JP7632198B2 - Drainage abnormality detection device, trained model generation device, drainage abnormality detection method, and trained model generation method - Google Patents

Description

The present disclosure relates to a trained model generation device that generates a trained model used to detect abnormalities in rainwater, a rainwater anomaly detection device that detects abnormalities in rainwater, a trained model generation method, and a rainwater detection method.

Conventionally, from the perspective of environmental conservation, rainwater discharged from various facilities is constantly inspected by humans to check for abnormalities before it leaves the premises. However, constant manual monitoring of rainwater places a heavy burden on the people doing the monitoring, so it would be desirable for the inspection of rainwater to be carried out automatically.

As a method for automatically inspecting water for abnormalities, for example, Patent Document 1 discloses an information processing method that can appropriately determine the quality of water quality when taking water from a river. This information processing method involves acquiring an image of water near a water intake, inputting the acquired image into a trained model that outputs a judgment result of whether the water intake is good or bad when an image of water near the intake is input, and outputting the judgment result from the trained model. This trained model is then generated by having a computer carry out learning using training data collected by preparing water of various turbidities.

JP 2020-125618 A

However, there is a possibility that mud and other substances may be mixed into wastewater such as rainwater discharged from facilities. Even though such wastewater is not actually abnormal, the apparent turbidity of the wastewater changes due to the coloring caused by the mud. Therefore, it is not possible to determine whether or not there is an abnormality in the wastewater based solely on turbidity, as in conventional technology. For this reason, when using a trained model generated using training data collected by the above method, there is a possibility that wastewater that is not actually abnormal may be mistakenly determined to be abnormal.

Therefore, there is a need for technology that can prevent wastewater colored with non-abnormal factor colors, which may be mistakenly recognized as abnormal if mixed into wastewater, from being over-detected as abnormal, and can improve the accuracy of detecting abnormalities in wastewater.

The present disclosure can be realized in the following forms.
According to one embodiment of the present disclosure, there is provided a drainage abnormality detection device that detects abnormality in drainage from a facility, which appears as a color. The drainage abnormality detection device includes an image acquisition unit that acquires a drainage image, which is an image of drainage, and a detection unit that detects abnormality in drainage from the facility using a trained model that detects abnormality in drainage from the facility using a trained model that receives drainage images as input and outputs an evaluation of the degree of abnormality, and detects the abnormality in drainage from the facility using a trained model that ...
According to one embodiment of the present disclosure, there is provided a trained model generation device for generating a trained model used to detect an abnormality in wastewater from a facility, which appears as a color. The trained model generation device includes a teacher data acquisition unit that acquires, as teacher data, correspondence information that associates a plurality of simulated wastewater images including captured images of a plurality of simulated wastewaters obtained by mixing an abnormality factor color component indicating an abnormality in the wastewater and a non-abnormality factor color component that is a color that may be mistakenly recognized as an abnormality when mixed into the wastewater, the simulated wastewater images each appearing in a different color due to a different concentration of the non-abnormality factor color component, with an evaluation of the degree of abnormality of each of the simulated wastewater images, and a learning unit that generates a trained model that uses the teacher data to learn and outputs an evaluation of the degree of abnormality.
The present disclosure can also be realized in the following forms.

(1) According to one embodiment of the present disclosure, there is provided a drainage abnormality detection device that detects an abnormality in drainage from a facility that appears as a color. The drainage abnormality detection device includes an image acquisition unit that acquires a drainage image that is an image of drainage, and a detection unit that detects an abnormality in drainage from the facility using a trained model that receives drainage images as input and outputs an evaluation of the degree of abnormality, and detects the abnormality in drainage from the facility using the trained model, which is generated by learning using a plurality of simulated drainage images that are imaged images of a plurality of simulated drainages that are obtained by mixing an abnormality factor color component that indicates an abnormality in the drainage and a non-abnormality factor color component that is a color that may be mistakenly recognized as an abnormality when mixed into the drainage, and that corresponds to an evaluation of the degree of abnormality of each of the simulated drainage images as teacher data.
According to the drainage abnormality detection device of the above-described form, it is possible to suppress overdetection of drainage colored with a non-abnormal factor color, which is a color that may be mistakenly recognized as an abnormality if mixed into water, as an abnormality, thereby improving the accuracy of detecting drainage abnormalities.
(2) In the drainage abnormality detection device of the above-described form, the image acquisition unit may acquire as the drainage image a white background image obtained by imaging drainage water flowing through a flow path having a white bottom under lighting that blocks external light.
According to this drainage abnormality detection device, it is possible to suppress disturbances such as external light and improve the accuracy of detecting drainage abnormalities.
(3) In the drainage abnormality detection device of the above embodiment, the image acquisition unit acquires, as the drainage images, a white background image obtained by imaging drainage flowing through a flow path having a white bottom under lighting that blocks external light, and a colored background image obtained by imaging drainage flowing through a flow path having a bottom colored a color other than white, and the detection unit may detect abnormalities in drainage from the facility by using, for the white background image acquired by the image acquisition unit, a trained model for white background that takes a white background image as an input and outputs an evaluation of the degree of abnormality, and, for the colored background image acquired by the image acquisition unit, a trained model for colored background that takes a colored background image as an input and outputs an evaluation of the degree of abnormality.
With this drainage abnormality detection device, for example, if white paint is mixed into the drainage, there is a high possibility of overdetection using a white background image, so by performing detection using a colored background image, the accuracy of detecting drainage abnormalities can be improved.
(4) In the drainage anomaly detection device according to any of the above aspects, the drainage may be rainwater, and the non-drainage anomaly factor color may be the color of mud mixed into the rainwater.
According to the drainage abnormality detection device of this embodiment, it is possible to improve the accuracy of detecting abnormalities in rainwater containing mud.
(5) According to another aspect of the present disclosure, there is provided a trained model generation device that generates a trained model used to detect an abnormality in wastewater from a facility, which appears as a color. The trained model generation device includes a teacher data acquisition unit that acquires, as teacher data, correspondence information that associates a plurality of simulated wastewater images, which are captured images of a plurality of simulated wastewaters each appearing in a different color, obtained by mixing an abnormality factor color component indicating an abnormality in the wastewater and a non-abnormality factor color component that is a color that may be mistakenly recognized as an abnormality when mixed into the wastewater, with an evaluation of the degree of abnormality of each of the simulated wastewater images, and a learning unit that generates a trained model that uses the teacher data to learn and outputs an evaluation of the degree of abnormality.
According to this form of trained model generation device, it is possible to generate a trained model that can suppress overdetection of wastewater colored with non-abnormal factor colors, which are colors that may be mistakenly recognized as abnormal if mixed into water, and can improve the accuracy of detecting abnormalities in wastewater.

The present disclosure can be realized in various forms, such as the drainage anomaly detection device and trained model generation device described above, as well as a drainage anomaly detection method, a trained model generation method, a program that causes a computer to execute a process for detecting drainage anomalies from a predetermined facility, and a program that causes a computer to execute a process for generating a trained model.

1 is a schematic diagram illustrating a configuration example of a rainwater anomaly detection system as a drainage anomaly detection system of a first embodiment. FIG. 10 is a flowchart illustrating an example of a procedure for a rainwater abnormality detection process performed by a control unit. FIG. 1 is a schematic diagram illustrating an example of the configuration of a trained model generation device. 13 is a flowchart illustrating an example of a procedure for a trained model generation process performed by a control unit. 1 is a portion of a table showing an example of multiple data sets used as training data. 13 is another part of a table showing an example of multiple data sets used as training data. FIG. 13 is a schematic diagram illustrating a configuration example of a rainwater anomaly detection system according to a second embodiment. FIG. 13 is a schematic diagram illustrating a configuration example of a rainwater anomaly detection system according to a third embodiment.

A. First embodiment:
FIG. 1 is a schematic diagram showing a configuration example of a rainwater abnormality detection system 10 as a drainage abnormality detection system of the first embodiment. The rainwater abnormality detection system 10 is a drainage abnormality detection system that automatically detects whether there is an abnormality in rainwater, taking rainwater discharged from various facilities such as a factory as an example of drainage, before the rainwater leaves the premises of the facility. For example, in a factory, the system automatically detects whether there is an abnormal coloring in rainwater due to the inclusion of a substance such as metal that has a negative effect on the environment. In the following, as shown in FIG. 1, the rainwater that flows through the gutter flow path 22 of the rainwater gutter 20 and before flowing into the rainwater tank 24 from the inlet 26 is described as the target of abnormality detection. In addition, in the rainwater gutter 20 shown in FIG. 1, the side wall on the front side of the gutter flow path 22 is omitted for ease of illustration, and the same applies to the figures shown below.

The rainwater anomaly detection system 10 includes a rainwater anomaly detection device 100 as a drainage anomaly detection device, and an imaging device 200. The imaging device 200 is connected so as to be controllable by the rainwater anomaly detection device 100.

The imaging device 200 can be configured with various cameras, and is a device that captures video or still images and outputs the captured images (video or still images) to the rainwater abnormality detection device 100. The imaging device 200 captures an image of rainwater in a flow path portion that corresponds to a preset imaging area IA of the gutter flow path 22.

The rainwater anomaly detection device 100 can be configured with various computer devices such as a server device or a personal computer. The rainwater anomaly detection device 100 acquires a rainwater image, which is a drainage image captured by the imaging device 200, and detects an abnormal state of rainwater for the input of the acquired rainwater image using a trained model 134 described later.

The rainwater abnormality detection device 100 has a control unit 110, an interface (IF) 120, and a memory unit 130.

IF120 is an interface that enables communication between the imaging device 200 and an external computer (not shown).

The storage unit 130 has a volatile storage area such as a static random access memory (SRAM), a dynamic random access memory (DRAM), or a flash memory, and a non-volatile storage area such as an EEPROM or a hard disk. The storage unit 130 stores the program 132 and data to be referenced during processing in advance. The program 132 stored in the storage unit 130 may be a program 132 read from a recording medium (not shown) that can be read by the rainwater anomaly detection device 100. The program 132 may also be a program downloaded from an external computer (not shown) connected to a communication network (not shown) and stored in the storage unit 130. The storage unit 130 also stores a file 134f that constitutes the trained model 134. The trained model file 134f may also be a trained model file 134f read from a recording medium that can be read by the rainwater anomaly detection device 100, like the program 132. The recording medium is created by recording the trained model file 134f generated by the trained model generation device 300 described below. The trained model file 134f may also be downloaded from the trained model generation device 300 connected to a communication network (not shown) and stored in the storage unit 130.

The control unit 110 has one or more arithmetic processing devices such as a central processing unit (CPU), a micro-processing unit (MPU), or a graphics processing unit (GPU). The control unit 110 reads and executes the program 132 stored in the storage unit 130, and reads the trained model file 134f to acquire (receive) the rainwater image output from the imaging device 200 and function as various functional blocks for detecting the abnormal state of the rainwater. Specifically, the control unit 110 functions as an image acquisition unit 112 that acquires the rainwater image received from the imaging device 200 via the IF 120, a detection unit 114 that detects the abnormality of the rainwater using the trained model 134 that receives the rainwater image as an input and outputs an evaluation of the degree of the abnormality of the rainwater, and a detection result output unit 116 that outputs the detection result of the abnormality of the rainwater. The trained model file 134f is generated in advance by the trained model generation device 300 (see FIG. 3) described later and stored in the storage unit 130 of the rainwater anomaly detection device 100.

Figure 2 is a flowchart showing an example of the procedure for the rainwater abnormality detection process (drainage abnormality detection method) by the control unit 110. The control unit 110 acquires a rainwater image captured by the imaging device 200 via the image acquisition unit 112 (step S110). The control unit 110 then inputs the acquired rainwater image to the trained model 134, detects an abnormality in the rainwater via the detection unit 114 according to the evaluation of the degree of the abnormality in the rainwater output by the trained model 134 (step S120), and outputs the detection result of the rainwater abnormality via the detection result output unit 116 (step S130).

As an output of the rainwater abnormality detection result, for example, the trained model 134 may output the presence or absence of rainwater abnormality as an evaluation of the degree of the rainwater abnormality, and may display the information on the presence or absence of rainwater abnormality on a display unit (not shown) to notify the monitor monitoring the rainwater abnormality. In particular, when a rainwater abnormality is detected, in addition to the display, the monitor may be notified by sound or light, etc., to prompt the monitor to check the display.

The trained model 134 can be generated by the trained model generation device 300 described below. FIG. 3 is a schematic diagram showing an example of the configuration of the trained model generation device 300. The trained model generation device 300 can be configured with various computer devices such as a personal computer. As described below, the trained model generation device 300 acquires training data and, by learning using the acquired training data, generates a trained model that outputs information that determines the abnormal state of rainwater in response to an input of a rainwater image that captures rainwater.

The trained model generating device 300 has a control unit 310, an interface (IF) 320, and a memory unit 330.

IF320 is an interface that allows input of a data set to be used for training data. The data set includes data of a simulated rainwater image as a simulated drainage image, which will be described later, and classification data as an evaluation result associated with the simulated rainwater image. The classification is data that indicates whether the simulated rainwater is in a normal or abnormal state, as an evaluation result of the abnormal state of the simulated rainwater.

The storage unit 330 has a volatile storage area and a non-volatile storage area, similar to the storage unit 130 (see FIG. 1) of the rainwater anomaly detection device 100. The storage unit 330 stores a program 332 and data to be referenced during processing in advance. The program 332 stored in the storage unit 330 may be a program 332 read from a recording medium (not shown) that is readable by the model generating device 300. The program 332 may also be a program downloaded from an external computer (not shown) connected to a communication network (not shown) and stored in the storage unit 330. In addition, the storage unit 330 stores a plurality of acquired data sets as teacher data 334, as described below, and stores a file 134f that constitutes the generated trained model 134.

The control unit 310 has one or more arithmetic processing units such as a CPU, MPU, GPU, etc., similar to the control unit 110 (see FIG. 1) of the rainwater anomaly detection device 100. The control unit 310 reads and executes the program 332 stored in the memory unit 330 to function as various functional blocks that take rainwater images captured of rainwater as input and generate a trained model that outputs an evaluation of the degree of anomaly (presence or absence of anomaly). Specifically, the control unit 310 functions as a teacher data acquisition unit 312 that acquires a data set used for teacher data via the IF 320, and a learning unit 314 that uses the teacher data to learn and generate a trained model 134 that takes rainwater images as input and outputs an evaluation of the degree of anomaly (presence or absence of anomaly).

Figure 4 is a flowchart showing an example of the procedure of a trained model generation process (model generation method) by the control unit 310. The control unit 310 acquires multiple data sets input via the IF 320 by the teacher data acquisition unit 312, and stores them in the memory unit 330 as teacher data 334 (step S310). The control unit 310 then generates a trained model 134 constructed by learning using the teacher data 334 by the learning unit 314, and stores a file 134f constituting the trained model 134 in the memory unit 330.

The learning unit 314 performs deep learning using multiple data sets as teacher data 334, which are correspondence information that associates the simulated rainwater image with a classification (abnormal or normal) as an evaluation result of the degree of abnormality of the simulated rainwater associated with the simulated rainwater image. As a result, the learning unit 314 constructs (generates) a neural network as the trained model 134, which receives the rainwater image as input and outputs the presence or absence of an abnormality in the rainwater shown in the rainwater image as an evaluation of the degree of abnormality. Note that the neural network is a CNN (Convolution Neural Network). However, the trained model 134 is not limited to a CNN, and may be a trained model 134 constructed with another learning algorithm, such as a neural network other than a CNN.

Figures 5A and 5B are tables showing examples of multiple data sets used as training data 334. The multiple data sets can be obtained as described below.

Because it is difficult to actually prepare rainwater in an abnormal state, multiple simulated rainwater samples are prepared to obtain multiple simulated rainwater images. The simulated rainwater images are obtained by capturing images of the prepared simulated rainwater using an imaging device similar to the imaging device 200 (see FIG. 1) used in the rainwater abnormality detection system 10.

In order to intentionally create color changes due to coloring according to abnormal conditions, the simulated rainwater is created by mixing abnormality factor color components that indicate abnormalities in rainwater with non-abnormality factor color components that may be mistakenly recognized as abnormal when mixed into rainwater. The combination of abnormality factor color, abnormality factor color density, and non-abnormality factor color density uses 18 types of simulated rainwater (see Figures 5A and 5B) created using an L18 orthogonal array in order to efficiently and comprehensively obtain multiple simulated rainwater images obtained by capturing multiple simulated rainwater images and their respective evaluation results. However, this is not limited to this, and any combination of abnormality factor color, abnormality factor color density, and non-abnormality factor color density may be used as long as multiple simulated rainwater images obtained by capturing multiple simulated rainwater images and their respective evaluation results can be comprehensively obtained.

As described above, the abnormality factor colors that indicate abnormality in rainwater are colors that are caused by the inclusion of metals or other substances that have a negative impact on the environment. In this example, three types of abnormality factor colors are used as examples: red, yellow, and black. The concentrations of the abnormality factor colors are three levels: the concentration of the original solution, the concentration of a 100-fold dilution of the original solution, and the concentration of a 1000-fold dilution of the original solution. As an example, the red original solution is a colored solution in which 1 L of water is colored with 2.5 g of red coloring (e.g., food coloring). The same applies to the yellow original solution and the black original solution. As an example, the color of mud is used as an example of the color of non-abnormality factor colors, and the concentrations of non-abnormality factor colors are determined by the amount of mud mixed in, and three levels of concentration are used: no mixing, little mixing, and much mixing. As an example, the mixing of 30 g of mud in 1 L of water is called little mixing, and the mixing of 60 g of mud is called much mixing. However, this is not a limitation, and it is sufficient if the densities of multiple non-anomalous factor colors can be combined with the color and density of the abnormal factor color so that multiple simulated rainwater images captured from multiple simulated rainwater samples and their respective evaluation results can be comprehensively obtained. Note that the abnormal factor colors may include colors that are not related to adverse effects on the environment and should not be shown by the rainwater that is the subject of observation, and that many people may recognize as colored.

The classification of the degree of abnormality of each simulated rainwater is performed, for example, by having a predetermined number of evaluators evaluate the captured simulated rainwater images according to a predetermined scoring method for each type of simulated rainwater, and classifying the evaluation results according to predetermined classification conditions. The classification shown in Figures 5A and 5B is performed as follows, as an example. Fifty evaluators who manage the rainwater each score three simulated rainwater images captured for each type of simulated rainwater, and the average score is obtained as the evaluation result. The scoring method is as follows: if it is judged that there is no problem, score 1 is given; if it is judged that the trend needs to be closely watched, score 2 is given; if it is judged that the site needs to be checked, score 3 is given; and if it is judged that the item needs to be collected immediately, score 4 is given. The classification conditions are that an average score of less than 2 is normal, and an average score of 2 or more is abnormal. However, the above scoring method and classification conditions are merely examples and are not limited to these. Any scoring method and classification conditions can be used as long as they can score the degree of abnormality in rainwater and classify the degree of abnormality in rainwater into normal and abnormal depending on the average value of the score.

The datasets used as the teacher data 334 may include not only the 18 types of datasets related to simulated rainwater described above, but also datasets related to actual rainwater that is determined to be in a normal state.

The file 134f constituting the trained model 134 generated by the trained model generating device 300 is recorded on a recording medium (not shown) in the trained model generating device 300 via a recording device (not shown), and can be stored in the memory unit 130 of the rainwater anomaly detection device 100 (see FIG. 1) via this recording medium. In addition, the file 134f may be stored in the memory unit 130 of the rainwater anomaly detection device 100 by being transmitted via a communication network to the rainwater anomaly detection device 100 connected via a communication network (not shown).

In the trained model generating device 300 described above, the trained model 134 is generated using multiple simulated rainwater obtained by mixing anomaly factor color components indicating drainage abnormalities with non-anomalous factor color components that may be mistakenly recognized as abnormal when mixed with drainage. Specifically, the trained model 134 is generated by learning teacher data 334 from multiple data sets as correspondence information that corresponds multiple simulated rainwater images obtained by capturing multiple simulated rainwater images with classifications (abnormal or normal) as an evaluation of the degree of abnormality of the simulated rainwater shown by each simulated rainwater image. Therefore, the rainwater anomaly detection device 100 using this trained model 134 can suppress overdetection of rainwater colored by mud contamination as abnormal, and can improve the detection accuracy of rainwater anomalies.

B. Second embodiment:
FIG. 6 is a schematic diagram showing a configuration example of the rainwater abnormality detection system 10B of the second embodiment. In the rainwater abnormality detection system 10 of the first embodiment (see FIG. 1), the imaging conditions of the imaging device 200 may change due to the influence (disturbance) of external factors such as sunlight, which is external light, and the captured rainwater image may change. Therefore, the rainwater abnormality detection system 10B of the second embodiment is configured to cover the space surrounding the path from the imaging device 200 to the imaging area IA with a black light-shielding device 230 composed of a light-shielding plate 232 and a light-shielding curtain 234, and to illuminate the light-shielded space area with an illumination device 240 such as an LED illumination, so that the illuminance of the flow path portion corresponding to the imaging area IA is constant. In addition, in order to clarify the color of the rainwater flowing in the flow path portion corresponding to the imaging area IA, for example, a white plate WB is arranged on the flow path bottom of the gutter flow path 22, so that the surface of the flow path bottom of the flow path portion corresponding to the imaging area IA is white.

The image acquisition unit 112 of the control unit 110 acquires a rainwater image of rainwater flowing through the white flow path portion where the white plate WB is placed as a white background image.

The trained model 134 of the control unit 110 is a trained model 134w for a white background that outputs information determining an abnormal state of rainwater in response to an input of a white background image, and the trained model file 134f stored in the memory unit 130 is a file 134wf constituting the trained model 134w for a white background.

The trained model 134w for the white background can be generated by the trained model generating device 300 (see FIG. 3) described above using a simulated rainwater image captured in an environment similar to the white channel portion corresponding to the imaging area IA of the above-mentioned gutter channel 22.

In the second embodiment, it is possible to improve the accuracy of detecting abnormalities in rainwater by suppressing the influence (disturbance) of external factors such as external light.

C. Third embodiment:
FIG. 7 is a schematic diagram showing a configuration example of the rainwater abnormality detection system 10C of the third embodiment. In the rainwater abnormality detection system 10B of the second embodiment (see FIG. 6), an abnormality in rainwater is detected by a white background image captured against the background of a white plate WB, so it is difficult to detect an abnormality in rainwater colored white, such as rainwater mixed with white paint. Therefore, the rainwater abnormality detection system 10C of the third embodiment is configured to include a shading device 230, a lighting device 240, and a white plate WB arranged at the bottom of the white channel of the side gutter channel 22, as in the rainwater abnormality detection system 10B of the second embodiment, and further includes a black colored plate BB arranged at the bottom of the colored channel of the side gutter channel 22 adjacent to the white plate WB. As a result, the rainwater image captured by the imaging device 200 includes both a white background image with the white plate WB arranged at the white channel part at the bottom of the white channel as the background, and a colored background image with the colored plate BB arranged at the colored channel part at the bottom of the colored channel as the background.

The control unit 110 of the rainwater anomaly detection device 100C also reads and executes the program 132C stored in the memory unit 130, and by reading the file 134wf of the trained model for the white background and the file 134bf of the trained model for the colored background, functions as a detection unit 114C that detects anomalies in rainwater and a detection result output unit 116 using the image acquisition unit 112C, the trained model for the white background 134w, and the trained model for the colored background 134b.

The image acquisition unit 112C acquires a rainwater image of rainwater in the imaging area IA from the imaging device 200, and extracts from the acquired rainwater image a white background image, which is a rainwater image with a white board WB as the background, and a colored background image, which is a rainwater image with a colored board BB as the background. The white background image is input to the trained model for white background 134w, and the colored background image is input to the trained model for colored background 134b. The trained model for white background 134w outputs an evaluation of the degree of abnormality of the rainwater for the input of the white background image. The trained model for colored background 134b outputs an evaluation of the degree of abnormality of the rainwater for the input of the colored background image. The detection unit 114C detects abnormalities in the rainwater based on the evaluation results output from the trained model for white background 134w and the evaluation results output from the trained model for colored background 134b. Specifically, if both evaluation results indicate that the rainwater condition is normal, the detection unit 114C detects that the rainwater condition is normal, and if at least one evaluation result indicates that the rainwater condition is abnormal, the detection unit 114C detects that the rainwater condition is abnormal.

The trained model for white background 134w and the trained model for colored background 134b can be generated by the trained model generating device 300 (see FIG. 3) described above using simulated rainwater images captured in an environment similar to the white channel portion and colored channel portion corresponding to the imaging area IA of the above-mentioned gutter channel 22, respectively.

In the third embodiment, abnormalities in rainwater colored in a white color, which would be difficult to detect in a white background image captured against a white plate WB and would be prone to overdetection, can be detected in a colored background image captured against a black colored plate BB, thereby improving detection accuracy.

In the above embodiment, the colored plate BB is described as black, but the present invention is not limited to this color. As long as it is possible to create an image that can detect rainwater colored in a color that is difficult to detect in a white image with the white plate WB as the background, the colored plate may be of another color, for example, an achromatic color such as gray that is less bright than white.

D. Other embodiments:
(D1) In the above embodiment, the case where the state of rainwater is detected as an abnormality is described as an example to evaluate the degree of abnormality of rainwater, but the present invention is not limited to this. The degree of abnormality of rainwater may be detected as an evaluation of the degree of abnormality of rainwater. In this case, the trained model may output the degree of abnormality of rainwater in response to the input of a rainwater image. In this case, the trained model may acquire teacher data by collecting multiple datasets including simulated rainwater images and information on the degree of abnormality as an evaluation result associated with the simulated rainwater images, and generate the acquired teacher data by training.

(D2) In the above embodiment, rainwater is used as an example of wastewater, but the present invention is not limited to this and can be applied to various types of wastewater discharged from facilities, for example, each treatment tank at a wastewater treatment plant. In the above embodiment, the color of mud mixed into rainwater is used as an example of a non-abnormal factor color that can be mistakenly recognized as an abnormality when mixed into wastewater, but the present invention is not limited to this and can be applied to various non-abnormal factor colors that can be mistakenly recognized as an abnormality when substances other than mud, such as proliferated algae, are mixed into wastewater, in the case of algae, and range from green to brown.

The present disclosure is not limited to the above-described embodiments, and can be realized in various configurations without departing from the spirit of the present disclosure. For example, the technical features of the embodiments corresponding to the technical features in each form described in the Summary of the Invention column can be replaced or combined as appropriate to solve some or all of the above-described problems or to achieve some or all of the above-described effects. Furthermore, if a technical feature is not described as essential in this specification, it can be deleted as appropriate.

10...Rainwater abnormality detection system, 10B...Rainwater abnormality detection system, 10C...Rainwater abnormality detection system, 20...Rainwater gutter, 22...Rainwater tank, 24...Inlet, 100...Rainwater abnormality detection device, 100C...Rainwater abnormality detection device, 110...Control unit, 112...Image acquisition unit, 112C...Image acquisition unit, 114...Detection unit, 114C...Detection unit, 116...Detection result output unit, 130...Memory unit, 132...Program, 132C...Program, 134...Trained model, 134f...Trained model file, 134b...Colored background trained model for, 134bf...file of trained model for colored background, 134w...trained model for white background, 134wf...file of trained model for white background, 200...imaging device, 230...shading device, 232...shading plate, 234...shading screen, 240...illumination device, 300...trained model generating device, 310...control unit, 312...teaching data acquisition unit, 314...learning unit, 330...storage unit, 332...program, 334...teaching data, RW...side drain channel, IA...imaging area, BB...colored plate, WB...white plate

Claims (7)

A drainage abnormality detection device that detects abnormalities in drainage from a facility, which appear as colors,
An image acquisition unit that acquires a drainage image, which is a captured image of drainage;
a detection unit that detects abnormalities in the drainage from the facility using a trained model that receives an input of a drainage image and outputs an evaluation of the degree of abnormality, and generates a learning result using correspondence information that corresponds a plurality of simulated drainage images including captured images of the simulated drainage obtained by mixing an abnormality factor color component that indicates an abnormality in the drainage with a non-abnormality factor color component that is a color that may be mistakenly recognized as an abnormality when mixed into the drainage, the simulated drainage images including captured images of the simulated drainage appearing in different colors due to different concentrations of the non-abnormality factor color components, and an evaluation of the degree of abnormality of each of the simulated drainage images; and
A drainage abnormality detection device comprising: The drainage abnormality detection device according to claim 1,
The image acquisition unit acquires, as the drainage image, a white background image obtained by imaging drainage flowing through a flow path having a white bottom under lighting that blocks external light.
Drainage abnormality detection device. The drainage abnormality detection device according to claim 1,
The image acquisition unit acquires, as the drainage image, a white background image obtained by imaging the drainage flowing through a flow path having a white bottom under lighting that blocks external light, and a colored background image obtained by imaging the drainage flowing through a flow path having a bottom colored in a color other than white;
The detection unit detects an abnormality in the wastewater from the facility by using, as the trained model, a trained model for a white background, which receives a white background image as an input and outputs an evaluation of the degree of abnormality, for the white background image acquired by the image acquisition unit, and by using, as the trained model, a trained model for a colored background, which receives a colored background image as an input and outputs an evaluation of the degree of abnormality, for the colored background image acquired by the image acquisition unit.
Drainage abnormality detection device. The drainage abnormality detection device according to any one of claims 1 to 3,
The drainage is rainwater,
The non-abnormal factor color is the color of mud mixed in the rainwater.
Drainage abnormality detection device. A trained model generation device that generates a trained model used to detect an abnormality in wastewater from a facility, the abnormality appearing as a color,
a teacher data acquisition unit that acquires, as teacher data, correspondence information that associates a plurality of simulated drainage images including captured images of the plurality of simulated drainages obtained by mixing an abnormality factor color component that indicates an abnormality in the drainage with a non-abnormality factor color component that is a color that may be mistakenly recognized as an abnormality when mixed into the drainage, each of which appears in a different color due to a different concentration of the non-abnormality factor color component, with an evaluation of the degree of abnormality of each of the simulated drainage images;
A learning unit that generates a trained model that receives a drainage image as an input and outputs an evaluation of the degree of abnormality by learning using the teacher data;
A trained model generation device comprising: A method for detecting an abnormality in drainage from a facility, the abnormality being manifested as a color, comprising:
A drainage image is obtained, which is an image of the drainage.
A plurality of simulated wastewater images including images of the simulated wastewater obtained by mixing an abnormality factor color component indicating an abnormality in the wastewater with a non-abnormality factor color component which is a color that may be mistakenly recognized as an abnormality if mixed into the wastewater, each of which appears in a different color due to a different concentration of the non-abnormality factor color component , are generated by learning using correspondence information that associates the simulated wastewater images with an evaluation of the degree of abnormality as training data, and an abnormality in the wastewater from the facility is detected using a trained model that takes the wastewater images as input and outputs an evaluation of the degree of abnormality.
A method for detecting drainage abnormalities. A trained model generation method for generating a trained model used to detect an abnormality in wastewater from a facility, which appears as a color, comprising:
A plurality of simulated drainage images including images of the simulated drainage obtained by mixing an abnormality factor color component indicating an abnormality in the drainage with a non-abnormality factor color component which is a color that may be mistakenly recognized as an abnormality when mixed into the drainage, each of which appears in a different color due to a different concentration of the non-abnormality factor color component , are obtained as teacher data, and correspondence information that corresponds the evaluation of the degree of abnormality of each of the simulated drainage images is obtained ;
A trained model is generated by training using the teacher data, the trained model inputting a drainage image and outputting an evaluation of the degree of abnormality.
How to generate trained models.

Images