (FIG. 15) in this embodiment is the same as the recording process (FIG. 10) in the first embodiment, except that S530 and S540 are executed by the processing device 100 instead of the management server 300. Therefore, this embodiment can provide the same various advantages as those provided by the first embodiment.

[0179] C. Third Embodiment: FIG. 16 is a flowchart illustrating a third embodiment of the recording process. The only difference from the recording process of FIG. 10 is that S510 is replaced with S510c and S550 is replaced with S550c. In this embodiment, in addition to the light source 910, an additional light source that illuminates the interior of the wastewater treatment device 800 is disposed within the wastewater treatment device 800. FIG. 17(A) is a diagram illustrating an example of the additional light source. Similar to FIG. 3(C), FIG. 17(A) illustrates a portion of the body 801, the disinfection tank 850, and the light source module 900 viewed from the side. The additional light source LT, like the digital camera CAM, is disposed above the standard water level WL5. For example, the additional light source LT is attached to the underside of the lid 893a (FIG. 2(A)), like the digital camera CAM. A power cable Lc for supplying power is connected to the additional light source LT. The power cable Lc is connected to the device interface 160 of the processing device 100 (FIG. 1).

[0180] In S510c (FIG. 16), the light source 910 is photographed. The only difference from S510 (FIG. 10) is that the processor 110 turns on the additional light source LT during photography. FIG. 17(B) is a diagram showing an example of a photographed image. The interior of the wastewater treatment device 800 is illuminated by the additional light source LT. The photographed image IMsc shows the exterior of the interior of the wastewater treatment device 800 in addition to the light from the light source 910. Although not shown, the aeration in the contact filter bed tank 830 may be biased toward certain areas. Bubbles may form on the water surface of the contact filter bed tank 830. Solid matter such as scum may float on the water surface of the treated water tank 840. The photographed image IMsc can show such an exterior of the interior of the wastewater treatment device 800. After photography, the processor 110 turns off the light source 910 and the additional light source LT.

[0181] Although not shown, the additional light source LT is also turned on during image capture in S110 of the training process ( FIG. 5 ). In both the training process and the recording process ( FIG. 16 ), a captured image showing the internal components of the wastewater treatment device 800, such as the captured image IMsc ( FIG. 17(B) ), is used. As shown in FIGS. 9(F) and 11 , a specific portion of the captured image, which includes the portion showing the light source 910 and the area surrounding the light source 910, is used to generate the training image and the target image. By extracting the specific portion from the captured image, parts of the image showing components other than the light source 910 are removed from the training image and the target image. Therefore, the influence of the portion showing components other than the light source 910 on the estimation of water quality is mitigated.

[0182] Furthermore, if the additional light source LT is excessively bright, the interior of the wastewater treatment device 800 may become brighter than the brightness of the light source 910. In this case, it may be difficult to estimate the water quality based on the appearance of the light source 910. The brightness of the additional light source LT is set to a dark brightness so that the captured image can appropriately show changes in the appearance of the light source 910 in response to changes in water quality. The brightness of the additional light source LT is determined experimentally in advance so that the water quality can be appropriately estimated using the machine learning model ML.

[0183] In S550c (FIG. 16), the processor 310 of the management server 300 records the data of the captured image together with the estimated transparency ET in a database. FIG. 18 is a diagram showing an example of the database. The only difference from the database DT in FIG. 12 is that the database DTc of this embodiment records data of the captured image IM in addition to the information ID, YMD, and ET.

[0184] The display process executed in this embodiment proceeds in the same manner as the display process of FIG. 13 . FIG. 19 is a diagram showing an example of a management page in this embodiment. The only difference from the management page WP in FIGS. 14(A) and 14(B) is that an image area IA has been added to the management page WPc in this embodiment. The device information transmitted in S660 ( FIG. 13 ) includes the date YMD, the estimated transparency ET, and the captured image IM. In S670, the terminal processor 210 displays a graph of the estimated transparency ET in the water quality display area MA and the captured image IM in the image area IA. By observing the water quality display area MA, the administrator can recognize the estimated transparency ET. Furthermore, by observing the image area IA, the administrator can recognize the internal appearance of the wastewater treatment device 800 (such as the generation of bubbles and the rise of scum).

[0185] D. Modifications: (1) The configuration of the light source (e.g., light source 910 (Figures 3(A)-3(D))) may be various configurations suitable for estimating water quality using the machine learning model ML. For example, the light-emitting element 912 may be various light-emitting diodes. The light-emitting diode may have a lens. The lens diameter may be various sizes, such as 3 mm, 5 mm, or 8 mm. The light color of the light-emitting diode may be various colors, such as white, red, green, or blue. The light source is not limited to light-emitting diodes, but may include various light-emitting elements such as incandescent bulbs. In either case, the light source is photographed from above the water surface. Therefore, it is preferable that the light source be configured to emit light that includes a component directed toward the water surface. Furthermore, the light source is not limited to being attached to the body of the wastewater treatment device (e.g., body 801 (Figure 3(C))), but may also be attached to any component of the wastewater treatment device (e.g., the wall of the water treatment tank).

[0186] (2) When water quality is estimated using photographed images of a light source, it is preferable that the appearance of the light source in the photographed image changes significantly in response to changes in water quality (the appearance of the light source includes, for example, the brightness of the light source and the distribution range of light centered on the light source). For this purpose, it is preferable that the light source be located far from the water surface, i.e., at a deep position. The depth of the light source from the water surface (for example, the depth Dp of the light source 910 from the water surface WS5 in Figure 3(C)) is, for example, preferably 15 cm or more, particularly preferably 20 cm or more, and most preferably 25 cm or more.

[0187] (3) The preprocessing is not limited to the process of FIG. 7 and may be any process that generates an image suitable for the machine learning model ML using the captured image. For example, the process of extracting a specific portion representing a light source from the captured image is not limited to the process of FIG. 8 and may be various processes. For example, a partial image of a rectangular area of ​​a predetermined size located at the center of gravity of the bright area Ab (FIG. 9(B)) may be used as the specific portion. Furthermore, binarization (S315) may be omitted. For example, the specific portion representing the light source 910 may be detected by template matching using a template image representing the light source 910. Such template matching may be performed using a grayscale image or a color image. In either case, when a specific portion is extracted from the captured image, the position of the light source in the captured image may differ between multiple wastewater treatment devices (e.g., wastewater treatment devices 800A and 800B).

[0188] The specific portion extraction process (S210) may be omitted from the preprocessing (FIG. 7). In this case, the training images used to train the machine learning model ML preferably include training images in which the positions of the light sources within the training images are different from each other. This allows the machine learning model ML to appropriately estimate water quality regardless of the position of the light source within the image.

[0189] (4) The water quality to be estimated is not limited to transparency T but may include various indices indicating the turbidity of water. For example, turbidity is an example of an index indicating the turbidity of water. Furthermore, the water quality to be estimated is not limited to an index indicating the turbidity of water but may include various water qualities that change the appearance of a light source depending on the water quality. For example, the water quality to be estimated may include the concentration of suspended solids (SS).

[0190] (5) The configuration of the wastewater treatment device is not limited to the configuration of the wastewater treatment device 800 in Figures 2(A) and 2(B), and various configurations may be used. For example, the discharge air lift pump 870 may be omitted. An anaerobic filter bed tank may be provided instead of the impurity removal tank 810. The water treatment tank for performing aerobic treatment may be a membrane treatment tank having a membrane separation device instead of the contact filter bed tank 830. The total number of water treatment tanks included in the wastewater treatment device may be any number equal to or greater than one.

[0191] The water treatment tank in which the light source is disposed is not limited to a disinfection tank, and may be various types of water treatment tanks. For example, the light source 910 may be disposed in the water of the treatment water tank 840. The wastewater treatment device may be a discharge pump tank that temporarily stores water to be discharged. The discharge pump tank may be provided with a light source and a digital camera. The light source may be attached to any component of the wastewater treatment device. Furthermore, the configuration for detachably attaching the light source to the wastewater treatment device is not limited to a configuration including the pipe holder 809 (FIGS. 3(C) and 3(D)) and the support portion 920, and may be any configuration. For example, the light source may be detachably attached to the wall of the water treatment tank using a clamp.

[0192] (6) The method for generating training data is not limited to the method described in S110-S135 of FIG. 5 , and various methods may be used. For example, the processes of S110-S135 may be performed using multiple wastewater treatment devices. That is, the training wastewater treatment device used to generate the training data may include multiple wastewater treatment devices. Furthermore, the model of the training wastewater treatment device may differ from the model of the wastewater treatment device whose water quality is to be estimated. In this case, it is preferable that the light source in each of the training wastewater treatment device and the wastewater treatment device whose water quality is to be estimated is placed in a water treatment tank that performs the same water treatment. Here, the water treatment may be selected from, for example, solid separation, anaerobic treatment, aerobic treatment, and disinfection. In either case, it is preferable that the difference in depth of the light source from the water surface between the multiple light sources of the multiple wastewater treatment devices is small. Here, if the difference in estimated water quality caused by the difference in depth is within a predetermined allowable error range of the estimated water quality, the difference in depth may be acceptable.

[0193] The water quality may be measured by various methods suitable for the water quality. For example, a sensor may be placed in the water. The water quality may then be measured using a signal obtained from the sensor. The water quality may also be measured by physical or chemical water analysis. The water quality may also be measured by sensory evaluation using human senses (e.g., vision). Repeated photographing and water quality measurement may be performed first, and the generation of training images and storage of a training set may be performed later. Water quality measurement may also be performed using equipment located far away from the wastewater treatment device. In this case, an operator may sample water in S115 ( FIG. 5 ) and measure the water quality separately. Repeated photographing and water sampling may be performed first, and the water quality measurement, generation of training images, and storage of a training set may be performed later.

[0194] In either case, in order to form an appropriate correspondence between the training images and the water quality, it is preferable that the time difference between "taking a photograph (e.g., S110)" and "measuring or sampling water (e.g., S115)" is small. Here, if the difference in water quality caused by the time difference is within a predetermined allowable error range of the estimated water quality, the time difference may be acceptable.

[0195] (7) The configuration of the machine learning model is not limited to the configuration of the machine learning model ML in FIG. 4 , and various models that estimate water quality using captured images of a light source may be used. For example, the machine learning model may be a regression model that estimates a numerical value representing the estimated water quality. The convolutional layer may be omitted from the machine learning model. Here, the machine learning model may include multiple fully connected layers. In either case, the machine learning model may be configured to estimate water quality using color image data (e.g., RGB bitmap data).

[0196] Furthermore, a machine learning model may be prepared for each model of the wastewater treatment device. In estimating the water quality, a machine learning model associated with the model of the wastewater treatment device may be used.

[0197] In either case, the method for training the machine learning model is not limited to the method described in S145-S170 of Fig. 5, and may be any method suitable for the machine learning model. Training of the machine learning model may be performed by a device other than the management server 300 (for example, a computer owned by the service provider).

[0198] (8) The water quality estimation results may be displayed in various formats, not limited to graphs (FIGS. 14(B) and 19). For example, a numerical value indicating the estimated water quality, an image of a meter with a needle that swings in response to the estimated water quality, or a color patch having a color corresponding to the numerical value indicating the estimated water quality may be displayed.

[0199] (9) The configuration of the data processing system that acquires target images, estimates water quality using a machine learning model, and outputs the estimated data is not limited to the configurations of the above-described embodiments and modifications, and may have various configurations. For example, the processing device 100 ( FIG. 1 ) and the management server 300 may each perform a portion of the multiple processes included in the recording process. Here, the roles of the processing device 100 and the management server 300 may differ from those shown in FIGS. 10 and 15 . For example, the processing device 100 may perform S510-S530, and the management server 300 may perform S540. In this case, a system including the processing device 100 and the management server 300 is an example of a data processing system. Here, the network used for communication between the processing device 100 and the management server 300 may include one or both of a cellular network CN and the Internet IT. In this way, multiple devices (e.g., computers) that can communicate with each other via a network may each perform a portion of the data processing functions of the data processing system, and as a whole, provide the data processing functions (a system including these devices corresponds to a data processing system). In either case, the data processing system may transmit data representing the estimation result to an external device, which may be a variety of devices different from the data processing system.

[0200] (10) The wastewater treatment device may be various types of wastewater treatment device instead of a decentralized wastewater treatment device. For example, the wastewater treatment device may be a centralized wastewater treatment device, such as a wastewater treatment facility connected to a sewer system. In this case, when photographing the light source, it is preferable to shield the water treatment tank so that light other than the light from the light source prepared for the photographing (e.g., the light source 910 in FIG. 3(C) or the light source 910 and the additional light source LT in FIG. 17(A)) is not irradiated onto the water treatment tank. In either case, photographing the light source for generating training data and photographing the light source for estimating water quality are preferably performed in an environment with minimal fluctuation in brightness within the digital camera's photographing range (i.e., the photographed area). This reduces the possibility of a decrease in the accuracy of the water quality estimation. Note that if the fluctuation in brightness within the photographing range is within a predetermined tolerance for the estimated water quality, the brightness fluctuation may be acceptable.

[0201] Furthermore, the operating mode of the digital camera (e.g., digital camera CAM (FIG. 2A)) may be set to auto mode (auto mode is a mode in which the shooting settings are automatically adjusted by the digital camera). When the brightness of the shooting range fluctuates little, the shooting settings adjusted in auto mode are less likely to change over multiple shootings. Therefore, the possibility of a decrease in the accuracy of the water quality estimation due to variations in the shooting settings is reduced. For example, as described in S510 (FIG. 10), the light source may be photographed by a digital camera placed inside the body of the wastewater treatment device when the manhole is covered with a lid. In this case, the inside of the body of the wastewater treatment device is dark, i.e., the brightness of the shooting range is roughly constant. Therefore, variations in the shooting settings are reduced compared to when the light source is photographed with the manhole open.

[0202] In each of the above embodiments, a part of the configuration realized by hardware may be replaced by software, and conversely, a part or all of the configuration realized by software may be replaced by hardware. For example, processing by a machine learning model (e.g., the machine learning model ML in FIG. 4) may be executed by a dedicated hardware circuit such as an Application Specific Integrated Circuit (ASIC).

[0203] Furthermore, when some or all of the functions of the present disclosure are realized by a computer program, the program can be provided in a form stored on a computer-readable recording medium (e.g., a non-transitory recording medium). The program can be used while stored on the same or a different recording medium (computer-readable recording medium) from when it was provided. The "computer-readable recording medium" is not limited to portable recording media such as memory cards and CD-ROMs, but can also include internal storage devices within a computer, such as various ROMs, and external storage devices connected to a computer, such as a hard disk drive.

[0204] The above-described examples and modifications can be combined as appropriate. The above-described examples and modifications are provided to facilitate understanding of the present disclosure and are not intended to limit the present invention. The present invention may be modified or improved without departing from the spirit thereof, and the present invention includes equivalents thereof.

[0205] The present invention can be suitably used for processing data relating to the water quality of a wastewater treatment device.

[0206] DESCRIPTION OF SYMBOLS 100A...first treatment device, 100B...second treatment device, 200...terminal device, 300...management server, 110, 210, 310...processor, 115, 215, 315...storage device, 120, 220, 320...volatile storage device, 130, 230, 330...nonvolatile storage device, 140, 240...display unit, 150, 250...operation unit, 160...device interface, 180, 280, 380...communication interface, 800A...first wastewater treatment device, 800B...second Wastewater treatment device, 801...body, 802, 803...partition wall, 804...inlet, 805...outlet, 809...pipe holder, 810...impurity removal tank, 812...inlet baffle, 814, 824, 836...opening, 820...anaerobic filter bed tank, 822...filter material, 830...contact filter bed tank, 832...contact material, 833...aerobic filter material, 834...air diffuser, 840...treated water tank, 850...disinfection tank, 854...chemical cylinder, 858...storage chamber, 870...discharge air lift pump, 872, 882... Intake port, 874...exhaust port, 880...circulating air lift pump, 891...first manhole, 892...second manhole, 893...third manhole, 891a-893a...lid, 900...light source module, 910, 910a, 910b...light source, 912...light emitting element, 914, 921...pipe, 916...connection part, 920...support part, 921u...upper end, 923...corner pipe, 930...cap, 1000...system, CN...cellular network, IT...internet Net, P11, P21, P31, P32, P33...program, CAM, CAMa, CAMb...digital camera, 9c, Lc...power cable, Cc...control cable, Ac, Acs...partial region, IMc, IMcs...specific part, ET...estimated transparency, IMt...target image, LT...additional light source, ML...machine learning model, MLc...pre-processing unit, MLf...post-processing unit, PR1...set processing, PR2...specific processing, T...transparency, WL, WS5...water surface, WL5...standard water level

Claims

1. A data processing system comprising: a target image acquisition unit configured to acquire data of a target image representing an image of a light source placed underwater in a water treatment tank included in a wastewater treatment device, the target image including a portion representing the light source and a surrounding portion surrounding the light source; an estimation unit configured to acquire an estimation result of the water quality of the water in the water treatment tank by inputting the data of the target image into a trained machine learning model, the trained machine learning model being trained to use data of a plurality of training images representing images taken with a light source placed underwater at different water qualities, and output an estimation result indicating a water quality that corresponds to the data of the training images input to the machine learning model; and a transmission unit configured to transmit data representing the estimation result to an external device.

2. A data processing system as described in claim 1, wherein the target image acquisition unit acquires data of the target image by using a specific portion of an image of the light source taken from above the water surface, the specific portion including a portion representing the light source and a surrounding portion surrounding the light source.

3. A data processing system according to claim 2, wherein the target image acquisition unit selects a portion of the captured image that includes an area where a predetermined number or more of pixels that are brighter than a threshold value are consecutive as the specific portion.

4. A data processing system according to any one of claims 1 to 3, wherein the water treatment tank is a disinfection tank configured to disinfect water.

5. A data processing system according to any one of claims 1 to 4, wherein the water quality includes an index indicating water turbidity.

6. A data processing system according to any one of claims 1 to 5, wherein the trained machine learning model has been trained using multiple sets of training image data and water quality measurement results obtained by specific processing including multiple set processes, the set processes including photographing a light source that is turned on and placed underwater in a first water treatment tank included in a training wastewater treatment device from above the water surface and measuring the water quality of the water in the first water treatment tank, the training image data representing the photographed image of the light source, and the multiple set processes including specific multiple set processes performed at multiple times when the water quality of the water in the first water treatment tank differs from each other.

7. A data processing system according to any one of claims 1 to 6, wherein the light source is detachably attached to the wastewater treatment device, the light source is positioned lower than a predetermined standard water level in the water treatment tank and is visible when observing the water treatment tank from above looking downward, and the light source includes an upward-facing light-emitting diode.

8. A data processing system according to any one of claims 1 to 7, wherein the wastewater treatment device is a distributed wastewater treatment device that is installed at a wastewater generation site and treats the wastewater.

9. A data processing system according to claim 8, wherein the wastewater treatment device includes a body having a manhole that houses the water treatment tank, and a lid that covers the manhole, and the target image represents the image taken inside the body with the manhole covered by the lid.

Images