Training device

Abstract

A training device (10) according to the present disclosure comprises: a training data creation unit (13) that creates training data in which a training image obtained by capturing an image of a corrosion region inside a metal pipeline is associated with a measured value of the thickness reduction amount of the metal pipeline in the corrosion region; and a training unit (14) that calculates a loss value using a predetermined loss function on the basis of an error between an estimated value of the thickness reduction amount in the corrosion region of the training image, estimated by an estimator that estimates the thickness reduction amount in the corrosion region inside the metal pipeline, and a measured value of the thickness reduction amount corresponding to the training image, and creates an estimator through training based on the calculated loss value. When the error between the estimated value and the measured value is less than a predetermined threshold value, the predetermined loss function yields the loss value that is calculated with the error treated as zero.

Description

Learning Device

[0001] The present disclosure relates to a learning device.

[0002] FIG. 13 is a cross-sectional view of a metal pipeline, such as a steel pipeline, viewed from the axial direction. As shown in FIG. 13 , as corrosion C progresses in the thickness direction of a metal pipeline, the wall thickness P1 of the metal material in the corroded area where corrosion C has occurred decreases (wall thinning) compared to the wall thickness P2 of the metal material in the area where corrosion has not occurred. Wall thinning of a metal pipeline can lead to a decrease in structural strength and the possibility of holes being created, allowing sediment to be drawn into the metal pipeline. Therefore, for the maintenance and management of facilities, it is important to identify wall thinning of metal pipelines through inspections and repair dangerous areas in advance. Non-Patent Document 1 describes a method for measuring the amount of wall thinning of metal pipelines due to corrosion using an ultrasonic measuring device or the like.

[0003] One possible method for inspecting metal pipelines is to estimate the amount of wall thinning in metal pipelines from images of the inside of the metal pipeline taken with a pipe camera inserted into the metal pipeline.In this method, an estimator that estimates the amount of wall thinning from images is created using machine learning, and the amount of wall thinning in the metal pipeline can be estimated by inputting images of the inside of the metal pipeline into the estimator.

[0004] In training an estimator (see, for example, Non-Patent Document 2) that outputs a numerical value such as the amount of wall thinning as an estimated value, a value (loss value) output according to a loss function called MAPE (Mean Absolute Percentage Error) is used as feedback. As shown in Fig. 14, MAPE is a linear function that outputs a loss value proportional to the amount of error between the estimated value of the estimator and the correct value.

[0005] Jun Murakoshi et al., "Study on the applicability of various measurement techniques to measuring the remaining thickness of corroded steel members," Journal of Structural Engineering, Vol. 59A, pp. 711-724; J.A. Park et al., "Learning-based image scale estimation using surface textures for quantitative visual inspection of regions-of-interest Comput Aided Civ Inf. 2021; 36: 227-241"

[0006] When inputting an image into an estimator to estimate a length scale such as the amount of wall thinning, there is a limit to the length equivalent to one pixel of the input image (pixel resolution), so it is not appropriate to use a minute value smaller than the pixel resolution as the loss value. Here, for example, assume that the length equivalent to one pixel (error resolution) is 0.1 mm. In this case, as shown in FIG. 14 , because MAPE is a linear function, if the error between the estimated value of the estimator and the correct value is in a range smaller than the image resolution (0.1 mm in the example shown in FIG. 14 ), the error may not be a valid value. As a result, the output loss value will also be invalid. Using such an invalid loss value for training will result in a decrease in the estimation accuracy of the estimator.

[0007] The purpose of the present disclosure, made in consideration of the above-mentioned problems, is to provide a learning device that can improve the accuracy of an estimator that estimates the amount of thinning of a metal pipeline from an image taken of the inside of the metal pipeline.

[0008] A learning device according to one embodiment is a learning device that creates an estimator that estimates the amount of metal thinning in a corroded region inside a metal pipeline, and includes: a learning data creation unit that creates learning data that associates training images of corroded regions inside a metal pipeline with measured values ​​of the amount of metal thinning of the metal pipeline in the corroded region; and a learning unit that calculates a loss value using a predetermined loss function based on the error between the estimated value of the amount of metal thinning in the corroded region of the training image by the estimator and the measured value of the amount of metal thinning corresponding to the training image, and creates the estimator by learning based on the calculated loss value, where the predetermined loss function is a function that, if the error between the estimated value and the measured value is less than a predetermined threshold, calculates the loss value assuming that the error is zero.

[0009] According to the present disclosure, it is possible to improve the accuracy of an estimator that estimates the amount of thinning of a metal pipeline from an image captured inside the metal pipeline.

[0010] 8 is a diagram illustrating an example of the configuration of an estimation system including a learning device according to a first embodiment of the present disclosure. FIG. 1 is a diagram for explaining a learning image input to a learning image input unit shown in FIG. 1. FIG. 2 is a diagram for explaining learning data created by a learning data creation unit shown in FIG. 1. FIG. 3 is a diagram for explaining an example of a loss function used by the learning unit shown in FIG. 1. FIG. 4 is a diagram for explaining creation of an estimator by the learning unit shown in FIG. 1. FIG. 5 is a diagram for explaining an example of output from a wall-thinning amount output unit shown in FIG. 1. FIG. 6 is a flowchart illustrating an example of the operation of the learning device shown in FIG. 1. FIG. 7 is a diagram for explaining an example of the configuration of an estimation system including a learning device according to a second embodiment of the present disclosure. FIG. 8 is a diagram for explaining another example of the loss function used by the learning unit shown in FIG. 8. FIG. 9 is a diagram for explaining creation of an estimator by the learning unit shown in FIG. 8. FIG. 10 is a diagram for explaining evaluation of the estimator by the evaluation unit shown in FIG. 10. FIG. 11 is a diagram for explaining an example of the configuration of a computer that functions as an estimation system according to the present disclosure. FIG. 12 is a diagram for explaining wall-thinning of a metal pipeline. FIG. 13 is a diagram for explaining a loss function used for learning the estimator.

[0011] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0012] First Embodiment Fig. 1 is a diagram illustrating an example of the configuration of an estimation system 1 including a learning device 10 according to a first embodiment of the present disclosure. The estimation system 1 according to this embodiment estimates the amount of wall thinning of a metal pipe that is a target for wall thinning estimation. As shown in Fig. 1, the estimation system 1 according to this embodiment includes the learning device 10 and an estimation device 20.

[0013] The learning device 10 creates an estimator that estimates the amount of metal thinning in a corroded region inside a metal pipeline from an image of the corroded region inside the metal pipeline. As shown in Figure 1, the learning device 10 includes a training image input unit 11, a metal thinning amount input unit 12, a training data creation unit 13, and a learning unit 14.

[0014] An image of a corroded area inside a metal pipeline (hereinafter referred to as a "training image") is input to the training image input unit 11. As shown in Figure 2, the training image is preferably an image that has been enlarged and cut out from an image of the inside of a metal pipeline so as to include the entire corroded area (an area showing corrosion products generated by corrosion). Referring back to Figure 1, the training image input unit 11 outputs the input training image to the training data creation unit 13.

[0015] The measurement value of the amount of thinning of the metal pipe in the corroded area captured in the learning image is input to the thinning amount input unit 12. The thinning amount input unit 12 outputs the input measurement value of the amount of thinning of the metal pipe in the corroded area to the learning data creation unit 13.

[0016] The learning data creation unit 13 creates learning data that associates learning images of corroded areas inside a metal pipeline, output from the learning image input unit 11, with measured values ​​of the amount of metal thinning of the metal pipeline in the corroded areas, output from the wall-thinning input unit 12. Specifically, as shown in FIG. 3 , the learning data creation unit 13 pairs learning images of corroded areas with measured values ​​of the amount of metal thinning of the metal pipeline in the corroded areas, and creates learning data in which the measured values ​​of the amount of metal thinning are correct values ​​for the learning images. When the learning data creation unit 13 receives multiple learning images of corroded areas and multiple measured values ​​of the amount of metal thinning of the metal pipeline in the corroded areas, it creates multiple pieces of learning data that associate the learning images with the measured values ​​of the amount of metal thinning. Referring again to FIG. 1 , the learning data creation unit 13 outputs the created learning data to the learning unit 14.

[0017] The learning unit 14 creates an estimator that estimates the amount of wall thinning of a metal pipeline in a corroded region from an image of the corroded region inside the metal pipeline, by machine learning of the learning data created by the learning data creation unit 13. Specifically, the learning unit 14 calculates a loss value using a predetermined loss function based on the error between the estimated value of the amount of wall thinning in the corroded region of the learning image by the estimator and the measured value corresponding to the learning image. The learning unit 14 then creates an estimator by learning based on the calculated loss value. The learning unit 14 creates the estimator by repeating learning so as to reduce the loss value.

[0018] Here, the predetermined loss function Loss used by the learning unit 14 is expressed by the following equation (1).

[0019]

[0020] In equation (1), x is the absolute error between the measured value and the estimated value by the estimator, α is a predetermined threshold, and k is the number of data. Furthermore, MAPE is expressed by the following equation (2).

[0021]

[0022]

[0023] FIG. 4 is a diagram illustrating the shape of the function f(x) shown in Equation (1). As shown in FIG. 4, if the absolute error between the estimated value of the estimator and the measured value is less than a predetermined threshold α, the output value of the function f(x) is 0, i.e., the absolute error between the estimated value of the estimator and the measured value is 0. If the absolute error between the estimated value of the estimator and the measured value is equal to or greater than the predetermined threshold α, the output value of the function f(x) is the MAPE calculated according to Equation (2). As such, in this embodiment, the predetermined loss function is a function that calculates a loss value assuming no error when the error between the estimated value of the estimator and the measured value is less than the predetermined threshold α. The predetermined threshold α is, for example, a length corresponding to one pixel in the training image.

[0024] When the error between the estimated value of the amount of wall thinning by the estimator and the measured value of the amount of wall thinning is less than a predetermined threshold value α, a function that calculates a loss value assuming no error is used as the loss function. This prevents minute errors that are equal to or smaller than the length scale that can be discerned from the training image from being taken into consideration in training the estimator. As a result, the accuracy of the estimator can be improved.

[0025] As shown in FIG. 5 , the learning unit 14 determines the estimator with the smallest loss function value (loss value) among the multiple estimators (estimators A-1, A-2, A-3, ..., A-n in the example shown in FIG. 5 ) obtained by performing learning multiple times as the estimator to be used to estimate the amount of wall thinning. In the example shown in FIG. 5 , the loss value of estimator A-1 is 0.8, the loss value of estimator A-2 is 0.2, the loss value of estimator A-3 is 0.5, ..., and the loss value of estimator A-n is 0.7. Since the loss value of estimator A-2 is the smallest, the learning unit 14 determines estimator A-2 as the estimator to be used to estimate the amount of wall thinning. In this way, the learning unit 14 repeats the calculation of the loss value and learning based on the calculated loss value multiple times, and determines the estimator with the smallest loss value among the multiple estimators created by the repeated learning as the estimator to be used to estimate the amount of wall thinning.

[0026] Referring back to FIG. 1 , the learning unit 14 outputs the created estimator (the estimator determined as the estimator to be used for estimating the amount of wall thinning) to the estimation device 20 .

[0027] An image (hereinafter referred to as an "estimated image") of a corroded region inside a metal pipeline, which is the target of estimating the amount of wall thinning, is input to the estimation device 20. The estimation device 20 inputs the input estimated image to an estimator created by the learning device 10, and estimates the amount of wall thinning of the metal pipeline in the corroded region of the estimated image. As shown in FIG. 1 , the estimation device 20 includes an estimated image input unit 21, an estimation unit 22, and a wall thinning output unit 23.

[0028] An estimated image obtained by capturing an image of a corroded region inside a metal pipeline, which is the target of estimating the amount of wall thinning, is input to the estimated image input unit 21. The estimated image input unit 21 outputs the input estimated image to the estimation unit 22. As with the training image, the estimated image is preferably an image that has been enlarged and cut out from an image capturing the inside of the metal pipeline so as to include the entire corroded region (a region showing corrosion products generated by corrosion).

[0029] The estimation unit 22 inputs the estimated image output from the estimated image input unit 21 to an estimator created by the learning device 10, and estimates the amount of metal loss in the corroded region of the estimated image. The estimation unit 22 outputs the estimation result of the amount of metal loss in the corroded region of the estimated image to the metal loss output unit 23.

[0030] The metal-reduction amount output unit 23 outputs the amount of metal-reduction in the corroded region of the estimated image estimated by the estimation unit 22. For example, as shown in Fig. 6 , the metal-reduction amount output unit 23 outputs the amount of metal-reduction in the corroded region of the estimated image by superimposing an estimated value of the amount of metal-reduction in the corroded region on the corroded region of the estimated image.

[0031] Next, the operation of the learning device 10 according to this embodiment will be described. Fig. 7 is a flowchart showing an example of the operation of the learning device 10 according to this embodiment, and is a diagram for explaining the learning method executed by the learning device 10 according to this embodiment.

[0032] The training data creation unit 13 creates training data that associates training images of corroded regions inside a metal pipeline with measurement values ​​of the amount of metal thinning of the metal pipeline in the corroded regions (step S11). Specifically, for a training image input to the training image input unit 11, the training data creation unit 13 creates pairs of a training image and a measurement value of the amount of metal thinning, with the measurement value of the amount of metal thinning of the corroded region in the training image being the correct value.

[0033] The learning unit 14 calculates a loss value using a predetermined loss function based on the error between the estimated value of the metal-thinning amount in the corrosion region of the learning image by the estimator and the measured value of the metal-thinning amount corresponding to the learning image.The learning unit 14 then creates an estimator by learning based on the calculated loss value (step S12).Here, the predetermined loss function is a function that calculates a loss value as if there is no error when the error between the estimated value by the estimator and the measured value of the metal-thinning amount is less than a predetermined threshold α.

[0034] As described above, the learning device 10 according to this embodiment includes a learning data creation unit 13 and a learning unit 14. The learning data creation unit 13 creates learning data that associates learning images of corroded regions inside a metal pipeline with measured values ​​of the amount of metal thinning of the metal pipeline in the corroded regions. The learning unit 14 calculates a loss value using a predetermined loss function based on the error between the estimated value of the amount of metal thinning in the corroded region of the learning image by the estimator and the measured value of the amount of metal thinning corresponding to the learning image. The learning unit 14 then creates an estimator by learning based on the calculated loss value. Here, the predetermined loss function is a function that calculates a loss value assuming no error when the error between the estimated value and the measured value is less than a predetermined threshold value α.

[0035] By using such a function, minute errors below the length scale that can be discerned from the training images are not taken into account in training the estimator, which results in improved accuracy of the estimator.

[0036] Second Embodiment Fig. 8 is a diagram illustrating an example of the configuration of an estimation system 1A including a learning device 10A according to a second embodiment of the present disclosure. In Fig. 8, the same components as those in Fig. 1 are denoted by the same reference numerals, and descriptions thereof will be omitted.

[0037] 8, the estimation system 1A includes a learning device 10A and an estimation device 20. That is, the estimation system 1A according to this embodiment differs from the estimation system 1 according to the first embodiment in that the learning device 10 is changed to a learning device 10A.

[0038] Similar to the learning device 10, the learning device 10A receives training images of corroded regions inside a metal pipeline and measurements of the amount of metal thinning in the corroded regions, and creates an estimator that estimates the amount of metal thinning in the corroded regions inside the metal pipeline from the images of the inside of the metal pipeline. As shown in FIG. 8 , the learning device 10A includes a training image input unit 11, a metal thinning amount input unit 12, a training data creation unit 13A, a learning unit 14A, and an evaluation unit 15. The learning device 10A according to this embodiment differs from the learning device 10 according to the first embodiment in that the training data creation unit 13 and the learning unit 14 are replaced by the training data creation unit 13A and the learning unit 14A, respectively, and in that an evaluation unit 15 is added.

[0039] Similar to the learning data creation unit 13, the learning data creation unit 13A receives learning images of corroded areas inside a metal pipeline and measurements of the amount of thinning in the corroded areas, and creates learning data that associates the learning images with the measurements of the amount of thinning. The learning data creation unit 13A creates multiple pieces of learning data, outputs some of them to the learning unit 14A, and outputs the rest to the evaluation unit 15.

[0040] Similar to the learning unit 14, the learning unit 14A creates an estimator that estimates the amount of wall thinning of a metal pipeline in a corroded region from an image of the corroded region inside the metal pipeline through machine learning of the learning data. As described above, the learning unit 14 calculates a loss value using a predetermined loss function based on the error between the estimated value of the amount of wall thinning in the corroded region of the learning image by the estimator and the measurement value corresponding to the learning image. The learning unit 14 then creates an estimator through learning based on the calculated loss value. In this embodiment, the learning unit 14A creates multiple estimators using multiple loss functions with different parameters.

[0041] FIG. 9A is a diagram illustrating an example of a loss function used by the learning unit 14A to create an estimator. For example, as shown in FIG. 9A , the learning unit 14A creates multiple estimators using multiple loss functions (loss functions A, B, C, ...) with different thresholds α at which loss values ​​are calculated assuming no error between the estimated wall-thinning amount by the estimator and the measured wall-thinning amount. Also, as shown in FIG. 9B , the learning unit 14A creates multiple estimators using loss functions (loss functions A, B, C, ...) with different thresholds α at which loss values ​​are calculated assuming no error between the estimated wall-thinning amount by the estimator and the measured wall-thinning amount, and in which the loss value when the error x≧α is expressed as a quadratic function of the error x rather than as a MAPE. In this way, the learning unit 14A creates multiple estimators using multiple loss functions with different parameters (the shape when the specified site α and the error x≧α are satisfied).

[0042] The learning unit 14A performs learning multiple times for each of multiple functions with different parameters so as to minimize the loss value calculated by the loss function, thereby creating multiple estimators. That is, as shown in FIG. 10 , the learning unit 14A performs learning n times for each of loss functions A, B, and C, creating estimators A-1 to C-n. Then, for each loss function, the learning unit 14A outputs the estimator with the smallest loss value among the multiple estimators created using that loss function to the evaluation unit 15. In the example shown in FIG. 10 , of the estimators A-1 to A-n created using loss function A, estimator A-2 has the smallest loss value, so the learning unit 14A outputs estimator A-2 to the evaluation unit 15. Furthermore, of the estimators B-1 to B-n created using loss function B, estimator B-3 has the smallest loss value, so the learning unit 14A outputs estimator B-3 to the evaluation unit 15. Furthermore, among the estimators C-1 to Cn created using the loss function C, the loss value of the estimator C-3 is the smallest, so the learning unit 14A outputs the estimator C-3 to the evaluation unit 15.

[0043] Referring again to FIG. 8 , the evaluation unit 15 receives, from the learning data generation unit 13A, learning data (hereinafter referred to as “evaluation data”) that has not been used to create the estimator in the learning unit 14A. The evaluation unit 15 also receives, from the learning unit 14A, multiple estimators (estimators created using multiple loss functions with different parameters). The evaluation unit 15 uses the evaluation data to evaluate the estimation accuracy of the multiple estimators output from the learning unit 14A. Specifically, the evaluation unit 15 inputs, as evaluation images, learning images of the evaluation data (learning data not used to create the estimator) to each of the multiple estimators. Then, the evaluation unit 15 determines one of the multiple estimators to be used to estimate the amount of wall-thinning based on a correlation coefficient between the estimated value of each of the multiple estimators and the measured value of the amount of wall-thinning corresponding to the evaluation image. 11, the evaluation unit 15 inputs the evaluation image to each of the multiple estimators (estimators A-2, B-3, and C-3) output from the learning unit 14A, and calculates a correlation coefficient between the estimated value of each of the multiple estimators and the measured value of the amount of thinning corresponding to the evaluation image. The evaluation unit 15 calculates the correlation coefficient, for example, using the following equation (3).

[0044]

[0045]

[0046] The evaluation unit 15 determines the estimator with the highest correlation coefficient among the multiple estimators that calculated the correlation coefficient as the estimator to be used to estimate the amount of wall thinning, and outputs the estimator to the estimation device 20. In the example shown in FIG. 11 , the correlation coefficient of estimator A-2 is 0.5, the correlation coefficient of estimator B-3 is 0.7, and the correlation coefficient of estimator C-3 is 0.6, with estimator B-3 having the highest correlation coefficient. Therefore, the evaluation unit 15 determines estimator B-3 as the estimator to be used to estimate the amount of wall thinning, and outputs the estimator to the estimation device 20.

[0047] As described above, the learning device 10A according to this embodiment includes a learning unit 14A and an evaluation unit 15. The learning unit 14A creates multiple estimators using multiple loss functions with different parameters. The evaluation unit 15 inputs training images of the training data that were not used to create the estimators as evaluation images to each of the multiple estimators, and calculates a correlation coefficient between the estimated values ​​of each of the multiple estimators and the measured values ​​of the wall-thinning amount corresponding to the evaluation image. Then, based on the calculated correlation coefficient, the evaluation unit 15 determines one of the multiple estimators as the estimator to be used to estimate the wall-thinning amount.

[0048] Image resolution varies depending on the resolution of the image or the distance to the subject. That is, the image resolution of the training image and the estimated image varies depending on the resolution of the image captured inside the metal pipeline or the position at which the training image or the estimated image is cut out from the image captured inside the metal pipeline. As in this embodiment, by creating an estimator using multiple loss functions with different parameters, it is possible to search for an optimal loss function and improve the estimation accuracy of the estimator.

[0049] The above-described estimation system 1 (learning devices 10, 10A, and estimation device 20) can be realized by a computer 30 shown in FIG. 12 . A program for causing the computer 30 to function as the estimation system 1 (learning devices 10, 10A, and estimation device 20) may be provided. The program may be stored in a storage medium or provided via a network. FIG. 12 is a block diagram showing a schematic configuration of the computer 30 functioning as the estimation system 1 (learning devices 10, 10A, and estimation device 20). The computer 30 may be a general-purpose computer, a dedicated computer, a workstation, a PC (Personal Computer), an electronic notepad, or the like. The program instructions may be program code, code segments, or the like for performing required tasks.

[0050] 12, the computer 30 includes a processor 31, a ROM (Read Only Memory) 32, a RAM (Random Access Memory) 33, a storage 34, an input unit 35, a display unit 36, and a communication interface (I / F) 37. Each component is communicably connected to one another via a bus 39. The processor 31 is specifically a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), a SoC (System on a Chip), or the like, and may be configured with multiple processors of the same or different types.

[0051] The processor 31 is a control unit that controls each component and performs various arithmetic operations. That is, the processor 31 reads a program from the ROM 32 or the storage 34 and executes the program using the RAM 33 as a work area. The processor 31 controls each component and performs various arithmetic operations in accordance with the program stored in the ROM 32 or the storage 34. In this embodiment, the ROM 32 or the storage 34 stores a program for operating the computer 30 as the estimation system 1 (the learning device 10, 10A, and the estimation device 20) according to the present disclosure. The program is read and executed by the processor 31, thereby realizing each component of the estimation system 1 (the learning device 10, 10A, and the estimation device 20).

[0052] The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), a USB (Universal Serial Bus) memory, etc. The program may also be provided in a form downloaded from an external device via a network.

[0053] The ROM 32 stores various programs and various data. The RAM 33 temporarily stores programs or data as a working area. The storage 34 is configured with an HDD (Hard Disk Drive) or an SSD (Solid State Drive) and stores various programs including the operating system and various data.

[0054] The input unit 35 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0055] The display unit 36 ​​is, for example, a liquid crystal display, and displays various information. The display unit 36 ​​may be a touch panel type and function as the input unit 35.

[0056] The communication interface 37 is an interface for communicating with other devices, for example, an interface for a LAN.

[0057] The following additional notes are provided regarding the above-described embodiments.

[0058] [Supplementary Item 1] A learning device that creates an estimator that estimates the amount of metal thinning in a corroded region inside a metallic pipeline, comprising a control unit, wherein the control unit is configured to: create learning data that associates learning images of corroded regions inside a metallic pipeline with measured values ​​of the amount of metal thinning of the metallic pipeline in the corroded region; calculate a loss value using a predetermined loss function based on the error between the estimated value of the amount of metal thinning in the corroded region of the learning image by the estimator and the measured value of the amount of metal thinning corresponding to the learning image; and create the estimator by learning based on the calculated loss value, wherein the predetermined loss function is a function that, when the error between the estimated value and the measured value is less than a predetermined threshold, calculates the loss value assuming that the error is zero.

[0059] [Supplementary Item 2] In the learning device described in Supplementary Item 1, the control unit creates a plurality of estimators using a plurality of loss functions with different parameters, inputs a learning image of learning data that has not been used to create the estimator into each of the plurality of estimators as an evaluation image, calculates a correlation coefficient between the estimated value of each of the plurality of estimators and the measured value of the amount of metal-reduction corresponding to the evaluation image, and determines, based on the calculated correlation coefficient, one of the plurality of estimators as the estimator to be used to estimate the amount of metal-reduction.

[0060] In the learning device described in Supplementary Item 1, the control unit repeats the calculation of the loss value and learning based on the calculated loss value multiple times, and determines, from among multiple estimators created by the repeated learning, the estimator with the smallest loss value as the estimator for estimating the amount of wall thinning.

[0061] [Supplementary Item 4] A learning method executed by a learning device that creates an estimator that estimates the amount of wall thinning in a corroded region inside a metallic pipeline, the learning method comprising: creating learning data that associates learning images of a corroded region inside a metallic pipeline with measured values ​​of the amount of wall thinning of the metallic pipeline in the corroded region; calculating a loss value using a predetermined loss function based on the error between the estimated value of the amount of wall thinning in the corroded region of the learning image by the estimator and the measured value of the wall thinning corresponding to the learning image; creating the estimator by learning based on the calculated loss value; and the predetermined loss function is a function that, when the error between the estimated value and the measured value is less than a predetermined threshold, calculates the loss value assuming that the error is zero.

[0062] [Supplementary Item 5] A non-transitory storage medium storing a program executable by a computer, the non-transitory storage medium storing a program that causes the computer to operate as the learning device described in any one of Supplementary Items 1 to 3.

[0063] Although the above-described embodiments have been described as typical examples, it will be apparent to those skilled in the art that many modifications and substitutions can be made within the spirit and scope of the present disclosure. Therefore, the present invention should not be interpreted as being limited by the above-described embodiments, and various modifications and alterations are possible without departing from the scope of the claims. For example, multiple building blocks shown in the block diagrams of the embodiments can be combined into one, or one building block can be divided.

[0064] 1, 1A Estimation system 10, 10A Learning device 11 Learning image input unit 12 Thinning amount input unit 13, 13A Learning data creation unit 14, 14A Learning unit 15 Evaluation unit 20 Estimation device 21 Evaluation image input unit 22 Estimation unit 23 Thinning amount output unit 30 Computer 31 Processor 32 ROM 33 RAM 34 Storage 35 Input unit 36 ​​Display unit 37 Communication I / F 39 Path

Claims

1. A learning device that creates an estimator that estimates the amount of metal thinning in a corroded area inside a metallic pipeline, comprising: a learning data creation unit that creates learning data that associates learning images of a corroded area inside a metallic pipeline with measured values ​​of the amount of metal thinning of the metallic pipeline in the corroded area; and a learning unit that calculates a loss value using a predetermined loss function based on the error between the estimated value of the amount of metal thinning in the corroded area of ​​the learning image by the estimator and the measured value of the amount of metal thinning corresponding to the learning image, and creates the estimator by learning based on the calculated loss value, wherein the predetermined loss function is a function that calculates the loss value assuming that the error is zero when the error between the estimated value and the measured value is less than a predetermined threshold.

2. A learning device according to claim 1, wherein the learning unit repeats the calculation of the loss value and learning based on the calculated loss value multiple times, and determines the estimator with the smallest loss value from among multiple estimators created by the repeated learning as the estimator for estimating the amount of wall thinning.

3. A learning device according to claim 1, further comprising an evaluation unit that generates a plurality of estimators using a plurality of loss functions with different parameters, inputs a training image of training data not used in generating the estimator into each of the plurality of estimators as an evaluation image, calculates a correlation coefficient between the estimated value of each of the plurality of estimators and the measured value of the amount of metal thinning corresponding to the evaluation image, and determines, based on the calculated correlation coefficient, one of the plurality of estimators as the estimator to be used to estimate the amount of metal thinning.

Images