Material analysis method

Abstract

The present invention relates to a material analysis method, which may comprise the steps of: (a) preparing a sample for structural analysis; (b) training an artificial intelligence model by using a training image marked with structural characteristics of any sample; (c) analyzing a plurality of to-be-analyzed images with unlabeled sample structure characteristics by using the trained artificial intelligence model, and removing the to-be-analyzed images classified as preset noise structure characteristics; and (d) using the trained artificial intelligence model to analyze the remaining to-be-analyzed images after the to-be-analyzed images classified as the noise structure characteristics are removed, and classifying the remaining to-be-analyzed images as at least one target structure characteristic.

Description

Technical Field

[0001] This invention relates to a material analysis method, and more specifically, to a method capable of analyzing the microstructure of materials using artificial intelligence. Background Technology

[0002] In the coking process of steel plants, coal is dry-distilled at high temperatures of 1000°C to 1300°C to produce coke suitable for use in blast furnaces. Coal used as a feedstock for coke is a mineral composed of carbonaceous material, consisting of a mixture of various types of macerals (components visible under a microscope).

[0003] Microscopic components are organic components distinguished primarily by size and shape using a microscope, and are used as the smallest unit for identification when classifying the microstructure of coal. The classification of microscopic components according to the KS standard (KS E ISO 7404-3) relies on the inspector's visual analysis (point counting method) and requires counting at least 500 points. This not only leads to analytical discrepancies between inspectors but also results in a lengthy process. Summary of the Invention

[0004] Technical issues This invention aims to solve various problems, including those mentioned above, and aims to provide a materials analysis method that automates visual analysis by a small number of skilled workers using a deep learning-based artificial intelligence model to distinguish and calculate the proportions of coal microstructures. This allows even unskilled workers to easily and quickly distinguish and calculate the proportions of coal microstructures. However, these problems are exemplary and do not limit the scope of this invention.

[0005] Technical solution According to an embodiment of the present invention, a material analysis method is provided. The material analysis method may include: (a) preparing a sample for analyzing microstructure; (b) acquiring multiple sample images by photographing a detection area disposed on a surface of the sample; (c) preprocessing the multiple sample images to obtain multiple analysis images; and (d) analyzing the multiple analysis images using a trained microstructure classification artificial intelligence model to classify them into at least one target microstructure pattern.

[0006] According to an embodiment of the present invention, step (a) may include: (a-1) crushing the material into particles of a predetermined size; and (a-2) mixing the particles with a binder to prepare a sample.

[0007] According to an embodiment of the present invention, the material may be coal, and the binder may be epoxy resin.

[0008] According to an embodiment of the present invention, step (c) may be a step of using a class activation map to extract the region of interest set by the artificial intelligence model for analyzing the sample image, thereby obtaining the analysis image.

[0009] According to an embodiment of the present invention, the region of interest may be a region corresponding to the central region of the sample image.

[0010] According to an embodiment of the present invention, step (c) may be a step of using a trained noise removal artificial intelligence model to analyze multiple sample images in which unlabeled samples have microstructure patterns, in order to remove the analysis images classified as preset noise microstructure patterns.

[0011] According to an embodiment of the present invention, the noise microstructure pattern can be an image in which the microstructure pattern of the sample is classified as an adhesive by a noise removal artificial intelligence model.

[0012] According to an embodiment of the present invention, the noise removal artificial intelligence model can be an artificial intelligence model using a deep learning network based on a residual neural network (ResNet) or MobileNet.

[0013] According to an embodiment of the present invention, in step (d), the microstructure classification artificial intelligence model can be an artificial intelligence model using an Inception-based deep learning network.

[0014] According to an embodiment of the present invention, prior to step (c), the method may further include: (f) using training images in which microstructure patterns of arbitrary samples are marked to train a noise removal artificial intelligence model or a microstructure classification artificial intelligence model.

[0015] According to an embodiment of the present invention, in step (d), the target microstructure pattern can be an image in which the microstructure pattern of the sample is classified by a microstructure classification artificial intelligence model into at least one of vitrified coal, exinite, fusinite, semi-fusinite, mineral, and combinations thereof.

[0016] According to an embodiment of the present invention, after step (d), the method may further include: (e) calculating the proportion of classified microstructure patterns in the sample.

[0017] Beneficial effects According to the above-described embodiments of the present invention, by setting a hexagonal detection area, the characteristics of the entire circular cross-section can be fully reflected and efficient analysis progress can be ensured. Moreover, it has the effect that microscopic structures of micro-components that are difficult for ordinary workers to distinguish can be easily and quickly distinguished using artificial intelligence.

[0018] Furthermore, by independently using an AI model for classifying epoxy resin images and an AI model for classifying the microstructure of microscopic components, it is possible to classify the microstructure of microscopic components with higher accuracy, and to improve accuracy by extracting and analyzing the central region of the sample image. Of course, the scope of this invention is not limited to these effects. Attached Figure Description

[0019] Figure 1 A flowchart is provided to illustrate the steps of the material analysis method according to an embodiment of the present invention in sequence.

[0020] Figure 2 A flowchart is provided to illustrate in more detail the various steps of the material analysis method according to an embodiment of the present invention.

[0021] Figure 3 An image illustrating the structure of a noise removal artificial intelligence model according to an embodiment of the present invention.

[0022] Figure 4 An image illustrating the structure of a microstructure classification artificial intelligence model according to an embodiment of the present invention.

[0023] Figure 5 Tables and images illustrating confusion matrices and microscopic component structures according to embodiments of the present invention are provided.

[0024] Figure 6 A table showing the accuracy of extracting the central region of the sample and inputting it into each microstructure classification AI model.

[0025] Figure 7 A confusion matrix image showing the degree of improvement in test accuracy when the central region of the sample is extracted.

[0026] Figure 8 An image showing the region of interest set by the artificial intelligence model for analyzing sample images.

[0027] Figure 9 A table showing the accuracy of an AI model by removing noise from its initial learning rate.

[0028] Figure 10 A table showing the accuracy of an AI model classifying microstructures based on its initial learning rate.

[0029] Figure 11 A conceptual diagram illustrating a material analysis apparatus according to an embodiment of the present invention is provided.

[0030] Figure 12 An image of a hexagonal detection region according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0031] Various preferred embodiments of the invention will be described in detail below with reference to the accompanying drawings.

[0032] The embodiments of the present invention are provided to explain the invention more fully to those skilled in the art, and the following embodiments can be modified in various different forms, and the scope of the invention is not limited to the following embodiments. Rather, these embodiments are provided to make the invention more faithful and complete, and to fully convey the spirit of the invention to those skilled in the art. In addition, for ease of explanation and clarity, the thickness or dimension of each layer in the drawings is enlarged.

[0033] Hereinafter, embodiments of the invention will be described with reference to the accompanying drawings, which schematically illustrate preferred embodiments of the invention. In the drawings, for example, the shapes shown may be deformed depending on manufacturing techniques and tolerances. Therefore, embodiments of the spirit of the invention should not be construed as limited to the specific shape of the areas shown in this specification, but should include, for example, variations in shape due to manufacturing processes.

[0034] Figure 1 A flowchart is provided to illustrate the steps of the material analysis method according to an embodiment of the present invention in sequence.

[0035] like Figure 1 As shown, the material analysis method according to an embodiment of the present invention may include: (a) preparing a sample for analyzing microstructure; (b) acquiring multiple sample images by photographing a detection area disposed on a surface of the sample; (c) preprocessing the multiple sample images to obtain multiple analysis images; and (d) analyzing the multiple analysis images using a trained microstructure classification artificial intelligence model to classify them as at least one target microstructure pattern.

[0036] More specifically, such as Figure 2 As shown, step (a) of preparing a sample for analyzing microstructure may include (a-1) crushing the material into particles of a predetermined size; and (a-2) mixing the particles with a binder to prepare the sample.

[0037] At this point, the material is coal and the binder is epoxy resin. The steps include crushing the coal into particles of a predetermined size, mixing the binder, and then molding the mixture into an integral cylindrical sample under pressure.

[0038] Subsequently, in step (b) of acquiring multiple images of the sample by photographing a detection area set on one surface of the sample, the detection area may be an area with multiple detection points set to form an overall hexagon. Such a detection area may be an area with multiple detection points set to form an overall hexagon, as described below. Figure 9 As shown in the image.

[0039] Furthermore, step (c) of preprocessing multiple sample images to obtain multiple analytical images may include: extracting regions of interest set by an artificial intelligence model for analyzing the sample images using class activation maps, thereby obtaining analytical images. This will be discussed later. Figure 8 Describe the area of ​​interest in more detail.

[0040] Furthermore, step (c) of preprocessing multiple sample images to obtain multiple analytical images may include: analyzing multiple sample images in which unlabeled samples exhibit microstructural patterns using a trained noise removal artificial intelligence model to remove analytical images classified as preset noise microstructural patterns. The noise microstructural pattern may be an image in which the microstructural pattern of the sample is classified by the noise removal artificial intelligence model as an adhesive (i.e., epoxy resin).

[0041] More specifically, the noise removal AI model can identify and classify images of the microstructure pattern of the sample as the binder as the noise microstructure pattern; while the microstructure classification AI model can identify and classify images of the microstructure pattern of the sample as at least one of the following: vitreous coal quality, shell coal quality, fibrous coal quality, semi-fibrous coal quality, mineral quality, and combinations thereof as the target microstructure pattern.

[0042] In addition, such as Figure 2 As shown, the material analysis method according to an embodiment of the present invention may further include: (f) training an artificial intelligence model using training images in which microstructure patterns of arbitrary specimens are marked, prior to step (c) of removing images classified as noise microstructure patterns.

[0043] Figure 2 Steps (a) through (e) shown do not necessarily have to be performed in sequence. For example, after training the artificial intelligence model with an arbitrary sample in step (f), a sample for analyzing the microstructure can be prepared in step (a).

[0044] More specifically, the step (f) of training an artificial intelligence model using training images may include (f-1) training a noise removal artificial intelligence model using noisy training images, and (f-2) training a microstructure classification artificial intelligence model using target microstructure training images.

[0045] At this point, in step (f-1) of training the noise removal artificial intelligence model using noise training images, the noise training images can be images labeled by classifying the microstructure pattern of the sample as an adhesive.

[0046] Furthermore, in step (f-2) of training the microstructure classification artificial intelligence model using the target microstructure training image, the target microstructure training image can be an image labeled by classifying the microstructure pattern of the sample into at least one of vitreous coal quality, hulled coal quality, fibrous coal quality, semi-fibrous coal quality, mineral quality, and combinations thereof.

[0047] The steps (c) to (d) of the material analysis method according to an embodiment of the present invention will be described in detail below.

[0048] According to an embodiment of the present invention, in step (c), the analysis image can be obtained by using class activation maps to extract the region of interest set by the artificial intelligence model for analyzing the sample image.

[0049] According to an embodiment of the present invention, in step (c), by using a noise removal artificial intelligence model to pre-remove images classified as preset noise microstructure patterns from the analysis images of multiple samples, the number of analysis images to be classified in step (d) can be reduced, and the remaining analysis images can be effectively classified as target microstructure patterns using a microstructure classification artificial intelligence model.

[0050] In step (c) of the material analysis method according to an embodiment of the present invention, the noise removal artificial intelligence model may be an artificial intelligence model using a deep learning network based on a residual neural network or MobileNet.

[0051] like Figure 3 As shown, the residual neural network artificial intelligence model can learn the amount of change in each layer by introducing skip connections, thereby learning the difference between the previous layer and the next layer, thus solving the problem of information loss that occurs as the depth of the neural network increases.

[0052] Furthermore, according to an embodiment of the present invention, the features of epoxy resin images (which exhibit relatively simple forms compared to the complex microstructures of microscopic components) are effectively extracted through the skip connection structure of the residual neural network, and these features are cascaded to the next layer. Even as the depth of each layer increases, the extracted features do not become blurred, thus enabling the classification of epoxy resin images with high accuracy.

[0053] Furthermore, although not shown in the figure, MobileNet (which is designed to analyze images on mobile devices in real time) is an AI model that maintains a small model size and low computational cost while ensuring high accuracy and can quickly and accurately classify only epoxy resin images.

[0054] In step (d) of the material analysis method according to an embodiment of the present invention, the microstructure classification artificial intelligence model can be an artificial intelligence model using an Inception-based deep learning network. More specifically, it can be an artificial intelligence model using an Inception V3-based deep learning network.

[0055] like Figure 4 As shown, the Inception V3 AI model can include Inception modules composed of a combination of convolutional layers, average pooling layers, max pooling layers, concatenation, dropout, and softmax. Each Inception module can reduce computational cost by computing convolutional layers segmented into sizes such as 1×1, 3×3, and 5×5, and extract features from the input image at multiple stages.

[0056] Furthermore, according to embodiments of the present invention, by using segmented convolutional layers (which are characteristic of the Inception structure), more diverse features can be discovered compared to artificial intelligence models with single layers, thereby enabling efficient classification of microstructure images of coal microcomponents containing various microstructure patterns.

[0057] Therefore, when using a single artificial intelligence model to analyze sample images, the accuracy of distinguishing epoxy resin is within 40%, which does not meet the required accuracy. However, by independently using an artificial intelligence model for noise removal to classify epoxy resin images and an artificial intelligence model for microstructure classification to classify the microstructure of micro-components, epoxy resin can be distinguished with a high accuracy of 99%, and the microstructure of micro-components can be classified with an accuracy of 94%.

[0058] Furthermore, step (f) of the material analysis method according to an embodiment of the present invention may further include (f-3) verifying the classification accuracy of the pre-trained artificial intelligence model using a confusion matrix. More specifically, step (f-3) may include verifying the classification accuracy of the noise removal artificial intelligence model and the classification accuracy of the microstructure classification artificial intelligence model using a confusion matrix.

[0059] Confusion matrices can visualize the predictions and classification data of artificial intelligence models in matrix form. For example, a confusion matrix is ​​a 2×2 matrix that combines the true / false of predicted values ​​with the true / false of actual values, resulting in combinations of true positive (TP) (where the actual positive value is the same as the predicted positive value), false positive (FP) (where a negative value is incorrectly predicted as a positive value), true negative (TN) (where the actual negative value is the same as the predicted negative value), and false negative (FN) (where a positive value is incorrectly predicted as a negative value).

[0060] At this point, the confusion matrix according to the embodiment of the present invention uses a 5×5 matrix obtained by extending the above-mentioned 2×2 confusion matrix, such as... Figure 5 As shown, the microscopic components (i.e., the components of coal visible under a microscope) can be classified into five categories. In this case, there are true values ​​(where the actual positive value is the same as the predicted positive value) along the diagonal direction from the top left to the bottom right, and the remaining values ​​may be combinations of false positives or false negatives.

[0061] like Figure 5 As shown, the composition of microscopic components in coal can be classified into vitreous coal quality, shell coal quality, fibrous coal quality, semi-fibrous coal quality, and mineral quality.

[0062] More specifically, vitreous coal qualities constitute the main part of the microstructure of coal that exhibits caking properties. Vitreous coal qualities originate from the lignin of plants and exhibit a smooth and uniform surface. The keratinous coal qualities used to control caking properties can be tissue derived from the cuticle of plant leaves or twigs. Keratinous coal qualities have high volatile matter and tar content and typically exhibit a dark brown, serrated shape. Inert coal qualities can be tissues with a three-dimensional structure; as an inert component, they do not soften or melt. The main types of inert coal qualities include fibrous coal qualities and semi-fibrous coal qualities. Minerals can be small amounts of minerals contained in coal and typically have a dark brown appearance.

[0063] In traditional visual analysis, it is difficult for ordinary workers to distinguish microscopic components, thus relying on the expertise of highly skilled workers. However, when the material analysis method according to embodiments of the present invention is applied, the microstructure of microscopic components that are difficult for ordinary workers to distinguish can be easily and quickly differentiated. For example, semi-siliceous coal quality, which has mixed characteristics of vitreous and siliceous coal quality, can be easily distinguished; and, for example, dark brown chalcanthite coal quality and mineral quality can be easily distinguished.

[0064] Furthermore, step (d) of classifying the analyzed image as the target microstructure pattern can classify the remaining analyzed image after removing the epoxy resin image as a noise microstructure pattern in step (c) as at least one of the above-mentioned vitreous coal quality, hulled coal quality, fibrous coal quality, semi-fibrous coal quality, mineral quality and combinations thereof.

[0065] Furthermore, the material analysis method according to an embodiment of the present invention may further include (e) calculating the proportion of the classification microstructure pattern of the sample after step (d). For example, the microstructure of the coal sample to be analyzed may be calculated as the vitreous coal quality, hulled coal quality, fibrous coal quality, semi-fibrous coal quality, and the percentage (%) of minerals in the total volume.

[0066] Figure 6 The table compares the results of extracting the region of interest (which is the region corresponding to the center region of the specimen image) in step (c) of the material analysis method according to an embodiment of the present invention and inputting it into a momentum stochastic gradient descent (SGDM), adaptive moment estimation (ADAM), or root mean square transfer (RMSPROP) solver (or optimizer) with the results of inputting the entire specimen image.

[0067] like Figure 6 As shown, it has been confirmed that the validation accuracy and test accuracy of the three deep learning networks have been improved to approximately 96%. Specifically, as... Figure 7 As shown, 16 images that were incorrectly classified as crustal coal when the entire sample image was input could be effectively classified when the central region was extracted.

[0068] Figure 8 To illustrate the use of class activation maps, in step (c) of the material analysis method according to an embodiment of the present invention, an artificial intelligence model sets up an image of a region of interest for analyzing a sample image.

[0069] It was found that, except for vitreous coal quality, all hull coal quality, filiform coal quality, semi-filiform coal quality, and mineral quality were analyzed using artificial intelligence to determine the regions corresponding to the central area. Therefore, in step (c) of the material analysis method according to an embodiment of the present invention, when obtaining the analysis image by extracting the region of interest (which is the region corresponding to the central area of ​​the sample image), the test accuracy can be ensured to reach approximately 96%.

[0070] Figure 9 The table in the document measures and compares the validation accuracy and test accuracy of the Residual Neural Network-50 and MobileNet V2 deep learning networks, which are deep learning networks for artificial intelligence models of material analysis methods according to embodiments of the present invention.

[0071] In step (c) of detecting noise (i.e., epoxy resin images), by applying the two types of deep learning networks described above to the noise removal artificial intelligence model of the material analysis method according to an embodiment of the present invention, epoxy resin can be distinguished with very high accuracy (100% verification accuracy and 99.86% test accuracy, respectively).

[0072] Figure 10 The table in the document measures and compares the validation accuracy and test accuracy of Inception V3 and other types of deep learning networks, which are deep learning networks for artificial intelligence models of material analysis methods according to embodiments of the present invention.

[0073] like Figure 10 As shown, for each deep learning network applied to the residual neural network, Inception-residual neural network, MobileNet, and Inception V3 of this invention, an initial learning rate of 0.01 or 0.001 was applied to the three types of solvers to measure validation accuracy and test accuracy. The measurement results confirm that Inception V3 exhibits the highest validation accuracy (93.18%, 92.64%) and the highest test accuracy (91.25%, 94.11%) with an initial learning rate of 0.001.

[0074] In other words, as a result of repeated experiments, the applicant has demonstrated that when an initial learning rate of 0.001 is applied to the ADAM or RMSPROP solver in the combination of the residual neural network AI model and the Inception V3 AI model, or the combination of the MobileNet AI model and the Inception V3 AI model, the microstructure of coal microcomponents can be classified with the highest accuracy.

[0075] In the following text, reference will be made to Figures 11 to 12 A material analysis apparatus 100 according to an embodiment of the present invention is described.

[0076] Figure 11 A conceptual diagram illustrating a material analysis apparatus 100 according to an embodiment of the present invention is shown. Figure 11 As shown, the material analysis apparatus 100 according to an embodiment of the present invention may include a stage 110, a stage moving device 120, an image acquisition device 130, and a microstructure classification device 140.

[0077] More specifically, the stage moving device 120 may include a stage forward and backward moving device 121 capable of moving the stage 110 in the forward and backward direction, and a stage left and right moving device 122 capable of moving the stage 110 in the left and right direction.

[0078] Furthermore, the image acquisition device 130 can capture images of the sample by setting a detection area on one surface of the sample, such as... Figure 12As shown in the diagram. In this case, the detection area can be an area with multiple detection points set to form an overall hexagon, such as... Figure 12 As shown in the image.

[0079] Therefore, the stage moving device 120 can, for example, move the stage 110 so that the image acquisition device 130 is located at the detection point of any vertex of the hexagon, and then move the stage 110 back and forth or left and right so that the image acquisition device 130 is located at the detection point of another adjacent vertex.

[0080] Traditionally, the detection area is usually set so that the detection points form a quadrilateral. However, in this invention, by setting the above-mentioned hexagonal detection area, the characteristics of the entire circular cross-section can be fully reflected. When the detection area is set to form a hexagonal or more polygonal shape, there is a problem that the detection time is prolonged due to the large number of detection points. Therefore, the most efficient analysis can be performed when the detection area is set as a hexagonal one.

[0081] Furthermore, the material analysis apparatus 100 according to an embodiment of the present invention may include a deep learning unit 150. In this case, the deep learning unit 150 may include a noise removal deep learning unit 151 and a microstructure classification deep learning unit 152. The noise removal deep learning unit 151 learns and classifies noise microstructure patterns to distinguish and classify noise appearing in the sample, and the microstructure classification deep learning unit 152 learns and classifies target microstructure patterns to distinguish and classify microstructure patterns appearing in the sample.

[0082] Therefore, using an artificial intelligence model trained with deep learning, the images of the sample detection points captured by the image acquisition device 130 can be classified into vitreous coal quality, shell coal quality, fibrous coal quality, semi-fibrous coal quality, and mineral quality.

[0083] The invention has been described with reference to the embodiments shown in the accompanying drawings, but these are merely exemplary, and those skilled in the art will understand that various modifications and equivalent embodiments can be derived therefrom. Therefore, the true scope of protection of the invention should be determined by the technical spirit of the appended claims.

Claims

1. A material analysis method, comprising: (a) Preparation of specimens for analysis of microstructure; (b) Acquire multiple images of the sample by photographing the detection area set on one surface of the sample; (c) Preprocessing multiple sample images to obtain multiple analytical images; (d) Analyze multiple analytical images using a trained microstructure classification artificial intelligence model to classify them as at least one target microstructure pattern.

2. The material analysis method according to claim 1, wherein, Step (a) includes: (a-1) Crushing the material into particles of a predetermined size; and (a-2) The particles are mixed with a binder to prepare a sample.

3. The material analysis method according to claim 2, wherein the material is coal, and in, The adhesive is epoxy resin.

4. The material analysis method according to claim 1, wherein, Step (c) includes: The region of interest set by the artificial intelligence model for analyzing sample images is extracted using class activation maps, thereby obtaining the analysis image.

5. The material analysis method according to claim 4, wherein, The region of interest is the area corresponding to the center region of the sample image.

6. The material analysis method according to claim 1, wherein, Step (c) includes: The trained noise removal AI model is used to analyze sample images of microstructure patterns of multiple unlabeled samples to remove the analyzed images classified as preset noise microstructure patterns.

7. The material analysis method according to claim 6, wherein, The noise microstructure pattern is an image in which the microstructure pattern of the sample is classified as an adhesive by a noise removal artificial intelligence model.

8. The material analysis method according to claim 6, wherein, The noise removal AI model is an AI model that uses a deep learning network based on residual neural networks or MobileNet.

9. The material analysis method according to claim 1, wherein, In step (d), the microstructure classification AI model is an AI model that uses an Inception-based deep learning network.

10. The material analysis method according to claim 1, further comprising: Prior to step (c), (f) a noise removal AI model or a microstructure classification AI model is trained using training images in which microstructure patterns of arbitrary specimens are marked.

11. The material analysis method according to claim 1, wherein, In step (d), the target microstructure pattern is an image in which the microstructure pattern of the sample is classified by a microstructure classification artificial intelligence model into at least one of vitreous coal quality, hulled coal quality, fibrous coal quality, semi-fibrous coal quality, mineral quality, and combinations thereof.

12. The material analysis method according to claim 1, further comprising: After step (d), (e) Calculate the proportion of classified microstructure patterns in the sample.

Images