Corrosion classification system and corrosion classification program

Abstract

This invention provides a corrosion classification system and program that can improve quantitative accuracy and objectivity in corrosion diagnosis, as well as enhance safety and efficiency. [Solution] The corrosion classification system 100 comprises a model generation unit 32 that generates a model for classifying metal-containing materials according to different corrosion degrees, which is obtained by learning the relationship between images of a plurality of metal-containing materials acquired in advance, sensor output values ​​acquired in advance when physical quantities related to corrosion are detected by a sensor 20 for each of the plurality of metal-containing materials, and the corrosion degree of the plurality of metal-containing materials acquired in advance; a modality acquisition unit 31 that acquires images of the metal-containing materials to be classified and sensor output values ​​acquired when physical quantities related to corrosion are detected by a sensor 20 for the metal-containing materials to be classified; and a corrosion classification unit 33 that classifies the metal-containing materials to be classified according to different corrosion degrees by applying the acquired images and sensor output values ​​to the generated model.

Description

[Technical Field]

[0001] This invention relates to a system and program for classifying metal-containing materials according to different degrees of corrosion. [Background technology]

[0002] Traditionally, visual inspections have been frequently used to diagnose corrosion in steel structures. To properly perform visual inspections at high places such as bridges, a device described in Patent Document 1 is known. This device comprises a long, extendable pipe and a camera fixed to its end. A pipe sway-preventing arm is also provided. This can improve work efficiency during visual inspections at high places. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 9-312837 [Overview of the project]

[0004] Incidentally, there is considerable room for improvement in quantitative accuracy and objectivity in diagnoses that rely on the aforementioned visual inspections and long-term exposure tests. The situation has arisen where the timing of repairs has been determined based on empirical rules. Therefore, there is a need for technologies that improve the reliability of inspections and appropriately determine the timing of repairs, painting, etc.

[0005] Furthermore, diagnostic methods that rely on the aforementioned visual inspections and long-term exposure tests have significant room for improvement in terms of safety and efficiency, particularly in inspections of high places and confined spaces. This makes it difficult to grasp the overall extent of corrosion in wide-area infrastructure. There are also issues such as increased costs associated with scaffolding installation and personnel deployment. In addition, there is a desire for real-time monitoring of corrosion progression, long-term deterioration prediction, and optimization of maintenance plans.

[0006] In view of the above, the object of the present invention is to provide a corrosion classification system and a corrosion classification program that can improve quantitative accuracy and objectivity in corrosion diagnosis, as well as improve safety and efficiency. [Means for solving the problem]

[0007] The technical part of the present invention for solving this technical problem is characterized by the following: The corrosion classification system of the present invention is a model obtained by learning the relationship between images of a plurality of metal-containing materials obtained in advance, sensor output values ​​obtained in advance when physical quantities related to corrosion are detected by sensors for the plurality of metal-containing materials, and the degree of corrosion of the plurality of metal-containing materials obtained in advance, and comprises: a model generation unit that generates a model for classifying metal-containing materials according to different degrees of corrosion; a modality acquisition unit that acquires images of metal-containing materials to be classified and sensor output values ​​obtained when physical quantities related to corrosion are detected by sensors for the metal-containing materials to be classified, respectively; and a corrosion classification unit that classifies the metal-containing materials to be classified according to different degrees of corrosion by applying the acquired images and the acquired sensor output values ​​to the generated model. The model generation unit uses, as images of the plurality of metal-containing materials, images obtained by applying the same amount of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after different exposure times, and / or images obtained by applying different amounts of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after the same exposure time. .

[0009] In the corrosion classification system of the present invention, the model generation unit generates a model that classifies the metal-containing material into at least two classification numbers according to different exposure times or different amounts of salt adhesion, and the corrosion classification unit classifies the metal-containing material to be classified according to different exposure times or different amounts of salt adhesion so as to correspond to the classification numbers of the model.

[0010] In the corrosion classification system of the present invention, the sensor output value is a value output by using at least one selected from among an ACM sensor, an ICM sensor, a surface salinity meter, an atmospheric corrosive gas monitor sensor, a wetness sensor, a rainfall and / or condensation sensor, a temperature and / or humidity sensor, a dissolved oxygen meter, a pH meter, an ultrasonic thickness gauge, a chloride ion meter, an infrared thermograph, an electrochemical measuring device, and a device used for gravimetric analysis.

[0011] The corrosion classification program of the present invention is a model obtained by having a computer learn the relationship between images of a plurality of metal-containing materials acquired in advance, sensor output values ​​acquired in advance when physical quantities related to corrosion are detected by sensors for each of the plurality of metal-containing materials, and the degree of corrosion of the plurality of metal-containing materials acquired in advance, and comprises: a model generation step of generating a model that classifies metal-containing materials according to different degrees of corrosion; a modality acquisition step of acquiring images of the metal-containing materials to be classified and sensor output values ​​acquired when physical quantities related to corrosion are detected by sensors for the metal-containing materials to be classified; and a corrosion classification step of classifying the metal-containing materials to be classified according to different degrees of corrosion by applying the acquired images and acquired sensor output values ​​to the generated model. The model generation step uses, as images of the plurality of metal-containing materials, images obtained by attaching the same amount of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after different exposure times, and / or images obtained by attaching different amounts of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after the same exposure time.

[0013] In the corrosion classification program of the present invention, the model generation step generates a model that classifies the metal-containing material into at least two classification numbers depending on different exposure times or different amounts of salt deposition, and the corrosion classification step classifies the metal-containing material to be classified according to different exposure times or different amounts of salt deposition, corresponding to the classification numbers of the model.

[0014] In the corrosion classification program of the present invention, the sensor output value is a value output by at least one selected from the group consisting of an ACM sensor, an ICM sensor, a surface salinity meter, an atmospheric corrosive gas monitor sensor, a wetness sensor, a rainfall and / or dew sensor, a temperature and / or humidity sensor, a dissolved oxygen meter, a pH meter, an ultrasonic thickness gauge, a chloride ion meter, infrared thermography, an electrochemical measurement device, and a device for gravimetric analysis.

Advantages of the Invention

[0015] According to the present invention, in the diagnosis of corrosion, quantification and objectivity can be improved, and safety and efficiency can also be improved. Therefore, the reliability of inspection can be improved, and the timing of repair, painting, etc. can be appropriately determined. In addition, it can contribute to improving the safety of infrastructure, reducing life cycle costs, and sustainable asset management.

Brief Description of the Drawings

[0016] [Figure 1] It is an overall schematic diagram of a corrosion classification system according to an embodiment of the present invention. [Figure 2] It is a diagram for explaining an example of the mode of the sensor shown in FIG. 1. [Figure 3] It is a diagram for explaining another example of the mode of the sensor shown in FIG. 1. [Figure 4] It is a diagram for explaining an outline of a classification scheme for classifying according to different amounts of salt deposition in the classification unit shown in FIG. 1. [Figure 5] It is a diagram for explaining an outline of a classification scheme for classifying according to different exposure times in the classification unit shown in FIG. 1. [Figure 6] It is a diagram for explaining an example of the mode of a galvanized steel sheet used when generating a model in the model generation unit shown in FIG. 1. [Figure 7] It is a diagram for explaining an example of the mode of an image obtained by imaging the surface of a galvanized steel sheet used when generating a model in the model generation unit shown in FIG. 1. [Figure 8] It is a flowchart showing the processing of a program executed by the terminal shown in FIG. 1. [Figure 9] This is a diagram illustrating an overview of a comparative example scheme, which classifies samples according to different salt deposition amounts when only images are used as the modality. [Modes for carrying out the invention]

[0017] Embodiments of the present invention will be described below with reference to the drawings.

[0018] As shown in Figure 1, the corrosion classification system 100 according to an embodiment of the present invention comprises an imaging device 10, a sensor 20, and a terminal 30. The imaging target of the imaging device 10 and the detection target of the sensor 20 are metal-containing materials. The imaging device 10 and the sensor 20 are each connected to the terminal 30 in a communication manner. The communication method may be wireless or wired. The image from the imaging device 10 and the sensor output value from the sensor 20 are transmitted to the terminal 30, respectively.

[0019] The imaging device 10 is a camera, scanner, etc. The imaging device 10 can be any form as long as it is capable of imaging the surface of a metal-containing material and outputting an image of the metal-containing material. For example, the imaging device 10 may be fixed via a stand, configured to be portable, or mounted on a flying object such as a drone.

[0020] The imaging device 10 is used to acquire images of metal-containing materials to be classified when classifying metal-containing materials. Furthermore, the imaging device 10 may also be used to acquire images of metal-containing materials in advance when creating a model, as described later.

[0021] Sensor 20 is applied to metal-containing materials and detects physical quantities related to corrosion. Sensor 20 may be of any form as long as it is capable of detecting physical quantities related to corrosion and outputting a sensor output value corresponding to the detection. For example, the sensor output value may be a value output by using at least one selected from among ACM sensors, ICM sensors, surface salinity meters, atmospheric corrosive gas monitor sensors, wetting sensors, rainfall and / or condensation sensors, temperature and / or humidity sensors, dissolved oxygen meters, pH meters, ultrasonic thickness gauges, chloride ion meters, infrared thermography, electrochemical measuring devices, and devices used in gravimetric methods.

[0022] Sensor 20 is used to detect physical quantities related to the corrosion of metal-containing materials to be classified and to obtain sensor output values ​​when classifying metal-containing materials. Furthermore, sensor 20 may also be used to pre-detect physical quantities related to the corrosion of metal-containing materials and to pre-obtain sensor output values ​​when creating the model described later.

[0023] As shown in Figure 2, if the sensor 20 is, for example, an ACM (Atmospheric Corrosion Monitor) sensor, an anode and a cathode are provided on the substrate. An insulating layer is interposed between the anode and the cathode. When a water film forms between the galvanic couple under relatively high humidity conditions, a current flows between the anode and the cathode. This current corresponds to the corrosion rate of the substrate. In other words, an ACM sensor is a sensor that outputs a current corresponding to the degree of corrosion.

[0024] The anode can be made of, for example, iron plate, galvanized steel plate, copper plate, zinc plate, aluminum plate, magnesium plate, tin plate, nickel plate, or chromium plate, and may be plated or alloyed. The cathode can be made of silver, carbon, copper, gold, platinum, etc., and may be molded from resin paste or laminated by sputtering or vapor deposition.

[0025] As the ACM sensor, for example, a configuration specified in JIS Z2384:2019 "Atmospheric Corrosion Monitoring Sensor" may be used. More specifically, for example, a Zn / Ag ACM sensor or an Fe / Ag ACM sensor corresponding to the said specification may be used. For the substrate, galvanized steel sheet or carbon steel may be used. For example, when a galvanized steel sheet is used as the substrate, a hot-dip galvanized steel sheet of the type specified in JIS G 3302, SGCC with a thickness of 0.80 mm ± 0.07 mm, or SGCH with a thickness of 0.80 mm ± 0.07 mm, which has not undergone chemical conversion treatment (chemical conversion treatment type symbol M), and a steel sheet with a plating adhesion amount indication symbol of Z22 or higher may be used.

[0026] As shown in Figure 3, if the sensor 20 is, for example, an ICM (Intelligent Corrosion Monitor) sensor (registered trademark), it is provided with two spiral-shaped electrodes. The gap between the electrodes is filled with epoxy resin or the like for insulation. In the ICM sensor, the impedance corresponding to the degree of corrosion is obtained using the AC impedance method.

[0027] As the ICM sensor, for example, one having the external shape of design registration No. 1741384, or a configuration and measurement method described in "Corrosion Monitoring of Aluminum by AC Impedance Method" (Proceedings of the 69th Symposium on Materials and Environment, Vol. 72, No. 5, 2023, pp. 164-168) may be used. More specifically, for example, a Zn / Zn versus ICM sensor may be used.

[0028] As shown in Figure 1, the corrosion classification system 100 includes a modality acquisition unit 31, a model generation unit 32, and a corrosion classification unit 33 at terminal 30. The modality acquisition unit 31, the model generation unit 32, and the corrosion classification unit 33 are composed of electrical and electronic circuits, a CPU, memory, stored programs, etc.

[0029] The modality acquisition unit 31 acquires images of metal-containing materials to be classified according to different degrees of corrosion, and sensor output values ​​obtained when the sensor 20 detects physical quantities related to corrosion on the metal-containing materials to be classified. The modality acquisition unit 31 acquires images obtained by imaging the metal-containing materials via the imaging device 10 as modalities. In addition, the modality acquisition unit 31 acquires sensor output values ​​obtained by detection on the metal-containing materials via the sensor 20 as modalities. In this way, the modality acquisition unit 31 acquires multiple modalities.

[0030] Furthermore, the modality acquisition unit 31 may be configured to acquire images and sensor output values ​​in advance in order to generate a model described later, before acquiring images of objects to be classified according to different degrees of corrosion, and sensor output values ​​of objects to be classified according to different degrees of corrosion.

[0031] The model generation unit 32 generates a model for classifying metal-containing materials according to their different degrees of corrosion. This model is obtained by learning the relationship between images of multiple metal-containing materials acquired in advance, sensor output values ​​acquired in advance by the sensor 20 detecting physical quantities related to corrosion for each of the multiple metal-containing materials, and the degrees of corrosion of the multiple metal-containing materials acquired in advance. The model may be, for example, a multimodal AI image class classification system using the above-mentioned image and sensor output values ​​as modalities. In this case, the images may be classified into classes according to the different degrees of corrosion of the metal-containing materials. More specifically, for example, the model may be generated so that multiple metal-containing materials to be classified are classified into those with a high degree of corrosion and those with a low degree of corrosion.

[0032] The corrosion classification unit 33 classifies metal-containing materials to be classified according to their different degrees of corrosion by applying the acquired images and sensor output values ​​to the generated model.

[0033] As shown in Figures 4 and 5, when classifying metal-containing materials using the model, the corrosion classification unit 33 performs the classification, going through processing in the input layer, hidden layer, and output layer. This classification is performed according to the degree of corrosion. In the examples in Figures 4 and 5, the amount of salt deposition and exposure time are used as the degree of corrosion, respectively, and either method may be adopted. In addition to the amount of salt deposition and exposure time, other factors and indicators related to corrosion may be used as the degree of corrosion. For example, corrosion factors such as acid, sulfide, condensation, electrolytes, and pH, the degree of corrosion wear, the degree of rust area, wetting time, environmental indicators of ISO 9223, corrosion forms such as stress corrosion cracking and localized corrosion, and, if painted steel sheets are used, the corrosion state under the paint film (state of spot rust, corrosion exposure area, etc.), the degree of paint film deterioration, the degree of peeling, the size and number of rust nodules may be used.

[0034] The amount of salt used in the example in Figure 4 is, for example, the amount of sea salt adhering to the surface of a metal-containing material (such as a galvanized steel sheet), and is the weight per unit area (g / m²) of the surface. 2 This is expressed as follows. Note that, instead of sea salt, for example, snowmelt salt (NaCl, CaCl2, MgCl2, etc.) may be used. The exposure time used in the example in Figure 5 is, for example, the time a metal-containing material (such as galvanized steel sheet) is exposed under a specified environment. The detailed conditions for salt deposition and exposure time will be described later.

[0035] In metal-containing materials, given a constant exposure time, a higher salt deposit leads to a higher degree of corrosion. Based on this relationship, the example in Figure 4 performs a classification system corresponding to three different salt deposit amounts, based on the degree of corrosion. More specifically, the metal-containing material being classified is assigned to one of three salt deposit classes: salt deposit class 1, salt deposit class 2, or salt deposit class 3.

[0036] Here, the salt adhesion amount class 1, the salt adhesion amount class 2, and the salt adhesion amount class 3 may be respectively associated with the ranges of the salt adhesion amount so that the salt adhesion amount increases in this order. For example, the salt adhesion amount class 1 corresponds to the range where the salt adhesion amount is 0 g / m 2 or more and less than 0.01 g / m 2 . The salt adhesion amount class 2 corresponds to the range where the salt adhesion amount is 0.01 g / m 2 or more and less than 0.1 g / m 2 . The salt adhesion amount class 3 corresponds to the range where the salt adhesion amount is 0.1 g / m 2 or more and less than 1.0 g / m 2 . Note that the range of the salt adhesion amount class 3 may also correspond to the case where the salt adhesion amount is 1.0 g / m 2 or more.

[0037] Note that the ranges of the adhesion amounts corresponding to the salt adhesion amount class 1, the salt adhesion amount class 2, and the salt adhesion amount class 3 are not limited to those described above. For example, they may be set to any ranges including 0.01 g / m 2 , any ranges including 0.1 g / m 2 , any ranges including 1.0 g / m 2 (for example, ranges including 1.0 g / m 2 or more and including 10 g / m 2 , etc.). Also, in this example, the number of classifications is three corresponding to the range of the salt adhesion amount. Instead of this, the number of classifications may be two, or may be four or more.

[0038] In the metal-containing material, when the adhesion amount is constant, the longer the exposure time, the higher the degree of corrosion. Based on this relationship, in the example of FIG. 5, as a classification according to the degree of corrosion, classification is performed so as to correspond to three different exposure times. More specifically, the metal-containing material to be classified is classified into any one of the exposure time class 1, the exposure time class 2, and the exposure time class 3.

[0039] Here, exposure time classes 1, 2, and 3 may be assigned exposure time ranges such that the exposure time increases in this order. For example, exposure time class 1 corresponds to an exposure time of 0 hours or more and less than 10 hours. Exposure time class 2 corresponds to an exposure time of 10 hours or more and less than 200 hours. Exposure time class 3 corresponds to an exposure time of 200 hours or more.

[0040] Furthermore, the time ranges corresponding to exposure time class 1, exposure time class 2, and exposure time class 3 are not limited to those described above. For example, they may be set to any range including 0 hours from the time of sea salt attachment, any range including 10 hours from the time of sea salt attachment, or any range including 200 hours from the time of sea salt attachment (for example, a range of 200 hours or more that includes 1,000 hours, 10,000 hours, etc.). Also, in this example, there are three classifications corresponding to the exposure time ranges, but instead, there may be two classifications or four or more classifications.

[0041] As shown in Figures 4 and 5, in the classification by the corrosion classification unit 33, first, the input layer receives the "image" and "sensor output value" acquired by the modality acquisition unit 31. The information of the "image" and "sensor output value" is subject to feature fusion. Before feature fusion, the "image" may be processed, for example, by a machine learning image classification method or by a deep learning type image classification method such as CoAtNet (Convolution-and-Attention Network), and the "sensor output value" may be processed, for example, by MLP (Multi-Layer Perceptron).

[0042] In this context, when machine learning (kernel methods) are used in image classification techniques, for example, SVM (Support Vector Machine), decision trees, Random Forests, k-NN, etc. may be used. Furthermore, the deep learning-type image classification technique is not limited to CoAtNet, but may also be used, for example, AlexNet, VGG, CNN (Convolutional Neural Network), GoogLeNet, ResNet, DenseNet, EfficientNet, ViT (Vision Transformer), Swin Transformer, ConvNeXt, etc.

[0043] Next, in the hidden layer, the feature-fused information described above is input from the input layer to the fully connected layer. Then, in the output layer, the input information is classified after passing through the fully connected layer. In terms of classification methods, when classifying according to the amount of sea salt attached, it is classified into one of the following classes: salt attachment class 1, salt attachment class 2, and salt attachment class 3 (see Figure 4). When classifying according to the exposure time from the time of sea salt attachment, it is classified into one of the following classes: exposure time class 1, exposure time class 2, and exposure time class 3 (see Figure 5). In the above processing for classification according to different degrees of corrosion by the corrosion classification unit 33, a model that has been pre-generated by the model generation unit 32 is used before the classification.

[0044] As shown in Figures 6 and 7, the model generation uses images of multiple metal-containing materials that have been acquired in advance. More specifically, the model generation unit 32 uses images of multiple metal-containing materials obtained by applying the same amount of sea salt to the surface of each of the multiple metal-containing materials, and then imaging the surface after different exposure times, and / or images obtained by applying different amounts of sea salt to the surface of each of the multiple metal-containing materials, and then imaging the surface after the same exposure time.

[0045] As the metal-containing material to be imaged, for example, galvanized steel sheet can be used. In this case, for example, a galvanized steel sheet (without chemical treatment) measuring 64 mm in length x 64 mm in width x 0.8 mm in thickness may be divided into four sections, and different amounts of artificial sea salt may be applied to the surface of each section, after which it may be dried and placed in a constant humidity chamber. Here, the relative humidity in the constant humidity chamber may be, for example, 80%. As the artificial sea salt, for example, Aquamarine manufactured by Yashima Pharmaceutical Co., Ltd. may be used.

[0046] In this case, the same amount of sea salt is applied to the surface of a galvanized steel sheet, and then the surface is imaged after different exposure times in a constant temperature chamber. The upper example in Figure 7 shows images when the amount of sea salt applied is constant, and the exposure time is varied from 0 to 200 hours. In addition, different amounts of sea salt are applied to the surface of a galvanized steel sheet, and then the surface is imaged after the same exposure time in a constant temperature chamber. The lower example in Figure 7 shows images when the exposure time is constant, and the amount of sea salt applied is 0.01 g / m². 2 ~1.0g / m 2 These are images showing the results when the settings are varied within a certain range.

[0047] In this embodiment, the amount of sea salt attached is 0.01 g / m². 2 , 0.1g / m 2 1.0g / m 2 Three conditions (all with an exposure time of 10 hours) were adopted (see Figure 4). More specifically, 120 galvanized steel sheets treated with the above three sea salt deposition amounts were prepared, and the surface of the galvanized steel sheets was imaged. In this case, 120 images were acquired. Of these 120 images, 90 may be used for training and 30 as test data. As for the images, 1051 × 1051 pixels may be compressed to 224 × 224 pixels and converted to 8-bit grayscale. Alternatively, instead of conditions with different sea salt deposition amounts, three conditions with exposure times of 0 hours, 10 hours, and 200 hours (all with a sea salt deposition amount of 0.1 g / m²) may be used. 2 ) may be adopted respectively (see Figure 5). In this case as well, the same image processing, number of training images, and number of test data may be used as described above.

[0048] Furthermore, in addition to the images described above, the model generation uses sensor output values ​​obtained in advance from sensors 20, which detect physical quantities related to corrosion for each of the multiple metal-containing materials. More specifically, the model generation unit 32 uses, as sensor output values ​​for the multiple metal-containing materials, the sensor output values ​​obtained by sensors 20 for each of the materials after applying the same amount of sea salt to the surface of each of the multiple metal-containing materials and then allowing each material to undergo different exposure times, and / or the sensor output values ​​obtained by sensors 20 for each of the materials after applying different amounts of sea salt to the surface of each of the multiple metal-containing materials and then allowing each material to undergo the same exposure time.

[0049] As described above, when galvanized steel sheets are used to acquire images, the sensor 20 may be used for galvanized steel sheets with the same conditions and number of levels. In this case, for example, a Zn / Ag vs. ACM sensor may be used as the sensor 20, and the Coulomb value of the ACM sensor may be used as the sensor output value.

[0050] As shown in Figure 4, the model generation uses the image and sensor output values ​​described above, as well as the corrosion degrees of multiple metal-containing materials acquired in advance. As mentioned above, when galvanized steel sheets are used to acquire the image and sensor output values, the amount of adhesion corresponding to the previously acquired image may be used as the corrosion degree.

[0051] During the learning process, as described above, information combining the characteristics of multiple pre-acquired images and sensor output values ​​is input into the fully connected layer of the hidden layer, and metal-containing materials are classified according to their different deposition amounts. Sea salt deposition amount: 0.01 g / m 2 , 0.1g / m 2 1.0g / m 2 When these three conditions are used, the amount of each deposit should be obtained in advance. Sea salt deposit amount: 0.01 g / m 2 Metal-containing materials that meet the following conditions are classified as salt adhesion class 1, with a sea salt adhesion amount of 0.1 g / m². 2Metal-containing materials that meet the following conditions are classified as salt adhesion class 2, with a sea salt adhesion amount of 1.0 g / m². 2 Metal-containing materials that meet the specified conditions are repeatedly trained to be classified into salt deposition class 3. The number of training iterations may be, for example, 500 (epochs=500, batch size=8).

[0052] As shown in Figure 5, the model generation uses the image and sensor output values ​​described above, as well as the corrosion degrees of multiple metal-containing materials acquired in advance. As mentioned above, when galvanized steel sheets are used to acquire the image and sensor output values, the exposure time corresponding to the previously acquired image may be used as the corrosion degree.

[0053] During training, as described above, information fused with the features of multiple pre-acquired images and sensor output values ​​is input to the fully connected hidden layer, and metal-containing materials are classified according to different exposure times. If three exposure time conditions are used: 0 hours, 10 hours, and 200 hours, each exposure time is acquired in advance. Metal-containing materials corresponding to the 0-hour exposure condition are classified into exposure time class 1, metal-containing materials corresponding to the 10-hour exposure condition are classified into exposure time class 2, and metal-containing materials corresponding to the 200-hour exposure condition are classified into exposure time class 3, and so on, through repeated training. The number of training iterations may be, for example, 500 (epochs=500, batch size=8).

[0054] As described above, the model generation unit 32 learns the relationship between images of multiple galvanized steel sheets acquired in advance, sensor output values ​​acquired in advance for multiple galvanized steel sheets by a Zn / Ag vs. ACM sensor, and the amount of sea salt deposited on or exposure time of multiple galvanized steel sheets acquired in advance, thereby obtaining a model for classifying galvanized steel sheets.

[0055] Furthermore, when acquiring images and sensor output values ​​in the creation of the model, the conditions and levels of salt deposition amount and exposure time are not limited to those described above, and any images obtained by actually exposing the test specimens can be used.

[0056] In this embodiment, galvanized steel sheet is used as an example of a metal-containing material. Various other materials can be used instead. For example, carbon steel, alloy steel (high-tensile steel, weathering steel), stainless steel, aluminum and aluminum alloys, copper and copper alloys, titanium and titanium alloys, aluminum-plated steel sheet, magnesium alloy, nickel alloy, Zn-Al alloy thermal spray, and other metals, plating, thermal spraying in general, painted steel sheets, etc., may be used. In this case, if, for example, a painted steel sheet is used as the metal-containing material, images showing the corrosion state under the coating (such as the state of spot rust and the area of ​​exposed corrosion) may be applied. More specifically, images showing the degree of coating deterioration, the degree of peeling, and the size and number of rust spots may be applied.

[0057] The actual operation of the corrosion classification system 100 will be explained with reference to the flowchart shown in Figure 8. The program in the corrosion classification system 100 causes the terminal 30 to execute the series of processes shown in Figure 8, thereby classifying the galvanized steel sheets. The program for these processes may be stored in a storage medium, and this storage medium may be installed in the terminal 30.

[0058] When performing classification, the process begins with step S1, in which the model generation unit 32 learns the relationship between the previously acquired image of the galvanized steel sheet, the sensor output value, and the amount of sea salt deposited (or exposure time), and generates a model for classifying the galvanized steel sheet according to different ranges of sea salt deposited (or exposure time).

[0059] Next, in step S2, the modality acquisition unit 31 acquires the image and sensor output values ​​of the galvanized steel sheet to be classified.

[0060] Then, in step S3, the corrosion classification unit 33 applies the image and sensor output values ​​acquired in step S2 as modalities to the model generated in step S1, and the galvanized steel sheet is classified (see Figures 4 and 5).

[0061] As described above, the corrosion classification system 100 according to the embodiment of the present invention has the following effects. The corrosion classification system 100 of this embodiment comprises a model generation unit 32 that generates a model for classifying metal-containing materials according to different corrosion degrees, which is obtained by learning the relationship between images of a plurality of metal-containing materials acquired in advance, sensor output values ​​acquired in advance when physical quantities related to corrosion are detected by the sensor 20 for each of the plurality of metal-containing materials, and the corrosion degree of the plurality of metal-containing materials acquired in advance; a modality acquisition unit 31 that acquires images of the metal-containing materials to be classified and sensor output values ​​acquired when physical quantities related to corrosion are detected by the sensor for each of the metal-containing materials to be classified; and a corrosion classification unit 33 that classifies the metal-containing materials to be classified according to different corrosion degrees by applying the acquired images and the acquired sensor output values ​​to the generated model.

[0062] According to this, the modalities used are images of the metal-containing material to be classified, and sensor output values ​​obtained by the sensor 20 detecting physical quantities related to corrosion on the metal-containing material to be classified. These multiple modalities can be integrated and used for classification. For classification, a model obtained by learning the relationship between previously acquired images, sensor output values, and the degree of corrosion is used.

[0063] As shown in Figure 9, as a comparative example, classification according to the degree of corrosion can also be performed using a unimodal AI that uses only images as a modality. In this comparative example, a CNN may be used as the deep learning type image classification method. The image, as a modality, is input to the input layer after processing in the convolutional layer and pooling layer. The input information is classified in the output layer after passing through the fully connected layer. The classification method may be the same as in this embodiment, such as classification according to the range of adhesion amount or the range of exposure time. Below, we will explain the comparison between the accuracy rates of the two comparative examples and the accuracy rate of this embodiment.

[0064] In the first comparative example, salt adhesion amount was classified into Class 1, Class 2, and Class 3. Here, as in this embodiment, sea salt adhesion amount: 0.01 g / m 2 , 0.1g / m 2 1.0g / m 2 Three conditions (all with an exposure time of 10 hours) were adopted, and 120 galvanized steel sheets were prepared, treated with the amount of sea salt deposited according to the three conditions. Images of the surface of these galvanized steel sheets were then taken. In this case, 120 images were acquired. Of these 120 images, 90 were used for training and 30 were used as test data. The images used were compressed from 1051×1051 pixels to 224×224 pixels and converted to 8-bit grayscale. In this case, after 500 training iterations (epochs=100, batch size=8), the accuracy rate on the test data was 67%.

[0065] In contrast, in this embodiment, when classifying salt adhesion amount into Class 1, Class 2, and Class 3 under the conditions described above, the accuracy rate on the test data was 93%. This embodiment achieved a higher accuracy rate compared to the first comparative example.

[0066] In the second comparative example, exposure time classification into exposure time class 1, exposure time class 2, and exposure time class 3 was performed. Here, as in the embodiment, three conditions were used for exposure time: 0 hours, 10 hours, and 200 hours (in all cases, sea salt deposition amount: 0.1 g / m 2 We adopted the same approach as in the first comparative example described above and prepared images similar to those used in this case. In this case, after 500 training iterations (epochs=100, batch size=8), the accuracy rate on the test data was 45%.

[0067] In contrast, in this embodiment, when classification of exposure time class 1, exposure time class 2, and exposure time class 3 was performed under the above-described conditions, the accuracy rate on the test data was 87%. This embodiment achieved a higher accuracy rate compared to the second comparative example.

[0068] Thus, the corrosion classification system 100 of this embodiment can achieve a higher accuracy rate compared to classification by unimodal AI. Therefore, this embodiment, which utilizes multiple modalities, can accurately classify metal-containing materials according to their degree of corrosion. This classification result can be used for corrosion diagnosis. As a result, quantitative and objective aspects of corrosion diagnosis can be improved, as well as safety and efficiency. Therefore, the reliability of inspections can be improved, and the timing of repairs and painting can be appropriately determined. Furthermore, it can contribute to improved infrastructure safety, reduced life cycle costs, and sustainable asset management.

[0069] Furthermore, in the corrosion classification system 100 of this embodiment, the model generation unit 32 uses, as images of the plurality of metal-containing materials, images obtained by applying the same amount of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after different exposure times, and / or images obtained by applying different amounts of salt to the surface of each of the plurality of metal-containing materials and then imaging the surface after the same exposure time.

[0070] According to this method, a large amount of discrete data with various conditions and levels can be easily prepared using a predetermined technique when generating a model. This large amount of data can then be used for training. Therefore, a model can be easily created, and this model can be used to perform accurate classification according to the degree of corrosion.

[0071] Furthermore, in the corrosion classification system 100 of this embodiment, the model generation unit 32 generates a model that classifies the metal-containing material into at least two classification numbers according to different exposure times or different amounts of salt deposition, and the corrosion classification unit 33 classifies the metal-containing material to be classified according to different exposure times or different amounts of salt deposition so as to correspond to the classification numbers of the model.

[0072] This allows for the creation of a model that classifies materials according to subdivided exposure time or deposition amount. This model enables the subdivision of metal-containing materials based on different exposure times or deposition amounts. Therefore, the subdivided classification results can be used for corrosion diagnosis.

[0073] Furthermore, in the corrosion classification system 100 of this embodiment, the sensor output value is a value output by using at least one selected from among ACM sensors, ICM sensors, surface salinity meters, atmospheric corrosive gas monitor sensors, wetting sensors, rainfall and / or condensation sensors, temperature and / or humidity sensors, dissolved oxygen meters, pH meters, ultrasonic thickness gauges, chloride ion meters, infrared thermography, electrochemical measuring devices, and devices used in gravimetric methods.

[0074] According to this method, the output values ​​of readily available and general-purpose sensors can be used as the sensor output values. Furthermore, a wide range of sensors can be selected. Therefore, sensor output values ​​can be easily acquired. [Explanation of Symbols]

[0075] 10...Imaging device, 20...Sensor, 30...Terminal, 31...Modality acquisition unit, 32...Model generation unit, 33...Corrosion classification unit, 100...Corrosion classification system

Claims

1. A model obtained by learning the relationship between images of multiple metal-containing materials acquired in advance, sensor output values ​​acquired in advance by sensors detecting physical quantities related to corrosion for each of the multiple metal-containing materials, and the degree of corrosion of the multiple metal-containing materials acquired in advance, comprising a model generation unit that generates a model for classifying metal-containing materials according to different degrees of corrosion, A modality acquisition unit acquires an image of the metal-containing material to be classified, and a sensor output value obtained when the sensor detects a physical quantity related to corrosion in the metal-containing material to be classified. A corrosion classification unit that classifies the metal-containing materials to be classified according to different degrees of corrosion by applying the acquired images and the acquired sensor output values ​​to the generated model, Equipped with, The aforementioned model generation unit, As images of the aforementioned multiple metal-containing materials, Images obtained by photographing the surfaces of the multiple metal-containing materials after applying the same amount of salt to each surface and then allowing them to undergo different exposure times, and / or images obtained by photographing the surfaces of the multiple metal-containing materials after applying different amounts of salt to each surface and then allowing them to undergo the same exposure time. Corrosion classification system.

2. In the corrosion classification system according to claim 1, The aforementioned model generation unit, A model is generated that classifies the metal-containing material into at least two categories depending on the exposure time or the amount of salt deposited. The aforementioned corrosion classification unit is Depending on the different exposure times or the different amounts of salt deposited, the metal-containing materials to be classified are classified to correspond to the number of classifications in the model. Corrosion classification system.

3. In the corrosion classification system according to Claim 1 or Claim 2, The output value of the aforementioned sensor is, The value is output by using at least one selected from among ACM sensors, ICM sensors, surface salinity meters, atmospheric corrosive gas monitor sensors, wetness sensors, rainfall and / or condensation sensors, temperature and / or humidity sensors, dissolved oxygen meters, pH meters, ultrasonic thickness gauges, chloride ion meters, infrared thermography, electrochemical measuring devices, and devices used in gravimetric methods. Corrosion classification system.

4. A computer, A model obtained by learning the relationship between images of multiple metal-containing materials acquired in advance, sensor output values ​​acquired in advance by sensors detecting physical quantities related to corrosion for each of the multiple metal-containing materials, and the degree of corrosion of the multiple metal-containing materials acquired in advance, comprising a model generation step of generating a model that classifies metal-containing materials according to different degrees of corrosion, A modality acquisition step that acquires an image of the metal-containing material to be classified, and a sensor output value obtained by the sensor when a physical quantity related to corrosion is detected on the metal-containing material to be classified, A classification step in which the metal-containing materials to be classified are classified according to different degrees of corrosion by applying the acquired images and the acquired sensor output values ​​to the generated model, Make it run, The aforementioned model generation step is: As images of the aforementioned multiple metal-containing materials, Images obtained by photographing the surfaces of the multiple metal-containing materials after applying the same amount of salt to each surface and then allowing them to undergo different exposure times, and / or images obtained by photographing the surfaces of the multiple metal-containing materials after applying different amounts of salt to each surface and then allowing them to undergo the same exposure time. Corrosion classification program.

5. In the corrosion classification program described in Claim 4, The aforementioned model generation step is: A model is generated that classifies the metal-containing material into at least two categories depending on the exposure time or the amount of salt deposited. The aforementioned classification step is, Depending on the different exposure times or the different amounts of salt deposited, the metal-containing materials to be classified are classified to correspond to the number of classifications in the model. Corrosion classification program.

6. In the corrosion classification program according to claim 4 or claim 5, The output value of the aforementioned sensor is, The value is output by using at least one selected from among ACM sensors, ICM sensors, surface salinity meters, atmospheric corrosive gas monitor sensors, wetness sensors, rainfall and / or condensation sensors, temperature and / or humidity sensors, dissolved oxygen meters, pH meters, ultrasonic thickness gauges, chloride ion meters, infrared thermography, electrochemical measuring devices, and devices used in gravimetric methods. Corrosion classification program.

Images