JP7600696B2 - Machine learning device and anomaly identification device - Google Patents

Description

The present invention relates to a machine learning device and an anomaly identification device.

Inspections of the manholes housing the underground power transmission lines mainly involve measuring the amount of cable movement using measuring instruments, visually checking the deterioration of metal fittings, and visually checking for cracks in the manhole body and water leaks at the pipe opening.

When measuring the amount of cable movement, the measurement value may vary depending on the person who measures it; for example, the position where the cable bending radius is maximum is first predicted by visual inspection, and then the actual position where it is maximum is found by applying a measuring tool. In addition, it is necessary to control it to the nearest few millimeters, and there is a large margin of error depending on the person who measures it.

The deterioration of metal fittings is judged based on samples of rust color and corrosive organisms that have developed, but the judgment may vary depending on the inspector.

When it comes to cracks in the manhole body or leaks from the pipe opening, tiny cracks or small leaks can be overlooked with human visual inspection alone, and the problem may have progressed significantly by the time of the next inspection.

For example, Patent Document 1 discloses a technology for identifying cracks on the surface of concrete using machine learning. However, the technology described in Patent Document 1 cannot automatically identify cracks in the manhole body, or the rust or corrosion state of metal fittings installed inside the manhole.

JP 2020-038132 A

The present invention was made in consideration of the above problems, and aims to provide a machine learning device and anomaly identification device that automatically identifies the rust and corrosion state of metal objects installed inside manholes.

To achieve the above objective, the present invention has the following configuration.

(1) A machine learning device that constructs a learning model for identifying anomalies in a manhole, the machine learning device comprising: a first input data acquisition means that acquires an on-site photograph that includes equipment in the manhole as a subject as first input data; a first label acquisition means that acquires a judgment result of a rust/corrosion state of metal objects included in the on-site photograph of the manhole as a first label; and a first learning model construction means that performs supervised learning using a pair of the first input data and the first label as teacher data to construct a first learning model for judging the rust/corrosion state of the metal objects included in a new on-site photograph that includes the equipment in the manhole as a subject.

(1) makes it possible to provide a machine learning device that can automatically identify the rust and corrosion state of metal fittings installed inside manholes.

(2) It is preferable that the machine learning device described in (1) further comprises a second input data acquisition means for acquiring an on-site photo including the equipment inside the manhole as a subject as second input data, a second label acquisition means for acquiring as a second label a determination result of a location where a crack has occurred in a wall included in the on-site photo including the equipment inside the manhole as a subject, and a second learning model construction means for constructing a second learning model for determining a location where a crack has occurred in a wall included in a new on-site photo including the equipment inside the manhole as a subject by performing supervised learning using a pair of the second input data and the second label as teacher data.

(2) makes it possible to provide a machine learning device that can automatically identify cracks in manhole structures and water leaks at pipe openings.

(3) In the machine learning device described in (1) or (2), it is preferable to further include a third input data acquisition means for acquiring first on-site condition data including the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal to be installed in the manhole, and information related to the material of the metal as third input data, a third label acquisition means for acquiring the corrosion level of the metal a predetermined period after the date and time indicated by the third input data as a third label, and a third learning model construction means for constructing a third learning model for predicting the corrosion level of the metal a predetermined period after the date and time indicated by the new first on-site condition data by performing supervised learning using a pair of the third input data and the third label as teaching data.

(3) makes it possible to provide a machine learning device that automatically predicts the corrosion level of metal fittings installed inside manholes.

(4) In the machine learning device described in (1) to (3), it is preferable to further include a fourth input data acquisition means for acquiring second on-site situation data including the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, information related to the material of the metal fittings, and an on-site photo including the equipment in the manhole as a subject as fourth input data, a fourth label acquisition means for acquiring the on-site photo after a predetermined period from the date and time indicated by the fourth input data as a fourth label, and a fourth learning model construction means for constructing a fourth learning model for generating a predicted photo including the equipment in the manhole as a subject after the predetermined period from the date and time indicated by the new second on-site situation data by performing supervised learning using a pair of the fourth input data and the fourth label as teacher data.

According to (4), it is possible to provide a machine learning device that automatically generates predicted photos that include equipment inside a manhole as a subject.

(5) An anomaly identification device using the first learning model constructed by the machine learning device according to any one of (1) to (4), comprising: an on-site photo acquisition means for acquiring a new on-site photo including the equipment inside the manhole as a subject; and a rust/corrosion state determination means for determining the rust/corrosion state of the metal fittings based on the on-site photo and the first learning model.

(5) makes it possible to provide an anomaly identification device that automatically determines the rust and corrosion state of metal objects installed inside manholes.

(6) In the anomaly identification device described in (5), it is preferable that the anomaly identification device further uses the second learning model constructed by the machine learning device described in (2), and further comprises an on-site photo acquisition means for acquiring a new on-site photo including the equipment inside the manhole as a subject, and a crack determination means for determining a location of a crack in a wall included in the new on-site photo including the equipment inside the manhole as a subject, based on the on-site photo and the second learning model.

(6) It is possible to provide an abnormality identification device that automatically detects cracks in a manhole body or water leakage from a pipe opening.

(7) In the anomaly identification device described in (6), it is preferable to further include a cable anomaly determination means for determining an anomaly in a cable installed in the manhole from an on-site photograph that includes the equipment inside the manhole as a subject.

(7) It is possible to provide an abnormality identification device that automatically determines the amount of movement of a cable inside a manhole body.

(8) In the abnormality identification device described in (7), it is preferable to identify an abnormality in the manhole based on at least one of the rust/corrosion state of the metal, the location of the crack, and the amount of movement of the cable.

(8) makes it possible to provide an anomaly identification device that automatically identifies abnormalities within a manhole structure.

(9) In the anomaly identification device described in (5) to (8), it is preferable that the anomaly identification device further uses the third learning model constructed by the machine learning device described in (3), and further comprises a first on-site condition data acquisition means for acquiring new first on-site condition data of the manhole, and a corrosion level prediction means for predicting the corrosion level of the metal fittings after the specified period from the date and time indicated by the new first on-site condition data based on the first on-site condition data and the third learning model.

(9) makes it possible to provide an anomaly identification device that automatically predicts the corrosion level of metal fittings installed inside manholes.

(10) In the anomaly identification device described in any one of (5) to (9), the anomaly identification device further uses the fourth learning model constructed by the machine learning device described in (4), and preferably further comprises a second on-site condition data acquisition means for acquiring new second on-site condition data for the manhole, and a predicted photo generation means for generating a predicted photo including, as a subject, the equipment in the manhole at a time the predetermined period after the date and time indicated by the new second on-site condition data, based on the second on-site condition data and the fourth learning model.

According to (10), it is possible to provide an anomaly identification device that automatically generates a predicted photograph that includes the equipment inside a manhole as a subject.

The present invention makes it possible to automatically identify the rust and corrosion state of metal fittings installed inside manholes.

1 is a diagram showing an overall configuration of an anomaly identification system according to an embodiment of the present invention; FIG. 2 is a functional block diagram of a first learning unit included in the machine learning device according to the embodiment of the present invention. 1 is a functional block diagram of an anomaly identification device according to an embodiment of the present invention. 4 is a flowchart showing an operation of the machine learning device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the machine learning device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the machine learning device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the machine learning device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the anomaly identification device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the anomaly identification device according to the embodiment of the present invention. 4 is a flowchart showing an operation of the anomaly identification device according to the embodiment of the present invention.

The following describes an embodiment of the present invention with reference to Figures 1 to 10.

[Configuration of 1 embodiment]
First, a configuration of an anomaly identification system 100 according to the present embodiment will be described. Fig. 1 is a block diagram showing the anomaly identification system according to the present embodiment. As shown in Fig. 1, the anomaly identification system 100 includes a machine learning device 10 and an anomaly identification device 20.

Here, the machine learning device 10 and the anomaly identification device 20 are paired one-to-one and connected to be able to communicate with each other. Although not shown in FIG. 1, the machine learning device 10 and the anomaly identification device 20 may be connected to each other via a network. The network may be, for example, a local area network (LAN), the Internet, a public telephone network, or a combination of these. There are no particular limitations on the specific communication method in the network, or whether the connection is wired or wireless. Alternatively, the machine learning device 10 and the anomaly identification device 20 may be directly connected via a connection unit, rather than communicating via a network.

The machine learning device 10 uses supervised learning to construct a learning model to be used by the anomaly identification device 20. To this end, the machine learning device 10 includes a first learning unit 11 to a fourth learning unit 14, as shown in FIG. 1.

The first learning unit 11 constructs a first learning model for determining the rust/corrosion state of metal objects contained in new site photos that include equipment inside a manhole as a subject. Note that this "equipment inside a manhole" also includes the inner wall of the manhole.

Figure 2 is a functional block diagram of the first learning unit 11. As shown in Figure 2, the first learning unit 11 includes a first input data acquisition unit 111, a first label acquisition unit 112, a first learning model construction unit 113, and a first learning model storage unit 114.

The first input data acquisition unit 111 acquires an on-site photograph that includes the equipment inside the manhole as a subject as the first input data. Here, it is preferable that the on-site photograph that includes the equipment inside the manhole as a subject includes the junction box, cable alignment, pipe entrance, and metal fittings installed inside the manhole.

The first label acquisition unit 112 acquires, as the first label, the judgment result of the rust/corrosion state of metal objects contained in an on-site photograph that includes the equipment inside the manhole as a subject. Note that this judgment result may be, for example, the rust/corrosion state of metal objects contained in the on-site photograph that includes the equipment inside the manhole as a subject, obtained by a skilled worker visually inspecting the on-site photograph that is used as the first input data.

Here, the corrosion state of the metal may include a corrosion level that indicates the degree of corrosion of the metal. For example, the corrosion level may be one of four levels, Level A to Level D, as shown below.

Level A is a level where crevice corrosion has occurred, causing significant damage to the cables inside the manhole. More specifically, this is a level that requires immediate repairs because within one year, there will be violations of technical standards for electrical equipment, insufficient design strength, and commercial operation will be impossible.

Level B is a level where damage to the cable is expected due to the occurrence of crevice corrosion. More specifically, repairs are necessary, detailed investigation is required, and the condition is likely to progress to Level A within one year.

At Level C, pitting and corrosion have occurred on the surface, and there is a possibility that the cable may be damaged before the next inspection. More specifically, compared to Level B, there are fewer products of corrosion, and it is thought that the cable will progress to Level A in a few years, so continued monitoring is required.

Level D is a condition where no pitting or corrosion has occurred, or where pitting or corrosion has occurred on the surface but there is no expected damage by the time of the next inspection, and so observation is required.

The first learning model construction unit 113 performs supervised learning using a pair of the first input data and the first label as training data to construct a first learning model for determining the rust/corrosion state of metal objects contained in site photos that include equipment inside a manhole as a subject, and transmits the constructed first learning model to the anomaly identification device 20.

The first learning model construction unit 113 can be realized, for example, using a support vector machine (hereinafter also referred to as SVM).

In this case, the first learning model construction unit 113 uses a binarized label indicating whether or not the data corresponds to a specific corrosion level as the first label, and calculates a hyperplane that separates the space including the first input data so as to maximize the margin as to whether or not the data corresponds to the specific corrosion level. Furthermore, the first learning model construction unit 113 can use the coefficients of this hyperplane as parameters of a learning model that the anomaly identification device 20 described below uses to determine the corrosion level.

The first learning model memory unit 114 stores the learning model constructed by the first learning model construction unit 113.

The second learning unit 12 constructs a second learning model for determining the location of a crack in a wall contained in a new site photo that includes equipment inside a manhole as a subject.
The second learning unit 12 includes a second input data acquisition unit 121, a second label acquisition unit 122, a second learning model construction unit 123, and a second learning model storage unit 124. Note that the configuration of the second learning unit 12 is similar to the configuration of the first learning unit 11 shown in FIG. 2, and therefore is not shown in the figure.

The second input data acquisition unit 121 acquires site photos that include the equipment inside the manhole as a subject as second input data.

The second label acquisition unit 122 acquires, as the second label, the judgment result of the location of the crack in the wall included in the site photograph that includes the equipment inside the manhole as a subject. Note that this judgment result may be, for example, the location of the crack included in the site photograph that includes the equipment inside the manhole as a subject, obtained by a skilled worker visually inspecting the site photograph that is used as the second input data.

The "location where the crack has occurred" may include not only the location where the crack has occurred, but also the location where water is leaking from the crack.

The second learning model construction unit 123 performs supervised learning using the pair of the second input data and the second label as training data to construct a second learning model for determining the location of a crack in a wall contained in a new site photo that includes the equipment inside the manhole as a subject, and transmits the constructed second learning model to the anomaly identification device 20.

The second learning model memory unit 124 stores the learning model constructed by the second learning model construction unit 123.

The third learning unit 13 constructs a third learning model for predicting the corrosion level of metal fittings a predetermined period of time after the date and time indicated by the new first on-site condition data.
The third learning unit 13 includes a third input data acquisition unit 131, a third label acquisition unit 132, a third learning model construction unit 133, and a third learning model storage unit 134. Note that the configuration of the third learning unit 13 is similar to the configuration of the first learning unit 11 shown in FIG. 2, and therefore is not shown in the figure.

The third input data acquisition unit 131 acquires the first local situation data, which includes the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, and information related to the material of the metal fittings, as the third input data.

The third label acquisition unit 132 acquires the corrosion level of the metal a predetermined period after the date and time indicated by the third input data as the third label.

For example, based on historical data consisting of the water level in the manhole at a specified date and time, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, information related to the material of the metal fittings, and the corrosion level of the metal fittings at the specified date and time, the third input data acquisition unit 131 may acquire the water level in the manhole at a first date and time, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, and information related to the material of the metal fittings as the third input data, and the third label acquisition unit 132 may acquire the corrosion level of the metal fittings at a second date and time after a specified period of time has elapsed in the same manhole as the third label.

The third learning model construction unit 133 performs supervised learning using the pair of the third input data and the third label as teacher data to construct a third learning model for predicting the corrosion level of metal a specified period of time after the date and time indicated by the new first on-site condition data, and transmits the constructed third learning model to the anomaly identification device 20.
In addition, the third learning model construction unit 133 may set multiple types of "specified periods" as this "specified period" and construct multiple learning models corresponding to each of these multiple types of "specified periods."

The third learning model memory unit 134 stores the learning model constructed by the third learning model construction unit 133.

The fourth learning unit 14 constructs a fourth learning model for predicting site photos a predetermined period after the date and time indicated by the new second site situation data.
The fourth learning unit 14 includes a fourth input data acquisition unit 141, a fourth label acquisition unit 142, a fourth learning model construction unit 143, and a fourth learning model storage unit 144. Note that the configuration of the fourth learning unit 14 is similar to the configuration of the first learning unit 11 shown in FIG. 2, and therefore is not shown in the figure.

The fourth input data acquisition unit 141 acquires, as fourth input data, second on-site condition data including the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, information related to the material of the metal fittings, and on-site photographs that include the equipment in the manhole as a subject.

The fourth label acquisition unit 142 acquires, as the fourth label, a site photo that includes the equipment inside the manhole as a subject, the photograph taken a predetermined period after the date and time indicated by the fourth input data.

For example, based on history data consisting of the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, information related to the material of the metal fittings, and site photos containing the equipment in the manhole as a subject at a specified date and time, the fourth input data acquisition unit 141 may acquire the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, and site photos containing the equipment in the manhole as a subject at a first date and time as the fourth input data, and the fourth label acquisition unit 142 may acquire the site photo of the same manhole at a second date and time after a specified period of time has passed as the fourth label.

It is preferable that the site photo containing the equipment inside the manhole as a subject, which is included in the fourth input data, and the site photo of the manhole after a predetermined period of time, which is included in the fourth label, are taken at the same angle. Alternatively, the site photo containing the equipment inside the manhole after a predetermined period of time may be corrected to be taken at the same angle as the site photo containing the equipment inside the manhole as a subject, which is included in the fourth input data, and then used as the fourth label.

The fourth learning model construction unit 143 performs supervised learning using the pair of the fourth input data and the fourth label as teacher data to construct a fourth learning model for predicting site photos a specified period after the date and time indicated by the new second on-site situation data, and transmits the constructed fourth learning model to the anomaly identification device 20.
In addition, the fourth learning model construction unit 143 may set multiple types of "specified periods" as this "specified period" and construct multiple learning models corresponding to each of these multiple types of "specified periods."

The fourth learning model memory unit 144 stores the learning model constructed by the fourth learning model construction unit 143.

Figure 3 is a functional block diagram of the anomaly identification device 20. The anomaly identification device 20 includes a control unit 21, a memory unit 22, and a display unit 23.

The control unit 21 is a part that controls the entire anomaly identification device 20, and realizes various functions in this embodiment by appropriately reading and executing various programs from a storage area such as a ROM, RAM, flash memory, or hard disk (HDD). The control unit 21 may be a CPU. The control unit 21 includes an on-site photo acquisition unit 211, a first on-site condition data acquisition unit 212, a second on-site condition data acquisition unit 213, a rust/corrosion determination unit 214, a crack determination unit 215, a cable anomaly determination unit 216, an anomaly identification unit 217, a corrosion level prediction unit 218, and a predicted photo generation unit 219.

The site photo acquisition unit 211 acquires new site photos that include the equipment in the manhole as a subject, separately from site photos that include the equipment in the manhole as a subject, which are included in the training data used when machine learning was performed by the machine learning device 10. For example, a worker may operate an input device such as a keyboard or mouse provided on the anomaly identification device 20 to load a new site photo that includes the equipment in the manhole as a subject into the anomaly identification device 20, and then the site photo acquisition unit 211 may acquire the new site photo that has been loaded.

The first local situation data acquisition unit 212 acquires new first local situation data for the manhole separately from the first local situation data included in the teacher data used when machine learning was performed by the machine learning device 10. For example, a worker may operate an input device such as a keyboard or mouse provided on the anomaly identification device 20 to load new first local situation data for the manhole into the anomaly identification device 20, and then the first local situation data acquisition unit 212 may acquire the new first local situation data that has been loaded.

The second local situation data acquisition unit 213 acquires new second local situation data for the manhole separately from the second local situation data included in the teacher data used when machine learning was performed by the machine learning device 10. For example, a worker may operate an input device such as a keyboard or mouse provided on the anomaly identification device 20 to load new second local situation data for the manhole into the anomaly identification device 20, and then the second local situation data acquisition unit 213 may acquire the new second local situation data that has been loaded.

The rust/corrosion determination unit 214 determines the rust/corrosion state of the metal fittings in the manhole based on the site photos acquired by the site photo acquisition unit 211 and the first learning model.

The crack determination unit 215 determines the location of a crack in the wall included in the new site photo of the manhole based on the site photo acquired by the site photo acquisition unit 211 and the second learning model.

The cable abnormality determination unit 216 determines an abnormality in the cable installed in the manhole from a site photo that includes the equipment inside the manhole as a subject and is acquired by the site photo acquisition unit 211.

For example, the allowable bending radius of the cable is determined based on the type of cable, and if the on-site photo acquired by the on-site photo acquisition unit 211 is a 3D photo, the difference between the survey data calculated using the 3D photo and the allowable bending radius of the cable is calculated, and if the difference is within a specified value, the cable is determined to be normal, and if it exceeds the specified value, the cable is determined to be abnormal.

It is preferable for the cable abnormality determination unit 216 to determine cable abnormalities from on-site photos taken at the current time, particularly at the time of abnormality determination by the abnormality identification unit 217 described below.

The abnormality identification unit 217 identifies abnormalities inside the manhole based on the rust/corrosion state of the metal fittings determined by the rust/corrosion determination unit 214, the location of the crack determined by the crack determination unit 215, and the abnormality determination related to the cable determined by the cable abnormality determination unit 216.
In particular, with regard to the rust/corroded condition of metal fittings, if there is a location where rust or corrosion has occurred and the cable is in contact with the cable and damage to the cable is anticipated, the anomaly identification unit 217 will identify the location of rust or corrosion as an abnormality inside the manhole.

The corrosion level prediction unit 218 predicts the corrosion level of metal a predetermined period of time after the date and time indicated by the first on-site condition data based on the first on-site condition data acquired by the first on-site condition data acquisition unit 212 and the third learning model. This "predetermined period" corresponds to the time lag between the date and time indicated by the third input data and the date and time indicated by the third label in the teacher data used when generating the third learning model.

The predicted photo generation unit 219 generates a predicted photo including, as a subject, the equipment inside the manhole a predetermined period after the date and time indicated by the second local situation data acquired by the second local situation data acquisition unit 213, based on the fourth learning model and the second local situation data acquired by the second local situation data acquisition unit 213. This "predetermined period" corresponds to the time lag between the date and time indicated by the fourth input data and the date and time indicated by the fourth label in the training data used when generating the fourth learning model.

The memory unit 22 stores the first to fourth learning models acquired from the machine learning device 10. The memory unit 22 may also store an on-site photo including the equipment in the manhole as a subject acquired by the on-site photo acquisition unit 211, the first on-site situation data acquired by the first on-site situation data acquisition unit 212, the second on-site situation data acquired by the second on-site situation data acquisition unit 213, the judgment results by the rust/corrosion judgment unit 214, the crack judgment unit 215, and the cable abnormality judgment unit 216, the identification result by the abnormality identification unit 217, the prediction result by the corrosion level prediction unit 218, and the predicted photo generated by the predicted photo generation unit 219.

The display unit 23 is a monitor that displays the judgment results from the rust/corrosion judgment unit 214, the crack judgment unit 215, and the cable abnormality judgment unit 216, the identification results from the abnormality identification unit 217, the prediction results from the corrosion level prediction unit 218, and the predicted photo generated by the predicted photo generation unit 219.

2. Operation of the embodiment
Next, a description will be given of the operation of the anomaly identification system 100 according to this embodiment. First, the operation of the machine learning device 10 will be described with reference to FIGS.

2.1 Operation of the Machine Learning Device 10
FIG. 4 is a flowchart showing the operation of the first learning unit 11.
In step S11, the first input data acquisition unit 111 acquires a site photo including the equipment inside the manhole as a subject as the first input data.

In step S12, the first label acquisition unit 112 acquires, as the first label, the judgment result of the rust/corrosion state of metal objects contained in the site photo that includes the equipment inside the manhole as a subject.

In step S13, the first learning model construction unit 113 sets the pair of the first input data and the first label as teacher data.

In step S14, the first learning model construction unit 113 performs supervised learning using the teacher data from step S13.

If machine learning has ended in step S15 (S15: YES), the process proceeds to step S16. If machine learning has not yet ended (S15: NO), the process proceeds to step S11.

In step S16, the first learning model construction unit 113 transmits the constructed first learning model to the anomaly identification device 20. After that, all processing is terminated.

FIG. 5 is a flowchart showing the operation of the second learning unit 12.
In step S21, the second input data acquisition unit 121 acquires, as second input data, a site photo that includes the equipment inside the manhole as a subject.

In step S22, the second label acquisition unit 122 acquires, as a second label, the determination result of the location of the crack in the wall included in the site photo that includes the equipment inside the manhole as a subject.

In step S23, the second learning model construction unit 123 sets the pair of the second input data and the second label as teacher data.

In step S24, the second learning model construction unit 123 performs supervised learning using the teacher data from step S23.

If machine learning has ended in step S25 (S25: YES), the process proceeds to step S26. If machine learning has not yet ended (S25: NO), the process proceeds to step S21.

In step S26, the second learning model construction unit 123 transmits the constructed second learning model to the anomaly identification device 20. After that, all processing is terminated.

FIG. 6 is a flowchart showing the operation of the third learning unit 13.
In step S31, the third input data acquisition unit 131 acquires, as the first on-site condition data for the manhole, the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings to be installed in the manhole, and information related to the material of the metal fittings as third input data.

In step S32, the third label acquisition unit 132 acquires the corrosion level of the metal a predetermined period after the date and time indicated by the third input data as the third label.

In step S33, the third learning model construction unit 133 sets the pair of the third input data and the third label as teacher data.

In step S34, the third learning model construction unit 133 performs supervised learning using the teacher data from step S33.

If machine learning has ended in step S35 (S35: YES), the process proceeds to step S36. If machine learning has not yet ended (S35: NO), the process proceeds to step S31.

In step S36, the third learning model construction unit 133 transmits the constructed third learning model to the anomaly identification device 20. Then, all processing is terminated.

FIG. 7 is a flowchart showing the operation of the fourth learning unit 14.
In step S41, the fourth input data acquisition unit 141 acquires, as fourth input data, second on-site condition data including the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings to be installed in the manhole, information related to the material of the metal fittings, and on-site photographs that include the equipment in the manhole as subjects.

In step S42, the fourth label acquisition unit 142 acquires, as the fourth label, a site photo that includes the equipment inside the manhole as a subject, the subject being a predetermined period of time after the date and time indicated by the fourth input data.

In step S43, the fourth learning model construction unit 143 sets the pair of the fourth input data and the fourth label as teacher data.

In step S44, the fourth learning model construction unit 143 performs supervised learning using the teacher data from step S43.

If machine learning has ended in step S45 (S45: YES), the process proceeds to step S46. If machine learning has not yet ended (S45: NO), the process proceeds to step S41.

In step S46, the fourth learning model construction unit 143 transmits the constructed fourth learning model to the anomaly identification device 20. Then, all processing is terminated.

2.2 Operation of the anomaly identification device 20
Next, the operation of the anomaly identification device 20 will be described with reference to FIGS.

First, the anomaly identification method using the anomaly identification device 20 will be described with reference to FIG.

In step S51, the site photo acquisition unit 211 acquires new site photos.

In step S52, the rust/corrosion determination unit 214 determines the rust/corrosion state of the metal fittings in the manhole based on the on-site photos acquired by the on-site photo acquisition unit 211 and the first learning model.

In step S53, the crack determination unit 215 determines the location of a crack in a wall included in a new site photo that includes the equipment inside the manhole as a subject, based on the site photo acquired by the site photo acquisition unit 211 and the second learning model.

In step S54, the cable abnormality determination unit 216 determines whether there is an abnormality in the cable from the current on-site photo.

In step S55, the anomaly identification unit 217 identifies an anomaly in the manhole based on the rust/corrosion state of the metal parts determined by the rust/corrosion determination unit 214, the location of the crack determined by the crack determination unit 215, and the cable anomaly determined by the cable anomaly determination unit 216. Then, all processing is terminated.

Next, the corrosion level prediction method using the anomaly identification device 20 will be explained with reference to FIG. 9.

In step S61, the first local situation data acquisition unit 212 acquires new first local situation data for the manhole separately from the first local situation data contained in the teacher data used when machine learning was performed by the machine learning device 10.

In step S62, the corrosion level prediction unit 218 predicts the corrosion level of the metal fittings a predetermined period of time after the date and time indicated by the first local situation data, based on the first local situation data acquired by the first local situation data acquisition unit 212 and the third learning model. Then, all processing is terminated.

Next, the method for generating a predicted photo using the anomaly identification device 20 will be described with reference to FIG. 10.

In step S71, the second local situation data acquisition unit 213 acquires new second local situation data for the manhole separately from the second local situation data contained in the teacher data used when machine learning was performed by the machine learning device 10.

In step S72, the predicted photo generation unit 219 generates a predicted photo including, as a subject, the equipment inside the manhole a predetermined period after the date and time indicated by the second local situation data, based on the second local situation data acquired by the second local situation data acquisition unit 213 and the fourth learning model. Then, all processing is terminated.

[3 Modifications]
Although the embodiment of the present invention has been described above, the present invention is not limited to the above-described embodiment. Furthermore, the effects described in the present embodiment are merely a list of the most preferable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the present embodiment.

3.1 Modification 1
In the above embodiment, the machine learning device 10 and the anomaly identification device 20 are shown as separate entities, but the present invention is not limited to this. For example, the machine learning device 10 may be incorporated into the housing of the anomaly identification device 20, thereby realizing the anomaly identification system 100 in an integrated manner.

3.2 Modification 2
In the above embodiment, the first input data acquisition unit 111 of the first learning unit 11 acquires a site photo including the equipment in the manhole as a subject as the first input data, and the second input data acquisition unit 121 of the second learning unit 12 acquires a site photo including the equipment in the manhole as a subject as the second input data, but is not limited to this. For example, the first input data acquisition unit 111 and the second input data acquisition unit 121 may acquire, as the first input data or the second input data, one or more of manhole interior data (dimensions of the manhole, distance from the wall surface of the junction box, position of the pipe entrance), dimensions of the junction box, and dimensions of the metal fittings, in addition to the site photo including the equipment in the manhole as a subject.

The anomaly identification method using the anomaly identification system 100 is realized by software. When realized by software, the programs that make up this software are installed on a computer. These programs may also be recorded on removable media and distributed to users, or may be distributed by being downloaded to the user's computer via a network.

LIST OF SYMBOLS 10 Machine learning device 11 First learning unit 12 Second learning unit 13 Third learning unit 14 Fourth learning unit 20 Anomaly identification device 21 Control unit 22 Memory unit 23 Display unit 111 First input data acquisition unit 112 First label acquisition unit 113 First learning model construction unit 114 First learning model storage unit 121 Second input data acquisition unit 122 Second label acquisition unit 123 Second learning model construction unit 124 Second learning model storage unit 131 Third input data acquisition unit 132 Third label acquisition unit 133 Third learning model construction unit 134 Third learning model storage unit 141 Fourth input data acquisition unit 142 Fourth label acquisition unit 143 Fourth learning model construction unit 144 Fourth learning model storage unit 211 Site photo acquisition unit 212 First on-site situation data acquisition unit 213 Second on-site situation data acquisition unit 214 Rust/corrosion determination unit 215 Crack determination unit 216 Cable abnormality determination unit 217 Abnormality identification unit 218 Corrosion level prediction unit 219 Prediction photo generation unit

Claims (9)

A machine learning device that constructs a learning model for identifying anomalies in a manhole,
A first input data acquisition means for acquiring a site photograph including the facility inside the manhole as a subject as first input data;
a first label acquisition means for acquiring, as a first label, a judgment result of a rust/corrosion state of a metal object included in an on-site photograph including the facility inside the manhole as a subject;
a first learning model construction means for constructing a first learning model for determining a rust/corrosion state of the metal object included in a new on-site photo that includes the equipment inside the manhole as a subject by performing supervised learning using a pair of the first input data and the first label as training data;
a third input data acquisition means for acquiring, as third input data, first on-site status data including information on the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, and the material of the metal fittings;
a third label acquisition means for acquiring a corrosion level of the metal object a predetermined period after the date and time indicated by the third input data as a third label;
a third learning model construction means for constructing a third learning model for predicting the corrosion level of the metal object after the predetermined period from the date and time indicated by the new first on-site condition data by performing supervised learning using a pair of the third input data and the third label as training data;
A machine learning device comprising: A machine learning device that constructs a learning model for identifying anomalies in a manhole,
A first input data acquisition means for acquiring a site photograph including the facility inside the manhole as a subject as first input data;
a first label acquisition means for acquiring, as a first label, a judgment result of a rust/corrosion state of a metal object included in an on-site photograph including the facility inside the manhole as a subject;
a first learning model construction means for constructing a first learning model for determining a rust/corrosion state of the metal object included in a new on-site photo that includes the equipment inside the manhole as a subject by performing supervised learning using a pair of the first input data and the first label as training data;
a fourth input data acquisition means for acquiring, as fourth input data, second on-site status data including the water level in the manhole, the pH and hydrogen sulfide concentration of the water stored in the manhole, the installation date and time of the metal fittings installed in the manhole, information related to the material of the metal fittings, and on-site photographs including the equipment in the manhole as subjects;
a fourth label acquisition means for acquiring the on-site photo taken a predetermined period after the date and time indicated by the fourth input data as a fourth label;
a fourth learning model construction means for constructing a fourth learning model for generating a predicted photograph including, as a subject, equipment in the manhole, the predetermined period after the date and time indicated by the new second on-site situation data, by performing supervised learning using a pair of the fourth input data and the fourth label as training data;
A machine learning device comprising: A second input data acquisition means for acquiring a site photo including the equipment inside the manhole as a subject as second input data;
a second label acquisition means for acquiring, as a second label, a determination result of a location where a crack has occurred in a wall included in a site photograph including the facility inside the manhole as a subject;
a second learning model construction means for constructing a second learning model for determining a location of a crack in a wall included in a new on-site photo including the facility inside the manhole as a subject by performing supervised learning using a pair of the second input data and the second label as teacher data;
The machine learning device according to claim 1 or 2 , further comprising: An anomaly identification device using the first learning model constructed by the machine learning device according to any one of claims 1 to 3 ,
An on-site photo acquisition means for acquiring a new on-site photo including the equipment inside the manhole as a subject;
A rust/corrosion state determination means for determining a rust/corrosion state of the metal object based on the on-site photograph and the first learning model;
An anomaly identification device comprising: An anomaly identification device further using the second learning model constructed by the machine learning device according to claim 3 ,
An on-site photo acquisition means for acquiring a new on-site photo including the equipment inside the manhole as a subject;
A crack determination means for determining a location of a crack in a wall included in a new on-site photo including the equipment inside the manhole as a subject based on the on-site photo and the second learning model;
The anomaly identification device according to claim 4 , further comprising: The abnormality identification device according to claim 4 or claim 5 , further comprising a cable abnormality determination means for determining an abnormality in a cable provided in the manhole from an on-site photograph that includes the equipment in the manhole as a subject. The present invention further includes a cable abnormality determination means for determining an abnormality in a cable provided in the manhole from an on-site photograph including the equipment in the manhole as a subject,
6. The abnormality identification device according to claim 5 , which identifies an abnormality in the manhole based on at least one of the rust/corrosion state of the metal fittings, the location of the crack, and the amount of movement of the cable. An anomaly identification device further using the third learning model constructed by the machine learning device according to claim 1 ,
A first on-site condition data acquisition means for acquiring new first on-site condition data of the manhole;
a corrosion level prediction means for predicting a corrosion level of the metal object after a predetermined period from a date and time indicated by new first on-site condition data based on the first on-site condition data and the third learning model;
The anomaly identification device according to claim 4 , further comprising: An anomaly identification device further using the fourth learning model constructed by the machine learning device according to claim 2 ,
A second on-site situation data acquisition means for acquiring new second on-site situation data of the manhole; and a predicted photo generation means for generating a predicted photo including, as a subject, equipment in the manhole after the predetermined period from the date and time indicated by the new second on-site situation data, based on the second on-site situation data and the fourth learning model.
The anomaly identification device according to claim 4 , further comprising:

Images