Information processing method, program, information processing device, and model generation method

Abstract

The present invention provides an information processing method, etc., capable of favorably estimating the state of a magnetic pipe. In the information processing method, a computer performs a process in which measurement data in which magnetic characteristic values of a magnetic pipe have been measured is acquired, and thickness information is estimated by inputting the acquired measurement data into a model which has been trained so as to estimate the thickness information relating to the thickness of the magnetic pipe when the measurement data has been input. Preferably, the magnetic characteristic values are measurement values measured by using an inspection probe provided with a magnet which generates a magnetic field, a yoke which is disposed on the opposite side of the magnetic tube with respect to the magnet, and a magnetic sensor which is disposed between the yoke and the magnetic pipe and measures the magnetic flux density passing through the yoke, magnet, and magnetic pipe. The output voltage of the magnetic sensor decreases as the magnetic flux density increases.

Description

Information processing method, program, information processing device, and model generation method

[0001] The present invention relates to an information processing method, a program, an information processing device, and a model generation method.

[0002] Known methods for non-destructively inspecting defects such as thinning in magnetic pipes include remote field eddy current testing (RFECT) and magnetic flux leakage (MFL).

[0003] For example, Patent Document 1 discloses a method for measuring defects in magnetic members by magnetic flux resistance (MFR).

[0004] Patent No. 6514592

[0005] However, these inspection methods predict the wall thickness of the magnetic pipe by applying the measured values ​​to a pre-prepared model formula, which requires manual creation of the model formula for predicting the wall thickness.

[0006] In one aspect, an object is to provide an information processing method and the like that can suitably estimate the state of a magnetic pipe.

[0007] In one aspect of the information processing method, a computer acquires measurement data that measures the magnetic property values ​​of a magnetic pipe, and inputs the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input, thereby performing a process of estimating the thickness information.

[0008] In one aspect, the state of the magnetic tube can be suitably estimated.

[0009] 1 is an explanatory diagram showing an example of the configuration of a defect estimation system. FIG. 1 is an explanatory diagram showing a measurement device. FIG. 1 is an explanatory diagram showing a measurement device. FIG. 2 is a block diagram showing an example of the configuration of a server. FIG. 2 is a block diagram showing an example of the configuration of a terminal. FIG. 3 is an explanatory diagram showing an example of the record layout of a user DB, a model DB, and a measurement DB. FIG. 4 is an explanatory diagram showing an overview of embodiment 1. FIG. 5 is an explanatory diagram showing a data extension process. FIG. 6 is an explanatory diagram showing a measurement data correction process. FIG. 7 is an explanatory diagram showing an image conversion process. FIG. 8 is a flowchart showing the procedure of an estimation model generation process. FIG. 9 is a flowchart showing the procedure of a thickness information estimation process. FIG. 10 is a diagram showing the procedure of a modification 1. FIG. 11 is a flowchart showing the procedure of an estimation model generation process according to modification 1. FIG. 12 is a flowchart showing the procedure of a thickness information estimation process according to modification 1. FIG. 13 is a diagram showing the procedure of a modification 2. FIG. 14 is a flowchart showing the procedure of a thickness information estimation process according to modification 2. FIG. 15 is an explanatory diagram showing the procedure of a modification 3. FIG. 16 is a diagram showing learning weighting in modification 3. FIG. 17 is a flowchart showing the procedure of an estimation model generation process according to modification 3. FIG. 18 is a flowchart showing the procedure of a thickness information estimation process according to modification 3. FIG. 19 is an explanatory diagram showing the influence of a Hall element voltage due to disturbance of magnetic flux in the MFR method. FIG. 10 is an explanatory diagram showing an overview of the dynamic time warping method in Modification 4. FIG. 11 is an explanatory diagram showing an overview of the dynamic time warping method in Modification 4. FIG. 12 is a diagram showing detected voltage and estimated thickness information in Modification 4. FIG. 13 is an explanatory diagram relating to data extension processing in Modification 5. FIG. 14 is an explanatory diagram showing an example of a measurement data upload screen. FIG. 15 is an explanatory diagram showing an example of a thickness information display screen. FIG. 16 is a flowchart showing the steps of a thickness information estimation process according to embodiment 2.

[0010] The present invention will be described in detail below with reference to the drawings showing embodiments thereof. (Embodiment 1) Fig. 1 is an explanatory diagram showing an example of the configuration of a defect estimation system. In this embodiment, a defect estimation system will be described that estimates defects such as wall thinning from measurement data obtained by measuring magnetic property values ​​of a magnetic pipe using a machine learning model. The defect estimation system includes an information processing device 1, a user terminal 2, a measuring device 3, and an inspection probe 4. The information processing device 1 and the user terminal 2 are communicatively connected via a network N such as the Internet.

[0011] The magnetic pipe to be measured in this embodiment is a pipe material made of a magnetic material such as carbon steel, ferritic stainless steel, or duplex stainless steel consisting of a ferrite phase and an austenite phase. Note that carbon steel is an example of a magnetic material, and the material (magnetic material) that makes up the magnetic pipe is not limited to these.

[0012] The measuring device 3 is a measuring device that measures the magnetic characteristic values ​​of a magnetic pipe and is a device for inspecting wall thinning of a magnetic pipe using the magnetic flux resistance method proposed by the applicant. The measuring device 3 measures the magnetic characteristic values ​​(magnetic flux density) at each position within the magnetic pipe by inserting an inspection probe 4 into the magnetic pipe and moving it within the pipe. As the measuring device 3, for example, a defect measuring device disclosed in Japanese Patent No. 6579840, Japanese Patent No. 6514592, and / or Japanese Patent Laid-Open No. 2019-100850 can be used.

[0013] 2A and 2B are explanatory diagrams of the measuring device 3. Fig. 2A is a cross-sectional view perpendicular to the longitudinal direction (axial direction) of the magnetic pipe, conceptually illustrating a cross-sectional view of the magnetic pipe with the inspection probe 4 inserted. Fig. 2B is a cross-sectional view perpendicular to the short side direction (radial direction) of the magnetic pipe, conceptually illustrating how the magnetic flux density of a defective portion of the magnetic pipe where thinning has occurred (hereinafter referred to as a "thinned portion") is measured. The magnetic flux resistance method will be briefly explained based on Figs. 2A and 2B.

[0014] The inspection probe 4 includes a yoke 41, a magnet 42, and a Hall element (magnetic sensor) 43. The yoke 41 is a hollow cylindrical magnetic member made of a high-magnetic-permeability metal such as carbon steel. The shape of the yoke 41 is not limited to a hollow cylindrical shape, and may be, for example, a rod, a plate, a column, or the like.

[0015] The magnets 42 and the Hall elements 43 are attached at equal intervals along the outer circumferential surface of the yoke 41. For example, the magnets 42 and the Hall elements 43 are provided at eight locations on the outer periphery of the yoke 41. The magnets 42 are arranged so that one magnetic pole faces the yoke 41 and the other magnetic pole faces the magnetic pipe, and are polarized in the direction facing the magnetic pipe.

[0016] 2A and 2B, the yoke 41 and the magnet 42 form a magnetic circuit. The Hall element 43 is provided on the magnetic circuit and outputs an output voltage proportional to the magnetic flux density passing through the Hall element 43. Specifically, the Hall element 43 outputs a higher output voltage as the magnetic resistance decreases and the magnetic flux density passing through the Hall element 43 increases.

[0017] Specifically, as shown in (a) of Figure 2B, when the magnet 42 and Hall element 43 are placed outside the magnetic pipe, the magnetic resistance is large, resulting in a small magnetic flux density, and the output voltage of the Hall element 43 is low. On the other hand, as shown in (c) of Figure 2B, when the magnet 42 and Hall element 43 are placed in a healthy part of the magnetic pipe (a part where no thinning has occurred), the magnetic resistance is small, resulting in a large magnetic flux density, and the output voltage of the Hall element 43 is high. As shown in (b) of Figure 2B, when the magnet 42 and Hall element 43 are placed in a thinned part, the wall thickness of the magnetic pipe is thinner than in a healthy part, resulting in a large magnetic resistance and a small magnetic flux density. Therefore, the output voltage of the Hall element 43 is low.

[0018] In this way, the measuring device 3 measures the magnitude of the magnetic flux density based on the level of the output voltage of the Hall element 43. The output voltage of the Hall element 43 varies depending on the thickness of the magnetic tube, and becomes lower as the thickness becomes thinner.

[0019] The magnetic flux resistance method has the advantage of being able to inspect a large number of magnetic pipes, and the inspection time per magnetic pipe is shorter than that of the ultrasonic immersion rotary ultrasonic thickness measurement method (IRIS; Internal Rotary Inspection System). Similar magnetic inspection methods, such as the remote field emission spectroscopy (RFECT) and magnetic flux leakage spectroscopy (MFL), have lower measurement accuracy than the magnetic flux resistance method.

[0020] In this embodiment, when estimating the wall thickness of the magnetic pipe from measurement data obtained by the magnetic flux resistance method, the wall thickness is estimated using a machine learning model.

[0021] Inspection probes 4 used in the magnetic flux resistance method include a P-type (Japanese Patent No. 6514592), in which there is a gap between the yoke 41 and the Hall element 43, as shown in FIG. 2B, and a PR-type (Japanese Patent Application Laid-Open No. 2019-100850), in which there is no gap between the yoke 41 and the Hall element 43 and the yoke 41 and the Hall element 43 are tightly attached. In both types, the output voltage is proportional to the magnetic flux density; however, in the P-type, the greater the magnetic resistance, the lower the output voltage, while in the PR-type, the greater the magnetic resistance, the higher the output voltage. For simplicity's sake, the measurement principle of the magnetic flux resistance method was explained above using a P-type as an example. However, the following description will focus on a case where a PR-type inspection probe 4 is used for measurement. Of course, either the P-type or the PR-type may be used.

[0022] Furthermore, in this embodiment, the measurement data is described as being measured by a flux-resistance method, but the measurement data is not limited to this. For example, eddy current values ​​measured by RFECT may be used as the measurement data. Furthermore, leakage magnetic flux (magnetic flux passing through a thinned portion) measured by MFL may be used as the measurement data. In other words, the measurement data may be any value that represents the magnetic characteristics of the magnetic tube, and is not limited to measurement data measured by a flux-resistance method.

[0023] Returning to FIG. 1 , the explanation continues. The information processing device 1 is an information processing device capable of various information processing and information transmission / reception, such as a server computer or a personal computer. In this embodiment, the information processing device 1 is assumed to be a server computer, and for simplicity, will be referred to as the server 1 below. The server 1 learns predetermined training data to generate an estimation model 50 (see FIG. 6 ) that estimates thickness information related to the wall thickness of a magnetic pipe when measurement data of the magnetic characteristic value of the magnetic pipe is input. The server 1 then acquires measurement data of the magnetic pipe to be measured from a user terminal 2 and inputs the data into the estimation model 50, thereby estimating the wall thickness information of the magnetic pipe. As will be described later, the wall thickness information is the wall thickness (thickness) of the magnetic pipe, but may also include the wall thinning width, wall thinning rate, wall thinning range, etc. of the magnetic pipe.

[0024] The user terminal 2 is a terminal device used by a user of the system, such as a personal computer or tablet terminal. Hereinafter, for simplicity, the user terminal 2 will be referred to as terminal 2. The user of the system is, for example, an inspection company that performs non-destructive testing of magnetic pipes, but is not limited to this. While only one terminal 2 is shown in FIG. 1 , it is assumed that multiple users each have their own terminals 2, 2, 2, ... connected to the server 1. The terminal 2 is connected to the measuring device 3, transmits measurement data of the magnetic pipe measured by the measuring device 3 to the server 1, and obtains wall thickness information estimated based on the estimation model 50 from the server 1.

[0025] In this embodiment, the server 1 on the cloud estimates the wall thickness information based on the estimation model 50, but the embodiment is not limited to this. For example, the local terminal 2 may install data of the estimation model 50 from the server 1 in advance, and estimate the wall thickness information by inputting measurement data acquired from the measurement device 3 into the estimation model 50. In this way, the device that generates the estimation model 50 and the device that estimates the wall thickness information based on the estimation model 50 may be different devices.

[0026] 3 is a block diagram showing an example configuration of the server 1. The server 1 includes a control unit 11, a main memory unit 12, a communication unit 13, and an auxiliary memory unit 14. The control unit 11 has one or more central processing units (CPUs), micro-processing units (MPUs), graphics processing units (GPUs), or other processing devices, and performs various information processing by reading and executing a program P1 stored in the auxiliary memory unit 14. The main memory unit 12 is a temporary storage area such as static random access memory (SRAM), dynamic random access memory (DRAM), or flash memory, and temporarily stores data necessary for the control unit 11 to execute information processing. The communication unit 13 is a communication module for performing communication-related processing, and transmits and receives information to and from the outside.

[0027] The auxiliary storage unit 14 is a non-volatile storage area such as a large-capacity memory or a hard disk, and stores the program P1 and other data necessary for the control unit 11 to execute processing. The auxiliary storage unit 14 also stores an estimation model 50, a user DB 141, a model DB 142, and a measurement DB 143. The estimation model 50 is a machine learning model that has learned predetermined training data, and is a model that estimates thickness information regarding the wall thickness of a magnetic pipe when measurement data measuring the magnetic property values ​​of the magnetic pipe is input. The estimation model 50 is expected to be used as a program module that constitutes part of artificial intelligence software.

[0028] The user DB 141 is a database that stores information about users of this system. The model DB 142 is a database that stores information about each of a plurality of estimation models 50 prepared according to the wall thinning form (shape pattern of the wall thinning portion), dimensions of the magnetic pipe, etc. As will be described later, in this embodiment, a plurality of estimation models 50 are prepared according to the wall thinning form, dimensions, etc. of the magnetic pipe to be estimated. The measurement DB 143 is a database that stores measurement data of the magnetic pipe acquired from the terminal 2.

[0029] The auxiliary storage unit 14 may be an external storage device connected to the server 1. The server 1 may be a multi-computer consisting of multiple computers, or may be a virtual machine virtually constructed by software.

[0030] In this embodiment, the server 1 is not limited to the above configuration and may include, for example, an input unit that accepts operational inputs, a display unit that displays images, etc. The server 1 may also include a reading unit that reads a non-transitory computer-readable recording medium 1a and reads the program P1 from the recording medium 1a. The program P1 may be executed on a single computer or on multiple computers interconnected via a network N.

[0031] 4 is a block diagram showing an example configuration of the terminal 2. The terminal 2 includes a control unit 21, a main memory unit 22, a communication unit 23, a display unit 24, an input unit 25, and an auxiliary memory unit 26. The control unit 21 is one or more arithmetic processing units such as a CPU or an MPU, and performs various information processing by reading and executing a program P2 stored in the auxiliary memory unit 26. The main memory unit 22 is a temporary storage area such as a RAM, and temporarily stores data necessary for the control unit 21 to execute information processing. The communication unit 23 is a communication module for performing processing related to communication, and transmits and receives information to and from the outside.

[0032] The display unit 24 is a display screen such as a liquid crystal display and displays images. The input unit 25 is an operation interface such as a keyboard or mouse and accepts operation input from a user. The auxiliary storage unit 26 is a non-volatile storage area such as a hard disk or large-capacity memory and stores the program P2 and other data required for the control unit 21 to execute processing.

[0033] The terminal 2 may include a reading unit that reads the non-transitory computer-readable recording medium 2 a and reads the program P2 from the recording medium 2 a. The program P2 may be executed on a single computer or on multiple computers interconnected via a network N.

[0034] 5 is an explanatory diagram showing an example of the record layout of the user DB 141, the model DB 142, and the measurement DB 143. The user DB 141 includes a user ID column, a user name column, a usage history column, and a usage fee column. The user ID column stores a user ID for identifying each user. The user name column, the usage history column, and the usage fee column each store, in association with the user ID, the user's usage history of the estimation model 50 (thickness information estimation history), and the usage fee (system usage fee) for the estimation model 50 determined according to the usage history. The usage fee will be described in detail in embodiment 2.

[0035] The model DB 142 includes a model ID column, a model name column, and a target column. The model ID column stores model IDs for identifying each estimation model 50 prepared according to the wall thinning form of the magnetic pipe, etc. The model name column and target column store the model name of the estimation model 50 and magnetic pipe information related to the magnetic pipe to be estimated, respectively. For example, the target column stores the wall thinning form, dimensions, and material of the magnetic pipe.

[0036] The measurement DB 143 includes a data ID column, a date and time column, a providing user column, a target column, a measurement data column, and a second measurement data column. The data ID column stores data IDs for identifying measurement data provided by each user. The date and time column, providing user column, target column, measurement data column, and second measurement data column each store, in association with a data ID, the date and time the measurement data was provided (acquisition date and time), the user name of the providing user, magnetic pipe information (wall thinning pattern, etc.) about the magnetic pipe being measured, measurement data of the magnetic pipe, and second measurement data. The second measurement data is measurement data of the wall thickness of the magnetic pipe, which is measured using a different method from the measurement data using the magnetic flux resistivity method, for example, measurement data using a water-immersion rotary ultrasonic thickness measurement method (IRIS; Internal Rotary Inspection System). The second measurement data will be described in detail in embodiment 2.

[0037] Fig. 6 is an explanatory diagram showing an overview of the first embodiment. Fig. 6 conceptually illustrates how the wall thickness of each portion of the magnetic pipe is estimated when measurement data of the magnetic characteristic values ​​of the magnetic pipe measured by the measurement device 3 (inspection probe 4) is input into an estimation model 50. The overview of this embodiment will be described based on Fig. 6.

[0038] As described above, the server 1 generates the estimation model 50 that estimates thickness information when measurement data of magnetic characteristic values ​​is input by learning predetermined training data. The estimation model 50 according to the present embodiment is a neural network generated by deep learning, such as a CNN (Convolutional Neural Network, ResNet, etc.).

[0039] The estimation model 50 may be a neural network other than a CNN. As in Modifications 1 and 2 described below, the estimation model 50 may be a model based on a learning algorithm other than a neural network, such as a k-Nearest Neighbor Algorithm (kNN), a Support Vector Machine (SVM), a random forest, or a decision tree.

[0040] The estimation model 50 according to this embodiment receives as input measurement data based on the magnetic flux resistance method, which is measured using the measurement device 3. As already mentioned, the measurement data is a value measured by the Hall element 43 of the magnetic flux density passing through the magnetic circuit formed by the yoke 41 and magnet 42 provided in the inspection probe 4 and the magnetic tube, and is an output voltage of the Hall element 43 that is proportional to the magnetic flux density.

[0041] When measuring a magnetic pipe, the user inserts the inspection probe 4 into an opening in the magnetic pipe and moves (sweeps) it to the opening on the opposite side. This measures the magnetic characteristic values ​​at each position along the longitudinal direction of the magnetic pipe. As described above, multiple Hall elements 43 are attached to the yoke 41 of the inspection probe 4, and the measurement device 3 measures the magnetic characteristic values ​​at each position on a cross section perpendicular to the longitudinal direction at each position along the longitudinal direction. Specifically, multiple Hall elements 43 are attached periodically on the outer periphery of the cylindrical yoke 41, and the measurement device 3 measures the magnetic characteristic values ​​at each position obtained by dividing the cross section of the cylindrical magnetic pipe into equal parts along the circumferential direction.

[0042] In this embodiment, the magnetic tube is described as being cylindrical, but the cross-sectional shape of the magnetic tube is not limited to being circular and may be, for example, rectangular. That is, the measuring device 3 only needs to be able to measure the magnetic characteristic values ​​at each position on a cross section perpendicular to the longitudinal direction of the magnetic tube using the multiple Hall elements 43 provided on the inspection probe 4, and the "each position on the cross section" is not limited to each position obtained by equally dividing the cross section of the cylindrical magnetic tube along the circumferential direction.

[0043] In the following description, the Hall elements 43 that function as magnetic sensors will be referred to as "channels" as appropriate, and each Hall element 43 will be assigned a channel number and described as "CH1," "CH2," "CH3," . . . "CH8."

[0044] By measuring the magnetic pipe as described above, waveform data of the magnetic characteristic values ​​(output voltages) of each channel that continues along the longitudinal direction of the magnetic pipe is observed, as shown on the left side of Figure 6. Note that the horizontal axis of the waveform data represents the position in the longitudinal direction (axial direction) of the magnetic pipe, and the vertical axis represents the output voltage of the Hall element 43. The server 1 uses this waveform data as input to the estimation model 50.

[0045] When the above measurement data is input, the estimation model 50 outputs (estimates) wall thickness information of the magnetic pipe. Specifically, the estimation model 50 estimates the wall thickness (thickness) of the magnetic pipe. In this embodiment, the problem handled by the estimation model 50 is defined as a classification problem, and the estimation model 50 classifies the wall thickness of the magnetic pipe in a predetermined numerical range (e.g., 1.0 to 2.3 mm) and in predetermined length increments (e.g., 0.1 mm increments). Of course, the estimation model 50 may also be a regression model.

[0046] Furthermore, in this embodiment, the estimation model 50 estimates the wall thickness of the magnetic pipe, but this embodiment is not limited to this. For example, the estimation model 50 may estimate a wall-thickness reduction width, which is calculated by subtracting the wall thickness of the thinned portion from the wall thickness of the sound portion, instead of the wall thickness of the magnetic pipe. Furthermore, for example, the estimation model 50 may estimate a wall-thickness reduction rate, which is calculated by dividing the wall thickness of the thinned portion by the wall thickness of the sound portion. Furthermore, for example, the estimation model 50 may estimate the position of the wall-thickness reduction portion within the magnetic pipe, the wall-thickness reduction range (the width of the wall-thickness reduction portion along the longitudinal direction and / or circumferential direction), the wall-thickness reduction form (the shape pattern of the wall-thickness reduction), etc.

[0047] Furthermore, in this embodiment, the target of estimation is thinning of the magnetic pipe, but this embodiment is not limited to this, and an increase in wall thickness due to rust or the like may also be the target of estimation.

[0048] In this way, the wall thickness information output from the estimation model 50 may be information related to the wall thickness of the magnetic pipe. The estimated wall thickness information is not limited to the wall thickness of the magnetic pipe itself, and the defect to be estimated is not limited to wall thinning.

[0049] For example, the estimation model 50 estimates the wall thickness of the magnetic pipe for each section obtained by dividing the magnetic pipe into fixed lengths along the longitudinal direction. The right side of Figure 6 conceptually illustrates the estimated wall thickness. The estimation model 50 divides the magnetic pipe into fixed lengths (e.g., 3 cm) and estimates the minimum wall thickness for each section. The estimation model 50 may also estimate the wall thickness continuously without dividing into fixed lengths, or may estimate a value other than the minimum value (e.g., the maximum or average wall thickness for each section). The estimation model 50 may also estimate the maximum wall thickness for each channel, or may estimate the minimum wall thickness for all channels combined as a representative value.

[0050] In the above description, the magnetic property values ​​of each position (each channel) on the cross section at each position in the longitudinal direction of the magnetic pipe are input to the estimation model 50, but this embodiment is not limited to this. For example, the server 1 may input only the magnetic property values ​​of a single position (cross section) in the longitudinal direction of the magnetic pipe to the estimation model 50 and estimate the thickness information at that position. Also, for example, the server 1 may input only the magnetic property values ​​of a single position (channel) on the cross section of the magnetic pipe to the estimation model 50 and estimate the thickness information at that position.

[0051] The server 1 generates an estimation model 50 using training data in which correct values ​​(labels) of wall thickness information are associated with a group of training measurement data. The training measurement data is actual measurement data of one or more magnetic pipes obtained by the measurement device 3, and is waveform data for each channel in which magnetic property values ​​are measured at each position in the longitudinal direction. Specifically, the training measurement data is data measured by inserting the inspection probe 4 into both ends of the magnetic pipe while changing the relative position of the inspection probe 4 in the circumferential direction so that CH1 is approximately 0°, 90°, 180°, and 270°, with a certain radial direction (position on the cross section) being 0°. In other words, the data is data obtained by performing 4 x 2 = 8 measurements per magnetic pipe. The correct values ​​are data obtained by measuring the wall thickness of the magnetic pipe using a 3D shape measuring device, and are data measured at intervals of approximately 23 μm at 0.2 mm intervals from the end of the magnetic pipe. The server 1 generates an estimation model 50 using training data in which correct values ​​of wall thickness for each portion are assigned to the training measurement data.

[0052] FIG. 7 is an explanatory diagram of the data expansion process. The server 1 may use only raw measurement data measured by the inspection probe 4 as input to the estimation model 50. However, in this embodiment, in order to improve the accuracy of estimating thickness information, the raw measurement data is subjected to a certain amount of preprocessing, and the preprocessed measurement data is used as input to the estimation model 50. The server 1 may also learn only the training data (actual measurement data). However, in this embodiment, the training measurement data (actual measurement data) is expanded to increase the amount of training data. The server 1 learns the expanded training data and generates the estimation model 50. FIGS. 7(a) to 7(d) conceptually illustrate how the preprocessed measurement data is expanded.

[0053] 7A conceptually illustrates preprocessed data in which summary statistics such as the mean value and standard deviation are added to the training measurement data, and data in which the magnetic property values ​​have been normalized and corrected is added. The preprocessing of the measurement data will be described with reference to FIG.

[0054] The server 1 uses the longitudinal position where the magnetic characteristic value was measured and the raw measurement value of each channel ("CH1," "CH2," ... "CH8" in Figure 7A) as raw measurement data. The reason for using the longitudinal position in addition to the measurement value is that the output voltage of the Hall element 43 changes drastically at the end (entrance / exit) of the magnetic pipe, so the position of the pipe end is also used as an explanatory variable. For the raw measurement data, the server 1 first normalizes the magnetic characteristic value of each channel so that the maximum magnetic characteristic value outside the magnetic pipe (in the air) and the minimum magnetic characteristic value in the healthy section are consistent across all channels (all positions along the circumferential direction) (middle part of Figure 7A). Specifically, the server 1 corrects the output voltage of the Hall element 43 outside the magnetic pipe to 1 V and the output voltage in the healthy section to 0 V. For the P-type, the output voltage is minimum outside the magnetic pipe and maximum in the healthy section. By normalizing the magnetic characteristic values, it is possible to correct variations in sensitivity inherent to each Hall element 43 and also correct the influence of temperature on the Hall elements 43 .

[0055] Next, the server 1 corrects the measurement data of each channel to minimize the influence of lift-off of the inspection probe 4 (and the entire surface thinning portion) (lower part of FIG. 7(a)).

[0056] FIG. 8 is an explanatory diagram of the measurement data correction process. The left side of FIG. 8 shows the measurement data before correction, and the right side of FIG. 8 shows the measurement data after correction. It is desirable for the inspection probe 4 to pass through the central axis of the magnetic pipe during measurement. However, as shown on the left side of FIG. 8, in reality, the position of the inspection probe 4 within the magnetic pipe depends on the user's operation, and the distance between each channel CH1, CH2, CH3, etc. and the pipe wall often varies. Therefore, the server 1 minimizes the effects of lift-off by correcting the magnetic property values ​​of each channel to data that would be obtained if the inspection probe 4 had passed through the central axis of the magnetic pipe.

[0057] Specifically, the server 1 smooths the measurement data by taking a longitudinal moving average for each channel and identifies the baseline for each channel. The server 1 then corrects the values ​​of each channel so that the average values ​​of the baselines for each channel match. In other words, the server 1 adjusts the height of the waveform so that the smoothed waveform shapes of each channel are approximately the same. The longitudinal length for taking the moving average and the number of averaging operations are adjusted according to the length (total length) of the magnetic tube.

[0058] Furthermore, the server 1 preferably corrects the measurement data to include only localized thinning areas by removing data on areas where the output voltage changes gradually, i.e., data on areas with thinning across the entire surface. Note that a thinned area across the entire surface refers to an area where the change in the output voltage (magnetic characteristic value) of the Hall element 43 is below a certain level. For example, the server 1 identifies the longitudinal range where the thinned area exists by calculating the average baseline value of all channels, and then replaces the output voltage in that range with the value of the healthy area (i.e., 0 V), thereby removing the data on the thinned area across the entire surface. This eliminates voltage changes in areas with thinning across the entire surface, allowing the data to be adjusted to include only localized thinning areas (areas where the change in output voltage is above a certain level).

[0059] 7 , the explanation will be continued. As described above, the server 1 generates data by normalizing the raw measurement data and corrected data by removing the effects of lift-off (and the entire surface thinning portion). The server 1 may use only the normalized data and / or the corrected data as input to the estimation model 50, but in this embodiment, all three types of data, the raw measurement data, the normalized data, and the corrected data, are used as input.

[0060] The server 1 uses data obtained by adding predetermined summary statistics to each of the three types of data in the estimation model 50. The summary statistics include, but are not limited to, the mean, standard deviation, skewness, and / or kurtosis. The server 1 calculates the mean, standard deviation, skewness, and / or kurtosis along the longitudinal and / or circumferential directions from the magnetic property values ​​at each position in the longitudinal and / or circumferential directions and adds them to the data.

[0061] That is, the server 1 calculates the mean, standard deviation, skewness, and kurtosis for each position (channel) along the circumferential direction of the magnetic pipe and adds them to the data. The server 1 also calculates the moving mean, moving standard deviation, moving skewness, and moving kurtosis for multiple positions (e.g., 10, 20, 30, 40, and 50 points) along the longitudinal direction of the magnetic pipe and adds them to the data. In the table of Figure 7, the "mean," "standard deviation," "skewness," and "kurtosis" correspond to summary statistics in the circumferential direction, while the "10-point moving average" and "20-point moving average" correspond to summary statistics (moving averages) in the longitudinal direction. For convenience of illustration, the moving standard deviation, moving skewness, and moving kurtosis are not shown in Figure 7. Adding these summary statistics can improve the accuracy of estimating wall thickness information. In particular, according to the inventor's research, adding skewness and kurtosis reduced estimation errors.

[0062] In this way, the server 1 performs a certain amount of preprocessing on the raw measurement data, and uses the preprocessed measurement data as input to the estimation model 50. The final preprocessed measurement data becomes matrix data with N rows and 275 columns (N is the number of measurement points along the longitudinal direction). Similarly, when actually estimating thickness information from measurement data of a magnetic pipe using the estimation model 50, the server 1 performs preprocessing on the measurement data acquired from the terminal 2, and inputs the preprocessed data to the estimation model 50.

[0063] During learning, the server 1 expands the measurement data shown in FIG. 7A to increase the number of learning data. As a first data expansion process, the server 1 generates multiple patterns of measurement data by shifting the positions on the cross section of the magnetic tube where the magnetic characteristic values ​​were measured along the circumferential direction. Specifically, the server 1 generates measurement data by shifting the channel numbers of the measurement data for each channel. FIG. 7B conceptually illustrates the first data expansion process. The server 1 first shifts the channel numbers by one, generating measurement data in which the data from CH1 is set to CH2, the data from CH2 to CH3, ..., and the data from CH8 to CH1. Next, the server 1 shifts the channel numbers by two, generating measurement data in which the data from CH1 is set to CH3, the data from CH2 to CH4, ..., and the data from CH8 to CH2. In the same manner, the server 1 shifts the channel numbers, increasing the number of data by eight times. This makes it possible to generate measurement data when a thinning portion is detected in CH2, measurement data when a thinning portion is detected in CH3, etc. based on the measurement data when a thinning portion is detected in CH1, and to correct any bias in the channel (position on the cross section of the magnetic tube) in which the thinning portion is detected.

[0064] Next, as a second data augmentation process, the server 1 generates multiple patterns of measurement data by adding multiple patterns of predetermined noise to the measurement data. The second data augmentation process is illustrated in (c) of Figure 7. For example, the server 1 adds 10 patterns of Gaussian noise (mean 0, variance 0.02) to the measurement data to increase the number of data by 10 times. This allows the estimation model 50 to learn from a slight misalignment of a wall-thinning area, such as when a wall-thinning area is located between one channel and another, as shown on the right side of (c) of Figure 7.

[0065] As a third data augmentation process, the server 1 generates measurement data with a modified sampling rate for measuring magnetic property values. (D) of FIG. 7 illustrates the third data augmentation process. The server 1 performs linear interpolation between two consecutive points in the longitudinal direction of the training measurement data to approximate the magnetic property value at the position between the two points. For example, the server 1 generates data with one-point interpolation, two-point interpolation, ..., 20-point interpolation to generate measurement data with a sampling rate of 1x to 6x. This allows changes in the movement speed (sweep speed) of the inspection probe 4 to be reproduced and trained by the estimation model 50.

[0066] For convenience, the tables in Figures 7(b) to (d) only show the measurement values ​​for each channel ("CH1", "CH2", ...), but it goes without saying that the data can be expanded to include the summary statistics and the like mentioned above.

[0067] By performing the first to third data expansion processes, the server 1 expands the number of measurement data by 8 × 10 × 6 = 480 times. This increases the amount of learning data and allows the estimation model 50 to learn about channel bias, minute positional deviations of thinned parts, changes in movement speed, and the like.

[0068] 6 , the server 1 uses the expanded measurement data as training data to generate an estimation model 50. While the server 1 may use the measurement data (magnetic property values) as input to the estimation model 50, in this embodiment, as shown in FIG. 6 , the server 1 converts the measurement data into an image and uses the image as input to the estimation model 50.

[0069] 9 is an explanatory diagram of the image conversion process. Fig. 9 illustrates the process of converting measurement data into an image. In this embodiment, the server 1 generates an image in which the magnetic property value at each position in the magnetic pipe is expressed by the pixel value of each pixel, and uses this image as input to the estimation model 50.

[0070] Specifically, the server 1 generates an image in which the magnetic property values ​​at each position are represented by the color (e.g., RGB) of each pixel. The vertical axis (first axis) of the image corresponds to the longitudinal position of the magnetic pipe, and the horizontal axis corresponds to the position on the cross section of the magnetic pipe perpendicular to the longitudinal direction, i.e., the channel. For example, the server 1 divides the measurement data (see FIG. 7 ) in matrix format, which represents each position along the longitudinal direction of the magnetic pipe as a row and each channel's raw measurement value, normalized value, correction value, and summary statistics as a column, into a predetermined number of rows (e.g., 224 rows). The server 1 then assigns the measurement data at each position to each pixel in the image by sequentially storing the measurement data from rows 1 to 224, rows 225 to 448, etc., starting from the left edge of the image. If there are any remaining pixels in the image, the server 1 sets the R, G, and B values ​​of the remaining portions to 0. Note that depending on the sampling rate during the third data extension, most of the remaining values ​​may be zero.

[0071] The server 1 generates hue images of each of the hues R, G, and B from the training measurement data, and generates a final input image by combining the hue images. Specifically, the server 1 calculates power values ​​of multiple patterns of magnetic characteristic values ​​with different exponents from the measurement data, and assigns the power value of each pattern to each hue to generate each hue image.

[0072] For example, the server 1 calculates the zeroth power value, the first power value, and the squared power value for the magnetic property value at each position in the magnetic pipe. The server 1 then assigns each power value to R, G, and B. FIG. 9 illustrates a case where the first power value is assigned to R, the squared power value to G, and the zeroth power value to B. The server 1 generates each hue image by storing the power value of the magnetic property value at each position as the pixel value of the pixel corresponding to that position. The server 1 combines the hue images to generate the final input image.

[0073] In the following description, an image obtained by combining the individual hue images will be referred to as a "composite image."

[0074] The above image conversion method is merely an example, and the present embodiment is not limited to this. For example, the server 1 may generate a composite image by treating the final composite image as a CMYK image rather than an RGB image and assigning values ​​from 0 to 4 to each hue. Furthermore, the hues to which power values ​​are assigned are not limited to primary colors (RGB). Furthermore, the exponent of the power value is not limited to a natural number, but may be a real number other than a natural number (e.g., 1.5). Furthermore, the server 1 may generate a monochrome image in which the magnetic characteristic value at each position within the magnetic pipe is simply stored in each pixel without calculating the power value. Furthermore, the server 1 may represent the magnetic characteristic value using a pixel value other than a color (e.g., grayscale). Thus, the server 1 is not particularly limited as long as it can generate an image in which a pixel value for each pixel is assigned according to the magnetic characteristic value at each position within the magnetic pipe.

[0075] During learning, the server 1 generates composite images for all combinations of assigning each power value to each hue to prevent overlearning due to hue. That is, in addition to a composite image in which a zeroth power value is assigned to R, a first power value to G, and a squared power value to B, the server 1 generates six patterns of composite images, such as a composite image in which a zeroth power value is assigned to G, a first power value to B, and a squared power value to R, a composite image in which a zeroth power value is assigned to B, a first power value to R, and a squared power value to G, etc. The server 1 prevents overlearning due to hue by having the estimation model 50 learn composite images of each combination.

[0076] 6 , the server 1 performs data expansion on each measurement data provided for training and converts each expanded measurement data into a composite image. The server 1 then generates an estimation model 50 using each composite image and the correct values ​​of the thickness information corresponding to each composite image as training data.

[0077] That is, the server 1 obtains an estimated value of the wall thickness information by inputting a training synthetic image into the estimation model 50, and compares the estimated value with a correct answer value. The server 1 updates parameters such as weights between neurons so that the estimated value and the correct answer value approximate each other. The server 1 sequentially performs learning using each pair of synthetic image and correct answer value, and finally generates an estimation model 50 with optimized weights, etc.

[0078] In this embodiment, the server 1 learns the measurement data according to the magnetic pipe that was the measurement target in the training measurement data, and generates multiple estimation models 50 according to various magnetic pipes. Specifically, the server 1 learns the measurement data separately according to magnetic pipe information related to the magnetic pipe that was the measurement target, and generates multiple estimation models 50 according to the magnetic pipe information.

[0079] Magnetic pipe information is information that indicates the state or attributes of a magnetic pipe, such as the thinning form, dimensions, and material of the magnetic pipe. Note that these are examples of magnetic pipe information, and other information may also be used. The thinning form is the state of the thinned part of the magnetic pipe, and is defined by the shape pattern of the thinned part (for example, whether the cross-sectional shape of the thinned part is rectangular or hemispherical). The dimensions are defined by the diameter and length of the magnetic pipe. The material is defined by the type of magnetic material that forms the magnetic pipe.

[0080] The server 1 learns the training data separately according to the magnetic pipe information, and generates an estimation model 50 according to the wall-thinning form, dimensions, etc. of the magnetic pipe. Note that, although a plurality of estimation models 50 are generated in this embodiment, it is also possible to have one estimation model 50 learn the training data regardless of differences in the magnetic pipe information, and estimate wall-thickness information using the single estimation model 50.

[0081] When actually estimating wall thickness information using the estimation model 50, the server 1 acquires measurement data of a magnetic pipe measured by a user from the terminal 2 and inputs the measurement data into the estimation model 50 to estimate the wall thickness information. Specifically, the server 1 acquires the measurement data and accepts, from the user via the terminal 2, specified input of magnetic pipe information regarding the magnetic pipe to be measured. The server 1 selects one of multiple estimation models 50 based on the specified magnetic pipe information. The server 1 generates hue images of each hue from the acquired measurement data and generates a composite image by combining the hue images. The server 1 inputs the composite image into the selected estimation model 50 to estimate the wall thickness of each portion obtained by dividing the magnetic pipe into fixed lengths along the longitudinal direction. The server 1 returns the estimated wall thickness information to the terminal 2, which displays it.

[0082] The operation (display screen, etc.) of the terminal 2 when estimating the wall thickness information will be described in detail in the second embodiment.

[0083] As described above, according to this embodiment, by using the estimation model 50, the state of thinning of the magnetic pipe (thickness information) can be suitably estimated and presented to the user.

[0084] FIG. 10 is a flowchart illustrating the steps of the process for generating the estimation model 50. The process for generating the estimation model 50 through machine learning will be described with reference to FIG. 10 . The control unit 11 of the server 1 acquires training data in which a group of measurement data measuring the magnetic property values ​​of the magnetic pipe is associated with a ground truth value of thickness information regarding the wall thickness of the magnetic pipe (step S11). The measurement data is data obtained by measuring each position on a cross section perpendicular to the longitudinal direction of the magnetic pipe (each position obtained by dividing the cross section of the cylindrical magnetic pipe equally along the circumferential direction). The magnetic property values ​​are measured using an inspection probe 4 including a magnet 42 that generates a magnetic field, a yoke 41 positioned on the opposite side of the magnetic pipe from the magnet 42, and a Hall element 43 (magnetic sensor) positioned between the yoke 41 and the magnetic pipe to measure the magnetic flux density passing through the yoke 41, the magnet 42, and the magnetic pipe. The output voltage of the Hall element 43 is proportional to the magnetic flux density. The wall thickness information is, for example, the wall thickness of each section obtained by dividing the magnetic pipe into fixed lengths along the longitudinal direction.

[0085] The control unit 11 performs preprocessing on the training measurement data (step S12). Specifically, the control unit 11 normalizes the magnetic characteristic values ​​so that the magnetic characteristic values ​​outside the magnetic tube and the magnetic characteristic values ​​in the healthy section match for each channel. The control unit 11 also determines the baseline for each channel by taking a moving average of the magnetic characteristic values ​​along the longitudinal direction, and corrects the magnetic characteristic values ​​to those when the inspection probe 4 passes through the central axis of the magnetic tube by matching the average values ​​of the baselines for each channel. The control unit 11 also adds summary statistics such as the mean value and standard deviation to the measurement data.

[0086] The control unit 11 expands the number of pieces of measurement data (step S13). Specifically, the control unit 11 generates multiple patterns of measurement data by shifting the channel numbers of the data measured in each channel. The control unit 11 also generates multiple patterns of measurement data by adding multiple patterns of predetermined noise (e.g., Gaussian noise) to the measurement data. The control unit 11 also generates multiple patterns of measurement data by performing linear interpolation between two consecutive points in the longitudinal direction.

[0087] The control unit 11 converts the measurement data with the expanded data count into an image in which the vertical axis (first axis) of the image corresponds to the longitudinal position and the horizontal axis (second axis) corresponds to the cross-sectional position, and assigns pixel values ​​to each pixel according to the magnetic property value at each position within the magnetic tube (step S14). Specifically, the control unit 11 calculates power values ​​of the magnetic property values ​​for multiple patterns with different exponents, assigns each power value to a different hue, and generates hue images for each hue. The control unit 11 then generates a composite image by combining the hue images. The control unit 11 generates composite images for all combinations of assigning each power value to each hue, and uses these as training data.

[0088] The control unit 11 generates an estimation model 50 that estimates thickness information when measurement data is input based on the generated composite image and the correct value of the thickness information (step S15). For example, the control unit 11 generates a neural network such as a CNN as the estimation model 50. The control unit 11 inputs the composite image generated in step S14 into the estimation model 50 to obtain an estimate of the thickness information, compares the obtained estimate with the correct value, and optimizes parameters such as weights of the estimation model 50 so that the two approximate each other. The control unit 11 then ends the series of processes.

[0089] 11 is a flowchart showing the procedure for estimating wall thickness information. The process for estimating wall thickness information using the estimation model 50 will be described with reference to FIG. 11. The control unit 11 of the server 1 receives, from the terminal 2, a selection input of the estimation model 50 to be used for estimating wall thickness information (step S31). Specifically, the control unit 11 receives a designation input specifying the wall thinning pattern, dimensions, and material of the magnetic pipe. The control unit 11 selects the estimation model 50 corresponding to the designated magnetic pipe information.

[0090] The control unit 11 acquires measurement data of the magnetic pipe from the terminal 2 (step S32). The control unit 11 performs preprocessing on the acquired measurement data (step S33). The control unit 11 converts the preprocessed measurement data into an image (step S34). Specifically, the control unit 11 assigns power values ​​of the magnetic characteristic value, each having a different exponent, to each hue, and generates a composite image by combining the images of each hue.

[0091] The control unit 11 estimates the thickness information of the magnetic pipe by inputting the image converted in step S34 into the estimation model 50 (step S35). For example, the control unit 11 estimates the thickness of each section obtained by dividing the magnetic pipe into fixed lengths along the longitudinal direction. The control unit 11 transmits the estimated thickness information to the terminal 2 (step S36), and the series of processes ends.

[0092] As described above, according to the first embodiment, the state of the magnetic pipe can be suitably estimated.

[0093] (Variation 1) In the first embodiment, a configuration was described in which measurement data of a magnetic pipe is converted into an image and input to the estimation model 50. In the first variation, a configuration will be described in which the original measurement data is used as is as input to the estimation model 50. Note that the same reference numerals are used to designate content that overlaps with the first embodiment, and description thereof will be omitted.

[0094] Fig. 12 is an explanatory diagram showing an overview of Modification 1. Fig. 12 illustrates how data on peak portions corresponding to thinned portions is extracted from measurement data of the magnetic pipe and the data on the peak portions is input to an estimation model 50 to estimate thickness information. An overview of this modification will be described with reference to Fig. 12.

[0095] In this modification, the server 1 constructs a model that can handle the original measurement data (magnetic property values ​​at each position in the magnetic pipe) as is, rather than an image, as the estimation model 50. For example, the server 1 constructs a model related to kNN (k-nearest neighbor method) as the estimation model 50.

[0096] In this modified example, the estimation model 50 is described as a model related to kNN, but the estimation model 50 may be any model that can handle the original measurement data, such as an SVM, a random forest, or a linear classifier.

[0097] The server 1 generates an estimation model 50 based on the same training data as in the first embodiment. Preferably, the server 1 performs principal component analysis or the like on the measurement data used for training to reduce the number of dimensions. Reducing the number of dimensions of the measurement data input to the estimation model 50 can reduce the processing load on the server 1. The user may be able to arbitrarily select whether or not to reduce the number of dimensions (principal component analysis).

[0098] Here, the server 1 may input the magnetic property values ​​at all positions along the longitudinal direction directly to the estimation model 50, but in this embodiment, only a portion of the measurement data is extracted and input to the estimation model 50. Specifically, the server 1 extracts data of peak portions corresponding to thinned portions and inputs the data to the estimation model 50.

[0099] 12 illustrates how peak data is extracted from the original measurement data. The server 1 extracts data on the portion where the output voltage is changing, i.e., the peak data, from the waveform of the output voltage of the Hall element 43, which corresponds to the measurement data.

[0100] The method for extracting peak data is not particularly limited. For example, the server 1 identifies a baseline (flat waveform) by taking a moving average of the output voltage, and identifies the peak based on the baseline. Specifically, the server 1 identifies a portion where the difference from the baseline is equal to or greater than a predetermined threshold as the peak. As shown in FIG. 12 , when the Hall element 43 approaches a thinning portion, the output voltage rises to a positive value and then drops to a negative value. In this modification, however, only a portion where the output voltage takes a positive value and the waveform is convex upward is identified as the peak.

[0101] The server 1 extracts data on the identified peak portion from the measurement data of the entire magnetic pipe and uses the extracted data as input to the estimation model 50. This makes it possible to reduce the amount of data processed by the estimation model 50 while suitably estimating thickness information on the thinned portion.

[0102] It is also possible to apply the data extraction of the peak portion in this modified example to the first embodiment, convert the input image to only the data of the peak portion, and use it as input to the estimation model 50 according to the first embodiment.

[0103] The server 1 extracts peak data from the training measurement data and generates an estimation model 50 based on the extracted peak data and the correct values ​​of thickness information at positions corresponding to the peaks. That is, the server 1 maps the peak data to which the correct values ​​of thickness information (correct labels indicating the correct thickness) are assigned to the feature space. The server 1 then constructs a dataset for estimating thickness information from the peak data by setting an arbitrary value k related to the k-nearest neighbor method.

[0104] When estimating thickness information, the server 1 extracts peak data from the measurement data acquired from the terminal 2 and maps it to the feature space. The server 1 then identifies k pieces of data (peak data to which correct labels have been assigned) located near the mapped position, and estimates the thickness information of the magnetic pipe to be estimated from the correct values ​​of the thickness information assigned to the k pieces of data.

[0105] 13 is a flowchart showing the steps of the process for generating the estimation model 50 according to the first modification. After expanding the number of measurement data (step S13), the server 1 executes the following process. The control unit 11 of the server 1 extracts data on peak portions corresponding to thinned portions from the measurement data on magnetic property values ​​at each position along the longitudinal direction of the magnetic pipe (step S201). Specifically, the control unit 11 identifies a baseline from the measurement data for the entire magnetic pipe, and extracts data on peak portions whose difference from the baseline is equal to or greater than a predetermined threshold.

[0106] The control unit 11 generates an estimation model 50 based on the peak portion data extracted in step S201 and the correct value of the thickness information (step S202). Specifically, the control unit 11 generates a model related to kNN. The control unit 11 maps the peak portion data to which the correct value of the thickness information (correct label) is assigned to a feature space and sets an arbitrary value k, thereby generating a data set as the estimation model 50 for estimating the thickness information from the peak portion data. The control unit 11 then ends the series of processes.

[0107] FIG. 14 is a flowchart showing the steps of the thickness information estimation process according to Modification 1. After preprocessing the measurement data (step S33), the server 1 executes the following process. The control unit 11 of the server 1 extracts data on peaks corresponding to thinned portions from the measurement data of magnetic property values ​​at each position along the longitudinal direction of the magnetic pipe (step S221). Specifically, the control unit 11 identifies a baseline from the measurement data of the entire magnetic pipe and extracts data on peaks whose difference from the baseline is equal to or greater than a threshold. The control unit 11 inputs the extracted peak data into the estimation model 50 to estimate thickness information on the thinned portions (step S222). Specifically, the control unit 11 maps the peak data to a feature space and estimates thickness information on the thinned portions of the magnetic pipe to be estimated from the correct values ​​of thickness information for k nearby data. The control unit 11 then proceeds to step S36.

[0108] As described above, according to Modification 1, it is possible to estimate thickness information from the original measurement data without imaging the measurement data. In particular, in this modification, by extracting data of peak portions corresponding to the thinned portions, it is possible to suitably estimate thickness information of the thinned portions.

[0109] (Modification 2) In this modification, an embodiment in which the estimation model 50 is generated using an ensemble learning technique will be described.

[0110] Fig. 15 is an explanatory diagram showing an overview of Modification 2. Fig. 15 illustrates how thickness information is estimated by extracting data for a plurality of data sections, each of which is a predetermined length of data section slightly shifted along the longitudinal direction of the magnetic tube, from the peak portion data described in Modification 1, and inputting the data for each section into an estimation model 50. An overview of this modification will be described based on Fig. 15.

[0111] In this modification, the server 1 uses an ensemble learning technique to generate the estimation model 50. Specifically, the server 1 uses a gradient boosting technique to generate a model related to a decision tree (for example, LightGBM) as the estimation model 50.

[0112] Gradient boosting is a type of ensemble learning, a technique for generating a final model by sequentially generating multiple weak classifiers (models). The server 1 generates one weak classifier (decision tree) using training data, and generates the next weak classifier based on the residual between the estimated value by the generated weak classifier and the correct value. The server 1 sequentially generates weak classifiers by referring to the gradient of a loss function defined by the error between the estimated value and the correct value, taking into account the learning result of the previous weak classifier, and generates a final classifier, i.e., an estimated model 50.

[0113] Although gradient boosting is used as the ensemble learning in this modification, other ensemble learning methods may be used, such as bagging, which generates multiple weak classifiers in parallel. Furthermore, ensemble learning may be applied to the estimation model 50 according to the first modification, and may be performed when generating a model such as a kNN or SVM.

[0114] The server 1 may input the peak portion data (or the measurement data of the entire magnetic tube) described in variant example 1 to the estimation model 50, but in this variant example, the peak portion is further divided into data sections of a predetermined length, and the data of each data section is input to the estimation model 50.

[0115] 15 illustrates how data for multiple data sections are extracted from the peak data. The data section is a section of measurement points along the longitudinal direction of the magnetic pipe, and constitutes part of the peak. For example, the server 1 considers 10 consecutive measurement points to be one data section. Note that the number of measurement points constituting a data section is not limited to 10.

[0116] The server 1 sequentially shifts the data section to be extracted along the longitudinal direction and extracts data for each data section from the peak portion. For example, the server 1 extracts data for 10 points from the start point of the peak portion where the difference from the baseline is equal to or greater than a threshold as data for the first data section. The server 1 shifts the start point of the data section to be extracted by a predetermined number (for example, one) at a time and extracts data for each data section. In this way, the server 1 extracts data for multiple data sections from the data for one peak portion.

[0117] When estimating thickness information for one peak portion, the server 1 estimates the thickness information by inputting the extracted data for the plurality of data sections into the estimation model 50. This makes it possible to more appropriately estimate the thickness of the thinned portion.

[0118] 16 is a flowchart showing the steps of the process for generating the estimation model 50 according to the second modification. After expanding the measurement data (step S13), the server 1 executes the following process. The control unit 11 of the server 1 extracts data of a peak portion corresponding to a thinned portion from the measurement data of the entire length of the magnetic pipe measured along the longitudinal direction (step S301). The control unit 11 then extracts data of multiple data sections from the data of the peak portion (step S302). Specifically, the control unit 11 sequentially shifts a data section of a predetermined length to be extracted along the longitudinal direction from the start point of the peak portion, and extracts data of each data section.

[0119] The control unit 11 generates an estimation model 50 based on the data of the multiple data intervals extracted in step S302 and the correct values ​​of the thickness information at the peak portions (thinned portions) (step S303). Specifically, the control unit 11 generates a decision tree using a gradient boosting technique. The control unit 11 generates one weak classifier (decision tree) using training data (the data of the multiple data intervals and the correct values ​​of the thickness information), and generates a next weak classifier based on the residual between the estimated value of the thickness information by the generated weak classifier and the correct value. The control unit 11 sequentially generates weak classifiers according to the gradient of a loss function defined by the residual between the estimated value and the correct value, and generates a final classifier, i.e., the estimation model 50. The control unit 11 ends the series of processes.

[0120] 17 is a flowchart showing the steps of the thickness information estimation process according to Modification 2. After preprocessing the measurement data (step S33), the server 1 executes the following process. The control unit 11 of the server 1 extracts data of peak portions corresponding to thinned portions from the measurement data of magnetic property values ​​at each position along the longitudinal direction (step S321). The control unit 11 extracts data of multiple data sections from the peak portion data (step S322). The control unit 11 inputs the extracted data of the multiple data sections into the estimation model 50 to estimate thickness information of the thinned portions (step S323). The control unit 11 then proceeds to step S36.

[0121] As described above, according to Modification 2, it is also possible to generate the estimation model 50 using ensemble learning related to gradient boosting. Furthermore, by dividing the data of the peak portion into data of a plurality of data sections and inputting the data into the estimation model 50, it is possible to more appropriately estimate the thickness information of the peak portion (thinned portion).

[0122] (Variation 3) An example using an ensemble learning method different from that of Variation 2 will be described. A known general ensemble learning method is to input the same data set to multiple models with different learning mechanisms, and use the average of the outputs from these multiple models or the majority output from these multiple models as the learning result.

[0123] In variant 3, unlike this type of learning, the data to be input is divided into multiple sets, each divided set is input to a different corresponding model, the output of each corresponding model is input to yet another model, and the output of that other model is used as the learning result.

[0124] Fig. 18 is an explanatory diagram showing an overview of Modification 3. As shown in Fig. 18, the estimation model in Modification 3 includes a thin-wall specialized model 50a (Bag Tree-A) trained mainly based on measurement data and thickness information in which the remaining wall thickness of the magnetic pipe is thinner than a predetermined value, a thick-wall specialized model 50b (Bag Tree-B) trained mainly based on measurement data and thickness information in which the remaining wall thickness of the magnetic pipe is thicker than a predetermined value, and a final model 50c (Bag Tree-C) that receives as input the outputs of the thin-wall specialized model 50a and the thick-wall specialized model 50b.

[0125] For example, the thin-wall specialized model 50a is a model that targets 90% of all data for which the thickness is 0.1 to 0.8 mm and 10% of all data for which the thickness is greater than 0.8 mm, and the thick-wall specialized model 50b is a model that targets the remaining data that is not used by the thin-wall specialized model 50a. The weighting relationship between the two models 50a and 50b in this example is shown in Figure 19.

[0126] Although the reference thickness for separating the thin-wall specialized model 50a from the thick-wall specialized model 50b is set to 0.8 mm, this reference thickness value is merely an example, and other numerical values ​​may be used. Furthermore, although 90% of the data whose thickness is equal to or less than a predetermined value and 10% of the data whose thickness is greater than the predetermined value are used as targets for the thin-wall specialized model 50a, this 90%:10% ratio is merely an example, and other ratios may be used.

[0127] 20 is a flowchart showing the steps of the process for generating estimation models (thin-wall specialized model 50a, thick-wall specialized model 50b, and final model 50c) in Modification 3. The control unit 11 of the server 1 acquires first training data including measurement data of the magnetic pipe, mainly thin-walled measurement data as described above, and correct values ​​of thickness information associated with the measurement data (step S401). Based on the acquired first training data (measurement data and correct values ​​of thickness information), the control unit 11 generates a thin-wall specialized model 50a that estimates and outputs first thickness information when measurement data is input (step S402).

[0128] The control unit 11 acquires second training data including the measurement data of the magnetic pipe, mainly the measurement data of the thick wall as described above, and the correct values ​​of the wall thickness information associated with the measurement data (step S403).The control unit 11 generates the thick-wall specialized model 50b that estimates and outputs the second wall thickness information when the measurement data is input, based on the acquired second training data (measurement data and the correct values ​​of the wall thickness information) (step S404).

[0129] The control unit 11 acquires third training data including the first thickness information output from the thin-wall specialization model 50a, the second thickness information output from the thick-wall specialization model 50b, and the correct values ​​of the associated thickness information (step S405).The control unit 11 generates a final model 50c that estimates and outputs thickness information when the first thickness information and the second thickness information are input based on the acquired third training data (the first and second thickness information and the correct values ​​of the thickness information) (step S406).

[0130] 21 is a flowchart showing the procedure for estimating wall thickness information in Modification 3. The control unit 11 of the server 1 receives, from the terminal 2, a selection input of an estimation model (thin wall specialized model 50a, thick wall specialized model 50b, and final model 50c) to be used for estimating wall thickness information (step S411). Specifically, the control unit 11 receives a designation input specifying a wall thickness boundary value that separates a thin wall model from a thick wall model, the proportion of thick walls included in a thin wall model, and the like. The control unit 11 selects an estimation model (thin wall specialized model 50a, thick wall specialized model 50b, and final model 50c) corresponding to the designation input.

[0131] The control unit 11 acquires measurement data of the magnetic pipe from the terminal 2 (step S412). The control unit 11 inputs the acquired measurement data into the thin-wall specialized model 50a and the thick-wall specialized model 50b (step S413). Next, the control unit 11 estimates thickness information of the magnetic pipe by inputting the output from the thin-wall specialized model 50a and the output from the thick-wall specialized model 50b into the final model 50c (step S414). The control unit 11 transmits the estimated thickness information to the terminal 2 (step S36), completing the series of processes.

[0132] In addition, in variant example 3, without using the final model 50c, the average of the output from the thin-wall specialized model 50a and the output from the thick-wall specialized model 50b may be calculated, and the calculated average value may be used as the estimated thickness information of the magnetic pipe.

[0133] According to the third modification, since different appropriate models are applied to the thin-wall side and the thick-wall side, it is possible to estimate thickness information more accurately. Generally, the amount of data is small on the thin-wall side, but even in such cases, it is possible to improve the accuracy of estimating thickness information on the thin-wall side where the amount of data is small.

[0134] (Variation 4) Variation 4 is an example of a technique for correcting measurement data used to train an estimation model using dynamic time warping to reduce the influence of Hall element voltage due to magnetic flux disturbance in the MFR method.

[0135] First, we will explain the effect of magnetic flux disturbance on the Hall element voltage in the MFR system. Figure 22 is an explanatory diagram illustrating this effect. As shown in the cross-sectional view of a magnetic pipe in Figure 22, assume that a deep localized thinning A has occurred in the magnetic pipe. When scanning the inside of such a magnetic pipe with an NFR inspection probe 4, the MFR waveform detected by the Hall element 43 should ideally be waveform B, depending on the change in the shape of the thinning. However, the actually detected MFR waveform may resemble waveform C, and a region independent of wall thickness (particularly the region surrounded by a circle) may appear near the thinned portion. This is because, as shown in the cross-sectional view of the magnetic pipe, the magnetic field lines are distorted around the thinned portion, creating a blank region of magnetic field lines (the region surrounded by the dashed line E).

[0136] In order to solve the above problems, the measurement data used to train the estimation model is corrected using a dynamic time warping method. Figures 23A and 23B are explanatory diagrams showing an overview of the dynamic time warping method in Modification 4. In Figures 23A and 23B, "a" shows the time series waveform of the measurement data (detected voltage), specifically, the time series waveform of the value obtained by the 3-peak method (in the case of 8 channels, the voltage value of the channel with the highest detected voltage and the voltage values ​​of the two channels on either side of that channel divided by 8). "b" shows the time series waveform of the wall thickness value as the actually measured target variable.

[0137] In the fourth modification, as shown in Fig. 23B, the time series points of the three peak values ​​(waveform a) are linked to the time series points of the measured thickness values ​​(waveform b) so that the sum of the Euclidean distances between the three peak values ​​(measurement data) and the measured thickness values ​​(thickness information) is minimized. For example, time series point i of waveform b changes to correspond to time series point i+2, which is shifted by two points on waveform a.

[0138] Fig. 24 shows the results of estimating wall thickness information by inputting measurement data into an estimation model trained using the corrected measurement data as described above. In Fig. 24, waveform A shows actual measurement data (MFR waveform), and wall thickness reduction shape B shows estimated wall thickness information (wall thickness reduction shape). In Fig. 24, the actual MFR waveform (corresponding to waveform C in Fig. 22) when dynamic time warping is not used is shown superimposed on waveform A by a dashed line, and the actual wall thickness reduction shape (corresponding to the wall thickness reduction shape shown in Fig. 22) is shown superimposed on estimated wall thickness reduction shape B by a dashed line.

[0139] In the fourth variant, the dynamic time warping method is used to correct the measurement data used to train the estimation model, thereby reducing the influence of the Hall element voltage caused by magnetic flux disturbance in the MFR method as described above, and enabling the thinning shape of the magnetic tube to be estimated with high accuracy.

[0140] (Modification 5) Modification 5 is an example in which the number of data items is increased by using a principal component analysis technique. Fig. 25 is an explanatory diagram of the data expansion process in Modification 5.

[0141] The server 1 may use, as input to the estimation model 50, raw measurement data measured by an inspection probe 4 having eight channels (CH1 to CH8) and first group data (in the example shown in Figure 25, including moving averages, moving standard deviations, moving skewness, moving kurtosis, and moving differences at multiple positions (10 to 50 points) along the longitudinal direction of the magnetic tube) obtained by performing preprocessing including arithmetic processing on this raw measurement data (see Figure 25(a)). However, in variant example 5, in order to improve the estimation accuracy of the thickness information, in addition to these raw measurement data and first group data, second group data obtained by performing principal component analysis on the raw measurement data and first group data is used as input to the estimation model 50 (see Figure 25(b)).

[0142] 25, the first group data and the second group data are moving averages, moving standard deviations, moving skewness, moving kurtosis, and moving differences, but are not limited to these. Also, although the second group data has the same data structure as the first group data, the data structures of the two groups may be different.

[0143] In the fifth modification, the number of data items is increased by adding the results of the principal component analysis (in the example shown in FIG. 25, the number of data items is increased from 169 variables to 338 variables), thereby improving the accuracy of estimating the thickness information.

[0144] (Embodiment 2) In the first embodiment, a defect estimation system was described that estimates defects such as thinning of a magnetic pipe using the estimation model 50. In the present embodiment, an embodiment in which a user actually uses this system to inspect defects in a magnetic pipe will be described.

[0145] Fig. 26 is an explanatory diagram showing an example of a measurement data upload screen. Fig. 26 illustrates a display screen when the terminal 2 transmits (uploads) measurement data to the server 1. The upload screen includes a model selection field 171, an upload button 172, and a provision / non-provision selection field 173.

[0146] The model selection field 171 is an input field for selecting the estimation model 50 used to estimate the wall thickness information, and is an input field for accepting specified input of magnetic pipe information related to the magnetic pipe that is the estimation target (measurement target). For example, the model selection field 171 includes a wall thinning form specification field 1711, a dimension specification field 1712, and a material specification field 1713. The terminal 2 accepts specified input of the wall thinning form, dimensions, and material of the magnetic pipe via each selection field.

[0147] The upload button 172 is a button for transmitting (uploading) the measurement data to the server 1. When an operation input to the upload button 172 is accepted, the terminal 2 transmits the measurement data to the server 1.

[0148] The provision selection field 173 is an input field for selecting whether or not to provide (use) the measurement data to this system. The terminal 2 accepts a selection input regarding whether or not to provide the measurement data in response to an operation input into the provision selection field 173. When a selection input to provide the measurement data is accepted, the server 1 stores the measurement data acquired from the terminal 2 in the measurement DB 143. As will be described later, the measurement data provided by the user is used to re-learn (update) the estimation model 50.

[0149] Regardless of whether provision is selected or not, all measurement data acquired from each terminal 2 may be used for re-learning the estimation model 50 .

[0150] Fig. 27 is an explanatory diagram showing an example of a display screen for wall thickness information. Fig. 27 illustrates a display screen for wall thickness information estimated based on the estimation model 50. The screen includes a magnetic pipe selection field 181, a wall thickness graph 182, a cross-sectional image 184, a download button 185, and a second measurement data upload button 186.

[0151] The magnetic pipe selection field 181 is an input field for selecting a magnetic pipe for which thickness information is to be displayed. The terminal 2 displays the thickness information of the magnetic pipe selected in the magnetic pipe selection field 181.

[0152] The wall thickness graph 182 is a graph showing the wall thickness of each portion of the magnetic pipe estimated based on the estimation model 50. In the wall thickness graph 182, the horizontal axis represents the longitudinal position of the magnetic pipe, and the vertical axis represents the wall thickness. For example, the terminal 2 displays the average value of the wall thickness for each channel on the wall thickness graph 182. Note that the terminal 2 may also display the wall thickness for each channel on the wall thickness graph 182.

[0153] The cross-sectional image 184 is an image that simulates a cross-section perpendicular to the longitudinal direction of the magnetic pipe, and reproduces the wall thickness at each position on the cross-section according to the wall thickness information. As described above, the server 1 estimates the wall thickness of each channel (each position on the cross-section) at each position along the longitudinal direction of the magnetic pipe. The server 1 compares the estimated wall thickness of each channel with the wall thickness of the healthy portion (the baseline of the wall thickness graph 182) to detect thinned portions that are thinner than the healthy portion. The server 1 then calculates the wall thickness reduction width, which is the difference in wall thickness between the healthy portion and the thinned portion. For example, the terminal 2 displays a cross-section of the magnetic pipe with CH1 oriented at 0 o'clock, and reproduces and displays a thinned portion (recess) according to the calculated wall thickness reduction width at a position corresponding to the channel where the thinned portion was detected (CH1 in FIG. 27).

[0154] For example, terminal 2 displays line 183 on wall thickness graph 182. Terminal 2 accepts an operation to move line 183 along the horizontal axis, thereby accepting a designation input that specifies a longitudinal position of the magnetic pipe at which to display cross-sectional image 184. Terminal 2 displays cross-sectional image 184 of the magnetic pipe corresponding to the position of line 183.

[0155] In addition, the terminal 2 displays the number of thinned portions detected based on the thickness information, the maximum value and average value of the thinned width, etc.

[0156] Note that the display screen for wall thickness information shown in FIG. 27 is merely an example, and the present embodiment is not limited thereto. For example, the terminal 2 may simply display the wall thickness at each position within the magnetic pipe, as in the wall thickness graph 182. Furthermore, for example, when displaying the wall thickness at each position within the magnetic pipe in the wall thickness graph 182 or the like, the terminal 2 may change the display mode (e.g., display color) of the wall thickness at each position depending on the reliability of the wall thickness estimation result at the corresponding position by the estimation model 50 (a probability value representing the likelihood of the estimation result). Furthermore, for example, if the terminal 2 detects multiple thinned portions from the wall thickness information, it may display a list of multiple cross-sectional images 184, 184, 184, ... corresponding to each thinned portion. Thus, the display mode of the wall thickness information is not limited to the mode shown in FIG. 19 .

[0157] In addition, in this embodiment, terminal 2 displays thickness information on a dedicated screen, but terminal 2 may simply download a file (e.g., a CSV file) storing the estimated results of thickness information, without displaying the thickness information.

[0158] The download button 185 is a button for downloading a file storing the estimation result of the wall thickness information. When an operation input to the download button 185 is received, the terminal 2 obtains the file storing the estimation result of the wall thickness information from the server 1.

[0159] The second measurement data upload button 186 is a button for uploading second measurement data, which is measurement data other than the above-mentioned measurement data (measurement data based on the magnetic flux resistivity method), to the server 1. When an operation input to the second measurement data upload button 186 is accepted, the terminal 2 transmits measurement data measured by the water immersion rotational ultrasonic thickness measurement method (IRIS) to the server 1 as the second measurement data. The second measurement data will be described later.

[0160] As described above, the server 1 acquires the measurement data of the magnetic pipe from each user's terminal 2 and estimates the wall thickness information. The server 1 then transmits the wall thickness information to the terminal 2 from which the measurement data was acquired and causes the terminal 2 to display the information. In this embodiment, the server 1 re-learns based on the measurement data acquired from each terminal 2 and updates the estimation model 50.

[0161] Specifically, when the server 1 receives a selection input indicating that the measurement data will be provided in the provision / non-provision selection field 173, the server 1 stores the measurement data acquired from the terminal 2 in the measurement DB 143. The server 1 uses the measurement data stored in the measurement DB 143 as training data for relearning. The server 1 updates the estimation model 50 based on the measurement data stored in the measurement DB 143 and the correct value of the wall thickness information corresponding to the measurement data.

[0162] The correct value of the thickness information during relearning may be manually set by the user via the screen shown in Fig. 27, but in this embodiment, the second measurement data provided by the user is used as the correct value. The second measurement data is data obtained by measuring the thickness at each position inside the magnetic pipe, such as measurement data related to the IRIS.

[0163] IRIS is a method for inspecting the wall thickness of a tube (magnetic tube) using ultrasound. A probe equipped with an ultrasonic probe is inserted into a tube filled with water, and the ultrasonic probe detects the reflected waves of an ultrasonic beam generated by the probe, thereby measuring the wall thickness. Since IRIS itself is well known, a detailed description thereof will be omitted in this embodiment. Compared to inspection methods that utilize magnetic properties, such as the magnetic flux resistance method according to this embodiment, RFECT, and MFL, IRIS has a slower inspection speed, but has the advantage of relatively high measurement accuracy.

[0164] In this embodiment, when the server 1 receives measurement data related to IRIS from the user, the server 1 uses the data as the correct value for the thickness information. That is, the server 1 sets a pair of the measurement data based on the magnetic flux resistivity method and the second measurement data based on IRIS as training data for relearning.

[0165] The server 1 compares the estimated value of the wall thickness information based on the measurement data provided by the user with the wall thickness indicated by the second measurement data, i.e., the correct value. The server 1 then updates parameters such as the weights between neurons so that the two values ​​approximate each other. The server 1 re-learns the measurement data of each magnetic pipe provided by the user and updates the estimation model 50.

[0166] The server 1 may retrain the measurement data of all users into a single estimation model 50, but it is more preferable to train the measurement data acquired from each user's terminal 2 separately and update the estimation model 50 for each user. That is, the server 1 provides the measurement data provided by each user to the estimation model 50 separately, and builds an estimation model 50 for each user. This allows the estimation model 50 to be individually tuned according to the tendencies of the magnetic pipes measured (inspected) by each user.

[0167] As described above, the estimation accuracy of the estimation model 50 can be improved through operation of this system.

[0168] The server 1 charges each user of the system a usage fee (fee) for the estimation model 50. The usage fee may be, for example, a fixed amount (subscription) per fixed period, but in this embodiment, it is determined based on the amount of calculation required for the thickness information estimation process based on the estimation model 50. The amount of calculation used as the basis for determining the usage fee is, for example, the number of times thickness information is estimated (the number of magnetic tubes), but it may also be the calculation time required for the estimation process, the amount of measurement data, etc. When the server 1 obtains measurement data from the terminal 2 and estimates thickness information, it determines the usage fee to be charged to the user based on the amount of calculation required for the estimation. The server 1 stores the determined usage fee in the user DB 141 and ultimately bills the user for the usage fee.

[0169] In this embodiment, the server 1 changes the usage fee depending on whether or not measurement data is provided (used). Specifically, the server 1 deducts the usage fee when providing measurement data (measurement data based on the magnetic flux resistivity method) that was the basis for estimating wall thickness information. Furthermore, the server 1 further deducts the usage fee when providing second measurement data (measurement data based on IRIS) corresponding to the measurement data. This allows for efficient collection of training data for re-learning.

[0170] FIG. 28 is a flowchart illustrating the steps of the wall-thickness information estimation process according to the second embodiment. The processing details of the server 1 and terminal 2 according to this embodiment will be described with reference to FIG. 28 . The control unit 21 of the terminal 2 accepts a selection input of the estimation model 50 used to estimate the wall-thickness information (step S501). Specifically, the control unit 21 accepts a designation input of magnetic pipe information related to the magnetic pipe to be estimated, such as the wall-thickness thinning pattern, dimensions, and material of the magnetic pipe. The control unit 21 also accepts a selection input of whether or not to use (provide) measurement data related to updating the estimation model 50 (step S502). The control unit 11 transmits the measurement data of the magnetic pipe along with the selections made in steps S501 and S502 to the server 1 (step S403). When the measurement data is acquired from the terminal 2, the control unit 11 of the server 1 executes the processes of steps S33 to S35 and transmits wall-thickness information estimated based on the estimation model 50 to the terminal 2 (step S36).

[0171] When the thickness information is acquired from the server 1, the control unit 21 of the terminal 2 displays the estimated thickness information on the display unit 24 (step S504). Specifically, the control unit 21 displays the thickness of each position (portion) along the longitudinal direction of the magnetic tube using a thickness graph 182 or the like, and also displays the thickness of each position (channel) on a cross section at any position in the longitudinal direction using a cross-sectional image 184 or the like.

[0172] In response to an operation input from the user, the control unit 21 transmits second measurement data, which is obtained by measuring the wall thickness of the magnetic tube using a method other than the magnetic flux resistivity method, to the server 1 (step S505). The second measurement data is, for example, measurement data related to IRIS, and is data obtained by inserting a probe equipped with an ultrasonic probe into a magnetic tube filled with water and detecting reflected waves of an ultrasonic beam generated from the probe.

[0173] After executing the process of step S36, the control unit 11 of the server 1 determines the usage fee (charge) to be charged to the user (step S506). Specifically, the control unit 11 determines the usage fee depending on the amount of calculation (number of calculations, calculation time, amount of measurement data, etc.) of the wall thickness information estimation process based on the estimation model 50. Furthermore, the control unit 11 deducts the usage fee depending on whether the measurement data and the second measurement data are provided (used or acquired).

[0174] When the control unit 11 receives a selection input from the user indicating that the measurement data can be provided, the control unit 11 stores the measurement data transmitted in step S403 in the measurement DB 143 (step S507). Furthermore, when the second measurement data is transmitted in step S405, the control unit 11 stores the second measurement data in the measurement DB 143 in association with the measurement data.

[0175] The control unit 11 updates the estimation model 50 based on the measurement data and the second measurement data stored in the measurement DB 143 (step S508). That is, the control unit 11 uses the measurement data measured by the user as input data for training, and uses the measured thickness values ​​in the second measurement data as correct values ​​for the thickness information, and updates parameters such as weights between neurons. For example, the control unit 11 individually learns the measurement data acquired from each user's device 2 and updates the estimation model 50 for each user. The control unit 11 then completes the series of processes.

[0176] As described above, according to the second embodiment, the present system can be preferably implemented, and the estimation model 50 can be optimized through the operation of the present system.

[0177] The embodiments disclosed herein are illustrative in all respects and should not be considered limiting. The scope of the present invention is defined by the claims, not by the above meaning, and is intended to include all modifications within the meaning and scope of the claims.

[0178] DESCRIPTION OF SYMBOLS 1 Server (information processing device) 11 Control unit 12 Main memory unit 13 Communication unit 14 Auxiliary memory unit P1 Program 142 Model DB 143 Measurement DB 2 Terminal (user terminal) 21 Control unit 22 Main memory unit 23 Communication unit 24 Display unit 25 Input unit 26 Auxiliary memory unit P2 Program 3 Measuring device 4 Inspection probe 41 Yoke 42 Magnet 43 Hall element (magnetic sensor) 50 Estimation model 50a Thin-wall specialized model 50b Thick-wall specialized model 50c Final model

Claims

1. An information processing method in which a computer executes a process of acquiring measurement data that measures the magnetic property values ​​of a magnetic pipe, and inputting the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input, thereby estimating the thickness information.

2. The information processing method according to claim 1, wherein the magnetic characteristic value is a measurement value measured using an inspection probe comprising a magnet that generates a magnetic field, a yoke that is positioned on the opposite side of the magnet from the magnetic tube, and a magnetic sensor that is positioned between the yoke and the magnetic tube and measures the magnetic flux density passing through the yoke, magnet, and magnetic tube, and is the output voltage of the magnetic sensor that is proportional to the magnetic flux density.

3. The information processing method according to claim 2, wherein the measurement data is data obtained by measuring the magnetic characteristic values ​​at each position along the longitudinal direction of the cylindrical magnetic pipe, by using the inspection probe having the magnets and magnetic sensors periodically attached on the outer periphery of the yoke, at each position obtained by equally dividing the cross section of the magnetic pipe along the circumferential direction; the magnetic characteristic values ​​are normalized so that the magnetic characteristic values ​​outside the magnetic pipe are consistent at all positions along the circumferential direction, and so that the magnetic characteristic values ​​in a healthy part of the magnetic pipe where no thinning has occurred are consistent at all positions along the circumferential direction; and the measurement data with the normalized magnetic characteristic values ​​is input into the model to estimate the wall thickness information.

4. The information processing method according to claim 2 or 3, wherein the measurement data is data obtained by measuring the magnetic characteristic values ​​at each position obtained by dividing the cross section of the cylindrical magnetic tube equally along the circumferential direction using the inspection probe having the magnets and magnetic sensors attached periodically on the outer periphery of the yoke, the magnetic characteristic values ​​at each position are corrected to values ​​when the inspection probe passes through the central axis of the magnetic tube, and the measurement data with the corrected magnetic characteristic values ​​is input into the model to estimate the wall thickness information at each position.

5. The information processing method according to any one of claims 2 to 4, wherein the measurement data is data obtained by measuring the magnetic property values ​​at each position along the longitudinal direction of the cylindrical magnetic tube, by using the inspection probe in which the magnets and magnetic sensors are periodically attached on the outer periphery of the yoke, and dividing the cross section of the magnetic tube into equal parts along the circumferential direction, and the method calculates the average value, standard deviation, skewness or kurtosis of the magnetic property values ​​along the longitudinal direction or circumferential direction from the magnetic property values ​​at each position along the longitudinal direction or circumferential direction, and estimates the wall thickness information by inputting the measurement data to which the average value, standard deviation, skewness or kurtosis has been added into the model.

6. An information processing method according to any one of claims 1 to 5, wherein the measurement data is data obtained by measuring the magnetic characteristic values ​​at each position along the longitudinal direction of the magnetic tube, and the measurement data is input into the model to estimate the thickness information of each portion obtained by dividing the magnetic tube into fixed lengths along the longitudinal direction.

7. The information processing method according to claim 6, further comprising: determining a baseline of the measurement data by taking a moving average of the magnetic property values ​​along the longitudinal direction; extracting data of peak portions from the measurement data where the difference between the baseline and the magnetic property values ​​is equal to or greater than a predetermined threshold; and inputting the extracted data of the peak portions into the model to estimate the thickness information of the peak portions.

8. The information processing method according to claim 7, wherein data for each data section of a predetermined length shifted from each other along the longitudinal direction is extracted from the measurement data, and the extracted data for each data section is input into the model, thereby estimating thickness information for the peak portion.

9. An information processing method according to any one of claims 1 to 8, wherein the measurement data is data obtained by measuring the magnetic property values ​​at each position along the longitudinal direction of the magnetic tube on a cross section perpendicular to the longitudinal direction, the measurement data is converted into an image in which a pixel value of each pixel is assigned according to the magnetic property value at each position of the magnetic tube, with a first axis of the image being the position along the longitudinal direction and a second axis being the position on the cross section, and the image is input into the model to estimate the wall thickness information.

10. The information processing method according to claim 9, further comprising: generating a plurality of hue images from the measurement data in which power values ​​of the magnetic characteristic values ​​of a plurality of patterns with different exponents are assigned to different hues; generating a composite image by combining the plurality of hue images; and inputting the generated composite image into the model to estimate the thickness information.

11. An information processing method according to any one of claims 1 to 10, comprising: acquiring magnetic pipe information relating to the magnetic pipe to be measured in the measurement data; selecting a model corresponding to the acquired magnetic pipe information from among a plurality of models which have learned different training data according to the magnetic pipe information; and inputting the measurement data into the selected model to estimate the wall thickness information.

12. The information processing method according to any one of claims 1 to 11, wherein the models include a first model trained mainly based on measurement data and thickness information where the remaining thickness of the magnetic pipe is thinner than a predetermined value, a second model trained mainly based on measurement data and thickness information where the remaining thickness of the magnetic pipe is thicker than the predetermined value, and a third model trained to output thickness information when thickness information output from the first model and thickness information output from the second model are input, and the thickness information is estimated by inputting the acquired measurement data into the first model and the second model and inputting the outputs from the first model and the second model into the third model.

13. An information processing method according to any one of claims 1 to 12, wherein the measurement data is corrected so that the sum of Euclidean distances between the measurement data and thickness information corresponding to the measurement data is minimized, and the model is trained based on the corrected measurement data and thickness information.

14. An information processing method according to any one of claims 1 to 13, wherein the measurement data includes sensor data obtained from a plurality of sensors provided circumferentially around the magnetic tube, first group data obtained by performing preprocessing including arithmetic processing on the sensor data, and second group data obtained by performing principal component analysis on the sensor data and first group data, and the wall thickness information is estimated by inputting the sensor data, first group data, and second group data into the model.

15. A program that causes a computer to execute a process of acquiring measurement data that measures the magnetic property values ​​of a magnetic pipe, and inputting the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input, thereby estimating the thickness information.

16. An information processing device comprising: an acquisition unit that acquires measurement data that measures the magnetic characteristic values ​​of a magnetic pipe; and an estimation unit that estimates thickness information by inputting the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input.

17. A model generation method in which a computer executes the following process: acquiring training data in which correct values ​​of thickness information regarding the thickness of a magnetic pipe are associated with measurement data that measures the magnetic property values ​​of the magnetic pipe; and generating a trained model that estimates the thickness information when the measurement data is input based on the training data.

18. A model generation method as described in claim 17, wherein the measurement data is data obtained by measuring the magnetic characteristic values ​​at each position obtained by dividing the cross section of the cylindrical magnetic tube equally along the circumferential direction, and the positions on the cross section of the magnetic tube where the magnetic characteristic values ​​were measured are shifted along the circumferential direction to generate multiple patterns of measurement data from the training data, and the trained model is generated using the multiple patterns of measurement data.

19. A model generation method according to claim 17 or 18, which generates multiple patterns of measurement data from the training data by adding multiple patterns of predetermined noise, and generates the trained model using the multiple patterns of measurement data.

20. A model generation method according to any one of claims 17 to 19, wherein the measurement data is data obtained by measuring the magnetic characteristic values ​​at each position along the longitudinal direction of the magnetic tube, and a plurality of patterns of the measurement data are generated from the training data by linearly interpolating the magnetic characteristic values ​​between two consecutive points along the longitudinal direction, and the trained model is generated using the plurality of patterns of measurement data.

21. A model generation method according to any one of claims 17 to 20, comprising: acquiring first training data mainly including measurement data and thickness information indicating that the remaining wall thickness of the magnetic pipe is thinner than a predetermined value; generating a first model based on the acquired first training data that outputs first thickness information when measurement data is input; acquiring second training data mainly including measurement data and thickness information indicating that the remaining wall thickness of the magnetic pipe is thicker than the predetermined value; generating a second model based on the acquired second training data that outputs second thickness information when measurement data is input; acquiring third training data including the first thickness information output from the first model, the second thickness information output from the second model, and thickness information; and generating a third model based on the acquired third training data that outputs thickness information when the first thickness information and the second thickness information are input.

22. A model generation method according to any one of claims 17 to 21, comprising: acquiring measurement data and thickness information corresponding to the measurement data; correcting the measurement data so that the sum of Euclidean distances between the acquired measurement data and thickness information is minimized; and training the model based on the corrected measurement data and thickness information.

23. An information processing method in which a computer executes the following processes: acquiring measurement data measuring the magnetic property values ​​of a magnetic pipe from each user terminal connected to the computer via a network; estimating thickness information by inputting the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input; and transmitting the estimated thickness information to the user terminal from which the measurement data was acquired.

24. The information processing method described in claim 23, wherein the magnetic characteristic value is a measurement value measured using an inspection probe comprising a magnet that generates a magnetic field, a yoke that is positioned on the opposite side of the magnet from the magnetic tube, and a magnetic sensor that is positioned between the yoke and the magnetic tube and measures the magnetic flux density passing through the yoke, magnet, and magnetic tube, and is the output voltage of the magnetic sensor that becomes lower as the magnetic flux density increases.

25. An information processing method according to claim 23 or 24, comprising: receiving specified input of magnetic pipe information relating to the magnetic pipe to be measured; selecting a model corresponding to the acquired magnetic pipe information from among a plurality of models which have learned different training data according to the magnetic pipe information; and estimating the wall thickness information by inputting the measurement data into the selected model.

26. An information processing method according to any one of claims 23 to 25, comprising storing the measurement data acquired from each user terminal in a memory unit, acquiring a correct value of the thickness information corresponding to the measurement data, and updating the model based on the measurement data stored in the memory unit and the correct value.

27. The information processing method according to claim 26, wherein the model is updated for each user who is the source of the measurement data using the measurement data acquired from the user terminal of that user.

28. An information processing method according to claim 26 or 27, wherein, when the thickness information is estimated, a usage fee for the model to be charged to the user is determined according to the amount of calculation required for the estimation process of the thickness information using the model.

29. An information processing method as described in claim 28, wherein, when the measurement data is obtained from the user terminal, a selection input is received from the user as to whether or not the measurement data can be used to update the model, and if a selection input is received indicating that the measurement data can be used, the measurement data is stored in the memory unit, and a usage fee for the model to be charged to the user is deducted depending on whether or not the measurement data can be used.

30. An information processing method as described in claim 28 or 29, further comprising: acquiring second measurement data from the user terminal, the second measurement data representing the correct value, the second measurement data being obtained by measuring the wall thickness of the magnetic tube using a measurement method different from that of the first measurement data; updating the model based on the second measurement data and the second measurement data; and deducting a usage fee for the model charged to the user depending on whether or not the second measurement data has been acquired.

31. The information processing method according to claim 8, wherein the second measurement data is measurement data obtained by a water immersion rotary ultrasonic thickness measurement method.

32. An information processing device comprising: an acquisition unit that acquires measurement data that measures the magnetic property values ​​of a magnetic pipe from each user terminal that is communicatively connected via a network; an estimation unit that estimates thickness information by inputting the acquired measurement data into a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input; and a transmission unit that transmits the estimated thickness information to the user terminal that is the source of the measurement data.

33. A program that causes a computer to execute the following processes: acquire measurement data that measures the magnetic property values ​​of a magnetic pipe from a measuring device; transmit the measurement data to an information processing device that estimates thickness information using a model that has been trained to estimate thickness information regarding the thickness of the magnetic pipe when the measurement data is input; acquire the estimated thickness information from the information processing device by inputting the measurement data into the model; and display the acquired thickness information on a display unit.

34. The program according to claim 11, wherein second measurement data representing the correct value of the wall thickness information, the second measurement data being obtained by measuring the wall thickness of the magnetic tube using a measurement method different from the first measurement data, is transmitted to the information processing device, and the wall thickness information estimated using the model updated based on the first measurement data and the second measurement data is obtained from the information processing device.

35. The program described in claim 33 or 34, wherein the measurement data is data obtained by measuring the magnetic property values ​​at each position on a cross section perpendicular to the longitudinal direction at each position along the longitudinal direction of the magnetic tube, displays a graph showing the thickness at each position along the longitudinal direction based on the thickness information showing the thickness at each position on the cross section at each position along the longitudinal direction, accepts a specified input specifying a position along the longitudinal direction on the graph, and displays a cross-sectional image of the magnetic tube that reproduces the thickness at each position on the cross section at the specified position.