Method for generating a phase extraction model for iron-containing ore, method for extracting phases from iron-containing ore, and method for producing agglomerate ore.
A method for generating a phase extraction model using image data processing and machine learning accurately extracts iron-containing ore phases, improving the production of agglomerate ore by optimizing manufacturing conditions.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- JFE STEEL CORP
- Filing Date
- 2025-10-14
- Publication Date
- 2026-06-02
Smart Images
Figure 2026090196000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for generating a phase extraction model of an iron-containing ore, a method for extracting a phase of an iron-containing ore, and a method for producing an agglomerated ore.
Background Art
[0002] The blast furnace process is an iron-making process suitable for high efficiency and mass production. As a raw material used in the blast furnace process, there is an agglomerated ore obtained by agglomerating an iron-containing ore, such as a natural iron ore. The quality of the iron-containing ore is one of the factors determining the operating state of the blast furnace, and the structure of the iron-containing ore affects the quality of the iron-containing ore. For example, the structure of sintered ore, which is one type of agglomerated ore, is generally composed of pores existing inside and phases such as hematite, magnetite, calcium ferrite, slag, etc. that constitute the matrix of the sintered ore. And the structure and fraction of each phase in the structure of the iron-containing ore affect the quality of the iron-containing ore. Therefore, quantitatively grasping the state of the structure of the iron-containing ore is important for controlling the quality of the iron-containing ore and leading to stable operation of the blast furnace.
[0003] Conventionally, studies have been made to quantitatively grasp the state of the structure of an iron-containing ore. For example, in Patent Document 1, the abundance ratio of each phase in sintered ore is determined by X-ray diffraction, an image is created by microscopically photographing the polished surface of the sintered ore, and the presence region of each phase is determined by drawing contour lines of luminance in the image.
[0004] Also, in Patent Document 2, a technique for classifying the structure of sintered ore by using a learning model machine-learned by deep learning is disclosed.
Prior Art Documents
Patent Documents
[0005]
Patent Document 1
Patent Document 2
[0006] However, it was difficult to accurately extract the microstructure of sintered ore using the technique disclosed in Patent Document 1.
[0007] Furthermore, in order to classify structures with high accuracy using the method disclosed in Patent Document 2, it is necessary to capture a large number of structural images from various sintered ores and learn the phase structure from hundreds or more image data. Thus, there was a problem that a large amount of data processing was required and was time-consuming.
[0008] The present invention has been made in view of the above circumstances, and aims to provide a method for generating an extraction model for iron-containing ore phases that can extract the phases of iron-containing ore simply and accurately, a method for extracting the phases of iron-containing ore, and a method for producing agglomerate ore using the same. [Means for solving the problem]
[0009] As a result of diligent research, the inventors have found that the above objective can be achieved by adopting the following configuration.
[0010] 1. Acquisition process for obtaining microstructure image data of iron-containing ore, A labeling step in which a label indicating which phase of the iron-containing ore corresponds to each pixel constituting the tissue image data, A channel selection step of selecting at least one channel from the channels that constitute the color space of the aforementioned tissue image data, A feature calculation step of calculating feature quantities from the grayscale values of the selected channel, A method for generating a phase extraction model for iron-containing ore, comprising: a model generation step of generating a phase extraction model for iron-containing ore that takes the aforementioned features as input and the aforementioned labels as output.
[0011] 2. A method for generating an extraction model of the phase of the iron-containing ore according to 1, further comprising a tone adjustment step for adjusting the tone of the tissue image data prior to the feature calculation step, In the acquisition step described above, multiple tissue image data are acquired, A method for generating an extraction model of an iron-containing ore phase, wherein the tonal range of a plurality of tissue image data is aligned in the aforementioned tone adjustment step.
[0012] 3. In the model generation process, A candidate model is generated that takes the aforementioned features as input and the aforementioned labels as output. A method for generating an extraction model of an iron-containing ore phase according to 1 or 2, wherein an extraction model of the iron-containing ore phase is determined from the candidate model based on the accuracy of the candidate model.
[0013] 4. A method for extracting phases of iron-containing ore, comprising an extraction step of extracting phases of iron-containing ore using an extraction model of iron-containing ore generated by the method for generating an extraction model of iron-containing ore described in any of 1 to 3 above.
[0014] 5. A determination step to determine whether the manufactured agglomerate ore satisfies predetermined conditions, If it is determined in the determination step that the conditions are not met, a manufacturing condition change step is performed to change the manufacturing conditions of the agglomerated ore. A method for producing agglomerated ore, comprising a manufacturing step for producing agglomerated ore using the manufacturing conditions changed in the manufacturing condition modification step, A method for producing agglomerated ore, wherein the determination in the determination step is made based on the result of extracting the phase of the agglomerated ore by the method for extracting the phase of the iron-containing ore described in item 4 above. [Effects of the Invention]
[0015] According to the present invention, it is possible to provide a method for generating an extraction model for iron-containing ore phases that can extract iron-containing ore phases simply and accurately, a method for extracting iron-containing ore phases, and a method for producing agglomerate ore using the same. [Brief explanation of the drawing]
[0016] [Figure 1] It is a diagram showing an example of the tissue image used for generating teacher data in the example. [Figure 2] It is a diagram showing an example of the tissue image used for extracting a phase in the example.
Mode for Carrying Out the Invention
[0017] Hereinafter, the present invention will be described. The following description shows a preferred embodiment of the present invention, and the present invention is not limited by the following description in any way.
[0018] <Method for Generating Phase Extraction Model of Iron-Containing Ore> The method for generating a phase extraction model of iron-containing ore according to an embodiment of the present invention includes an acquisition step, a labeling step, a channel selection step, a feature quantity calculation step, and a model generation step.
[0019] (Iron-containing ore) The extraction model according to the present embodiment extracts the phase of iron-containing ore. The iron-containing ore used for generating the extraction model is not particularly limited, and any iron-containing ore can be used. The iron-containing ore may be, for example, massive ore, iron ore (natural iron ore), etc. Examples of iron ore include lumpy ore. Examples of massive ore include sintered ore and pellets. Sintered ore is produced by sintering raw materials containing iron ore (especially powdered ore). On the other hand, pellets are produced by granulating raw materials containing iron ore (especially powdered ore) and binding and solidifying them thermally or cold, and making the shape, for example, spherical. The number of iron-containing ores used for generating the extraction model is not particularly limited, and it may be one, or two or more.
[0020] (Phase) The phases of iron-containing ore are, in other words, the phases that constitute the structure of the iron-containing ore. Examples of phases of iron-containing ore include hematite phase, magnetite phase, calcium ferrite phase, slag phase, and pores. In this invention, pores are included in the phases of iron-containing ore. The structure of agglomerate ore generally contains the above five phases, and may also be composed of the above five phases. Other examples of phases of iron-containing ore include other-component calcium ferrite phase, wustite phase, and goethite phase. That is, the structure of iron-containing ore may further contain one or more of the other-component calcium ferrite phase, wustite phase, and goethite phase.
[0021] (Acquisition process) In the acquisition process, microstructure image data of the iron-containing ore is acquired. The microstructure image data is not particularly limited as long as it is image data that can identify the microstructure of the iron-containing ore, but it can be acquired by taking a microstructure image of the iron-containing ore. Here, an example of how to take a microstructure image is described. For example, first, if necessary, the iron-containing ore sample is crushed, and the crushed sample is then prepared to be used as an observation surface by, for example, polishing, cutting, or laser, and the observation surface is polished until it becomes mirror-like. At this time, for irregularly shaped and easily crumbled samples, it is preferable to embed the sample in resin beforehand. Next, the observation surface of the polished sample is observed and a microstructure image is taken. The microstructure image can be taken using, for example, a microscope. Examples of microscopes include optical microscopes, scanning electron microscopes (SEMs), and stereomicroscopes. When taking a microstructure image using an SEM, the type of image is not limited and can be a secondary electron image or a backscattered electron image, etc. The imaging conditions when taking a microstructure image, such as magnification, are not particularly limited and can be set arbitrarily, but it is preferable to use imaging conditions that allow the operator to identify the phases in the entire microstructure image.
[0022] The tissue image is not particularly limited, but it is preferable that it be a color image. A color image may provide more information for extracting the phase.
[0023] Tissue image data consists of multiple pixels, and each pixel contains information representing color. In tissue image data, color is represented by the tonal values in the channels that make up the color space (or, if the color space consists of two or more channels, by the combination of tonal values in each channel). When the tissue image is a color image, examples of color spaces include RGB, HSV, HSL, and CMY.
[0024] The number of tissue image data points is not limited and may be as little as one, but multiple points are preferable to improve accuracy. Multiple tissue image data points can be obtained by taking multiple tissue images.
[0025] (Labeling process) In the labeling step, a label is assigned to each pixel constituting the microstructure image data acquired in the acquisition step, indicating which phase of the iron-containing ore that the pixel corresponds to. A label is assigned to each pixel constituting the microstructure image. The type of phase represented by the label may be at least one phase constituting the microstructure of the iron-containing ore, or it may be all phases constituting the microstructure of the iron-containing ore. The type of label is not limited, but examples include numerical data where the numerical value corresponds to the type of phase, such as 1 for pores and 2 for the hematite phase.
[0026] The specific method used in the labeling process is not limited. For example, a computer may display a tissue image (or an equivalent image) based on the tissue image data acquired in the acquisition process, and may also accept input of the phase type corresponding to each pixel constituting the tissue image data. The operator simply inputs the phase type to the computer. The computer then assigns a label based on the input. The computer may further superimpose the input phase type onto the displayed tissue image, etc., and may accept further input. The manner in which the phase types are displayed is not limited, but for example, each phase type may be colored with a predetermined color.
[0027] (Channel selection process) In the channel selection process, at least one channel is selected from the channels that make up the color space of the tissue image data. For example, in the RGB color space, one of R (red), G (green), B (blue), RG, GB, RB, or RGB can be selected. Alternatively, depending on the color space, at least one channel may be selected from H (hue), S (saturation), V (value), L, C, M, Y, etc.
[0028] Tissue images differ in appearance (color, etc.) depending on imaging conditions such as the type of microscope and lens used. Therefore, by selecting different channels for each imaging condition, it is possible to evaluate using images that are highly relevant to the phase to be extracted, further improving extraction accuracy.
[0029] (Feature extraction process) In the feature extraction process, features are calculated from the tonal values of the channels that constitute the color space of the tissue image data selected in the selection process. For example, if only R is selected, features can be calculated from the tonal values of the R channel, and these features correspond to the features of the R image. If multiple channels are selected in the selection process, the tonal values of each channel can be used as the subject of the calculation. For example, if RG is selected in the selection process, features can be calculated from the tonal values of the R channel and the G channel, and these features correspond to the features of the RG image.
[0030] The type of feature is not particularly limited, but it is preferable to use features that reflect the information of the phase to be extracted, for example, (1) to (9) below. The computer may have filters that can calculate features implemented, and it is preferable that the features are calculated using at least one of these filters.
[0031] (1) Identity features The identity feature is a feature that represents the grayscale value itself.
[0032] (2) Mean features The Mean feature is a feature that represents the average value of the grayscale values within a given range (number of pixels x × number of pixels y).
[0033] (3) Gaussian features A Gaussian feature is a feature that represents the average value of grayscale values, with weights increasing as the pixel gets closer to the center of a given range (x pixels × y pixels). Here, the weighting can be set arbitrarily. However, typically, a weighting based on a single Gaussian function is set.
[0034] (4) Median features The median feature is a feature that represents the median value of the grayscale values within a given range (number of pixels x × number of pixels y).
[0035] (5) Max feature The Max feature is a feature that indicates the maximum value of the grayscale value within a given range (number of pixels x × number of pixels y).
[0036] (6) Min Features The Min feature is a feature that indicates the minimum value of the grayscale value within a given range (number of pixels x × number of pixels y).
[0037] (7) Prewitt features Prewitt features are one type of feature obtained by performing a difference calculation on the grayscale values within a predetermined range (number of pixels x number of pixels y). Specifically, the difference in grayscale values between pixels located horizontally to the left and right within the above range, and the difference in grayscale values between pixels located vertically above and below within the above range are calculated. Feature quantities can be obtained using these differences.
[0038] (8) Sobel features Sobel features are a type of feature obtained by performing a difference calculation on the grayscale values within a predetermined range (number of pixels x number of pixels y). Specifically, the difference in grayscale values between pixels located horizontally to the left and right within the above range is calculated, and weighting is applied so that the contribution of pixels closer to the center is greater. Similarly, the difference in grayscale values between pixels located vertically above and below within the above range is calculated, and weighting is applied so that the contribution of pixels closer to the center is greater. Features can be obtained using these differences.
[0039] (9) Laplacian features Laplacian features are a type of feature obtained by performing a difference calculation on the grayscale values within a predetermined range (number of pixels x number of pixels y). Specifically, the difference in grayscale values between the pixel at the center of the range and the pixels at eight surrounding positions is calculated, and weighting is applied so that the contribution of pixels close to the center is larger. Features can be obtained using these differences.
[0040] Regarding the features described in (2) to (9) above, it is preferable that the number of pixels x and y be 2 or more in order to remove unavoidable noise contained in tissue image data. On the other hand, if x and y are made too large, information about the type of phase may be lost due to the influence of adjacent different phases, and the phase boundaries may become unclear, so it is preferable that x and y be 10 or less.
[0041] The above-mentioned features are classified into features (1) to (6) above, which deal with the grayscale values themselves, and features (7) to (9) above, which deal with the differences in grayscale values. It is preferable to use at least one feature selected from (1) to (6) above. It is also preferable to use at least one feature selected from (7) to (9) above, as this improves the accuracy of identifying the boundaries of the phases to be extracted. As features, features that combine operations to obtain each feature (features calculated by applying two or more filters) may be used.
[0042] The calculated features are not limited to one type; multiple types of features may be calculated. Increasing the number of feature types increases the amount of data in the training set for model generation, which can further improve the accuracy of the model. For example, it is preferable to use both at least one feature selected from (1) to (6) above and at least one feature selected from (7) to (9) above.
[0043] (Model generation process) In the model generation process, a phase extraction model for iron-containing ore is generated, using the features calculated in the feature calculation process as input and the labels assigned in the labeling process as output. The phase extraction model for iron-containing ore is a model that extracts phase regions from tissue image data based on the relationship between the above features and labels; in other words, it is a classification model for iron-containing ore phases. The phase extraction model for iron-containing ore can be generated by machine learning. The machine learning method is not particularly limited, but it is preferable to use decision trees, random forests, gradient boosting, or neural networks. It is more preferable to use random forests because they can learn even with a small amount of training data.
[0044] In the model generation process, a model is generated that takes the features calculated in the feature calculation process as input and the labels assigned in the label assignment process as output. This generated model may then be used as the extraction model for the phases of iron-containing ore. However, to further improve accuracy, it is preferable to generate candidate models that take the aforementioned features as input and the labels as output, and then determine the extraction model for the phases of iron-containing ore from the candidate models based on the accuracy of the candidate models.
[0045] When determining an extraction model for iron-containing ore phases from candidate models, the accuracy of the candidate model can be determined by using microstructure image data in which the aforementioned labels have been pre-assigned to each of the constituent pixels. Specifically, feature quantities are calculated from the above-mentioned microstructure image data using the method described above, input into the candidate model, and labels are output. Next, for each pixel, the pre-assigned labels and the labels output using the model are compared to determine whether they are correct or incorrect. Then, the accuracy rate (%) for each phase is calculated from the accuracy determination of each pixel. A higher accuracy rate for the phase to be extracted indicates higher accuracy.
[0046] One or more candidate models may be generated. When generating multiple candidate models, the generation of candidate models may be repeated by changing the conditions of at least one of the following steps: the acquisition step, the arbitrary gradation adjustment step, the channel selection step, and the feature calculation step. For example, the conditions of the channel selection step may be changed, specifically by changing the channels or combinations of channels selected. To improve accuracy, it is preferable to generate two or more candidate models with different selected channels or combinations, and more preferably three or more candidate models. On the other hand, there is no particular upper limit to the number of candidate models with different selected channels or combinations, but it is preferable to be 10 or less. Here, when generating multiple candidate models with different selected channels or combinations, the feature calculation methods may be the same or different. For example, the pixel counts x and y used to calculate the features may be the same or different for all candidate models. Also, for example, the conditions of the feature calculation step may be changed, in which case at least one of the type or number of features may be changed. For example, the number of features may be increased.
[0047] The specific method for determining the phase extraction model for iron-containing ore from candidate models is not limited. For example, the accuracy of one or more candidate models can be calculated, and if an arbitrarily set accuracy rate cannot be obtained, candidate models can be generated again using the same method as when generating multiple candidate models, and the generation of candidate models can be repeated until the accuracy rate is obtained. Another example is to create multiple candidate models, compare their accuracy, and determine the one with the highest accuracy as the phase extraction model for iron-containing ore. Yet another example is that if the accuracy of a candidate model is low, microstructure image data can be obtained for an iron-containing ore different from the one used to create the candidate model, and a newly created candidate model using the obtained microstructure image data can be determined as the extraction model. Here, the iron-containing ore from which microstructure image data is obtained can be an iron-containing ore corresponding to the iron-containing ore that is the target of phase extraction, for example, one with similar components or component ratios, or one with the same manufacturing conditions.
[0048] (Gradation adjustment process) The method for generating an extraction model of an iron-containing ore may further include a tone adjustment step for adjusting the tone of the tissue image data. The tone adjustment step is performed prior to the labeling step. In the tone adjustment step, the tone of multiple tissue image data is aligned.
[0049] To standardize the tonal range, the mean and standard deviation of the tonal values calculated for each pixel constituting the tissue image data should be standardized. These tonal values may be the tonal values of each channel constituting the color space of the tissue image data, or they may be the tonal values of each channel constituting another color space obtained by converting the color space of the tissue image data. Examples include the grayscale value in a grayscale image, the tonal values of R, G, and B in the RGB color space, and the luminance value. Furthermore, by standardizing the standard deviation of the tonal values, the contrast of the tissue image can be standardized.
[0050] The mean and standard deviation of the above-mentioned grayscale values can be obtained by calculating the above-mentioned grayscale values for all pixels that make up the tissue image data. Then, the grayscale of the tissue image data can be adjusted so that the same mean and standard deviation are obtained among multiple tissue image data. For example, the grayscale of the tissue image data can be made uniform by performing calculations for grayscale adjustment on the tissue image data and changing the coefficients of the calculation formula for each tissue image data. For example, when using luminance as the grayscale, first, one tissue image data can be selected and its mean and standard deviation of luminance values can be calculated. Next, calculations can be performed to adjust the luminance and contrast of the other tissue image data so that the mean and standard deviation of the luminance values of the other tissue image data are the same as those of the first selected tissue image data. The mean and standard deviation of grayscale values can be adjusted in a similar manner for other grayscale values as well.
[0051] Adjustments can be made for multiple types of grayscale. In that case, a histogram should be created for each grayscale value, and the mean and standard deviation should be calculated for each of the above grayscale values.
[0052] By incorporating a gradation adjustment process, it becomes possible to homogenize images based on statistical information, further improving the accuracy of the extraction model. For example, depending on how the light source of an optical microscope strikes the tissue, it may be possible to obtain both tissue images with many pixels with low V (brightness) (dark) and tissue images with many pixels with high V (bright). However, the overall bias of the histogram is often not important information when extracting phases from a tissue image. Therefore, by homogenizing the brightness according to this process, for example, the accuracy of the extraction model can be further improved.
[0053] To achieve the above effects, the grayscale adjustment process is performed prior to the feature calculation process. It is preferable to perform the grayscale adjustment process prior to the labeling process because it makes it easier for the operator to identify phases and improves the accuracy of the assigned labels.
[0054] <Method for extracting phases from iron-containing ore> Another embodiment of the present invention provides a method for extracting phases from iron-containing ore, comprising an extraction step. According to this embodiment, phases of iron-containing ore can be extracted simply and with high accuracy. The method according to this embodiment is a method for extracting regions from microstructure image data, in other words, a method for classifying phases of iron-containing ore.
[0055] (extraction process) In the extraction process, the phases of iron-containing ore are extracted (classified) using the iron-containing ore phase extraction model generated by the method described above. Specifically, for the iron-containing ore from which the phases are to be extracted, tissue image data is acquired, and feature quantities are calculated from the acquired tissue image data using the method described above. By inputting the calculated feature quantities into the extraction model, a label is output for each pixel constituting the tissue image data. The output format for the extracted results is not particularly limited and may be either a text file or an image file.
[0056] In this embodiment, the iron-containing ore used for phase extraction is not particularly limited, but it is preferable to use multiple iron-containing ores with varying manufacturing conditions. By changing the manufacturing conditions when natural iron ore is converted into agglomerated ore, the grain size, properties, etc., can be adjusted. Furthermore, according to this embodiment, the relationship between manufacturing conditions, etc., and phase composition can be understood, making it possible to find the phase composition and manufacturing conditions necessary to obtain the target characteristics.
[0057] The type of phase to be extracted is not limited, but at least one phase should be extracted. It is preferable to extract all phases that constitute the structure of the iron-containing ore.
[0058] Feature calculation can be performed in accordance with the acquisition process, tone adjustment process, channel selection process, and feature calculation process described above. That is, when calculating features, the same processing as the acquisition process, tone adjustment process, channel selection process, and feature calculation process should be performed on the iron-containing ore targeted for phase extraction. For example, for the iron-containing ore targeted for phase extraction, tissue image data can be obtained by taking a tissue image under the same imaging conditions as used in the acquisition process described above. Furthermore, in order to further improve accuracy, the tone of the tissue image data may be adjusted in the same way as the processing performed in the tone adjustment process described above.
[0059] <Method for producing agglomerated ore> Another embodiment of the present invention provides a method for producing agglomerated ore, comprising a determination step, a manufacturing condition modification step, and a manufacturing step.
[0060] (Judgment process) In the determination step, it is determined whether the manufactured agglomerate ore satisfies predetermined conditions. Furthermore, the determination in the determination step is made based on the results of extracting the phase of the agglomerate ore using the iron-containing ore phase extraction method described above.
[0061] The determination method is not limited, but for example, values such as phase fractions can be obtained from the results of extracting the phases of the agglomerated ore (in other words, labels output from the agglomerated ore phase extraction model), and it can be determined whether these values are within a predetermined range or are within a predetermined range. Here, the values obtained from the results of extracting the phases of the agglomerated ore are not limited to phase fractions; for example, the particle size of the particles constituting the phases may be used. Furthermore, the determination conditions are not limited, but for example, it can be determined whether the values such as phase fractions satisfy the conditions for obtaining the target characteristics (strength, etc.). The conditions for obtaining the target characteristics may be determined from the correlation between the value and the characteristic. The aforementioned relationship can be obtained for multiple agglomerated ores by determining the phase fractions of the phases of the agglomerated ore and measuring the characteristics of the agglomerated ore, and then performing linear approximation, nonlinear approximation, deep learning using a neural network, etc. However, it may also be obtained by other methods; for example, if the relationship between phase fractions and characteristics is conventionally known, that can be used.
[0062] (Manufacturing condition change process) In the manufacturing condition change process, if the conditions are determined not to be met in the judgment process, the manufacturing conditions for agglomerate ore are changed.
[0063] The manufacturing conditions to be changed in the manufacturing condition change process are not particularly limited and may include, for example, the mixing ratio of raw materials, the ratio of components in the mixed raw materials, the particle size of the raw materials, or the granulation time, firing temperature, firing time, etc.
[0064] In the manufacturing condition change process, the manufacturing conditions should be changed so that the conditions in the judgment process are satisfied. The specific method for doing so is not limited, but for example, the manufacturing conditions can be changed based on the relationship between the manufacturing conditions and the judgment criteria (e.g., phase fraction) in the judgment process. This relationship can be obtained by determining the values of the phase fraction used as the judgment criteria for multiple agglomerates with different manufacturing conditions. However, it may also be obtained by other methods; for example, if the relationship between manufacturing conditions and the phase fraction is known, that can be used.
[0065] (manufacturing process) In the manufacturing process, agglomerate ore is produced using the manufacturing conditions modified in the manufacturing condition modification process. In other words, when producing sintered ore by sintering raw materials containing iron ore, or when producing pellets by granulating raw materials containing iron ore and binding and solidifying them hot or cold, the manufacturing conditions modified in the manufacturing condition modification process are used. If the conditions are determined to be met in the determination process, the agglomerate ore may be produced using the conventional manufacturing conditions. This makes it possible to produce agglomerate ore with the desired properties. [Examples]
[0066] The present invention will be described below based on examples.
[0067] (Example 1) First, an extraction model was generated using the method according to the present invention, and it was confirmed that the extraction model could easily and accurately extract the phase of sintered ore as an iron-containing ore.
[0068] Sintered ore 1 was used as the iron-containing ore. First, the collected sintered ore 1 was split in half to expose the cross-section, which was then embedded in resin and polished. The polishing was carried out in the following order: First, it was polished with SiC paper from #120 to #400. Next, it was polished sequentially with diamonds of 9 μm, 3 μm, 1 μm, and 0.25 μm. Finally, it was polished with colloidal silica.
[0069] A cross-section of the polished sintered ore 1 was captured in color at 50x magnification using an optical microscope equipped with imaging capabilities, and acquired as microstructure image data of the sintered ore. In addition, four microstructure image data (microstructure image data 1 to 4) were acquired by changing the observation area of the sintered ore (acquisition process). An example of the acquired microstructure image is shown in Figure 1. Here, a histogram of the brightness (represented by a grayscale value from 0 to 255) of each pixel constituting the microstructure image data was obtained, and the brightness values of the microstructure image data were adjusted so that the mean value was 75 and the standard deviation was 16 (grayscale adjustment process).
[0070] For each obtained tissue image, a label was assigned to each pixel constituting the tissue image data, indicating which phase of the iron-containing ore that the pixel corresponds to. The operator entered the labels into a computer, and the computer assigned labels based on the input (labeling process). In this process, one of the following labels was assigned to each pixel: hematite phase, magnetite phase, calcium ferrite phase, slag phase, or pores.
[0071] Of the obtained tissue image data, three fields of view (Tissue Image Data 1-3) were used for model creation, and the remaining field of view (Tissue Image Data 4) was used to validate the obtained model. For Tissue Image Data 1-3, the R and G channels were first selected from the RGB color space (channel selection step). Next, identity features, Gaussian features, and Sobel features were calculated using filters on the grayscale values of the R and G channels (feature calculation step). Subsequently, machine learning using a random forest was performed with the above three types of features (features of the RG image) as input and labels as output to generate trained model A. The above operations were similarly performed when the G and B channels were selected, when the R and B channels were selected, and when the R, G, and B channels were selected to generate trained models B, C, and D. In other words, feature calculation and model generation were performed similarly for GB images, RB images, and RGB images.
[0072] For the obtained models A to D, features were calculated and input using tissue image data 4 under the same conditions as when the models were generated. The labels output by the models were compared with the labels assigned by the aforementioned workers, and the accuracy rate was calculated for each phase. The results are shown in Table 1.
[0073] [Table 1]
[0074] Table 1 shows that Model A exhibited the highest accuracy for the tissue image data 4 used in this study. Therefore, Model A was selected as the phase extraction model for sintered ore 1 (model generation process).
[0075] Using Model A, microstructure image data was acquired under similar imaging conditions for sintered ore 2, which was produced under different conditions (change in the amount of coagulant) than sintered ore 1 used to generate the model, and phase extraction was performed (extraction process). An example of the microstructure image used is shown in Figure 2.
[0076] Here, the brightness of the acquired tissue image data was adjusted under the same conditions as when the model was generated. Then, identity features, Gaussian features, and Sobel features were calculated from the grayscale values of the R and G channels of the tissue image data. These three types of features (RG image features) were input into Model A to output labels. Additionally, a worker visually classified the phases within the tissue images and assigned labels to the tissue image data. The labels extracted by the model and the labels assigned by the worker were compared to calculate the accuracy rate. The obtained accuracy rates are shown in Table 2. Using the model obtained with this method, phase extraction was achieved with an accuracy of over 90%.
[0077] [Table 2]
[0078] (Example 2) Next, it was confirmed that agglomerate possessing the desired properties can be produced by manufacturing agglomerate according to the method of the present invention.
[0079] First, sintered ore 3 was prepared as an iron-containing ore, and using Model A described in Example 1, tissue image data of four fields was acquired under the same imaging conditions as sintered ore 1, and phase extraction was performed (extraction step). Here, sintered ore 3 is a sintered ore produced under the same manufacturing conditions as sintered ore 1, but with a different manufacturing chance. Specifically, in the extraction step, the brightness of each acquired tissue image data was adjusted under the same conditions as when the model was generated. Then, identity features, Gaussian features, and Sobel features were calculated from the grayscale values of the R and G channels of each tissue image data. Labels were output by inputting the three types of features (RG image features) calculated above into Model A. Then, based on the obtained labels, the areas of hematite, magnetite, calcium ferrite, slag, and pores were calculated, and the phase fraction was calculated for each tissue image data.
[0080] In a non-patent document (Tetsu-to-Hagane Vol. 72 (1986), No. 7, pp. 775-782), it was reported that the ratio S of the area of calcium ferrite to the sum of the areas of hematite, magnetite, and calcium ferrite correlates with the reduction rate, which is one of the important characteristics of sintered ore. Specifically, it is thought that increasing the above ratio S, specifically to 20.0% or more, will increase the reduction rate. Therefore, the above ratio S was calculated for each microstructure image using the phase fraction obtained for sintered ore 3. The results are shown in Table 3. Next, it was determined whether the average of the above ratio S for each microstructure image was 20.0% or more. From Table 3, the average of the above ratio S was 18.9%, and the condition was not met (determination step).
[0081] [Table 3]
[0082] The manufacturing conditions for the sintered ore were changed because the conditions were not met during the evaluation process (manufacturing condition change process). The change involved increasing the oxygen concentration, referring to the relationship between oxygen concentration in gaseous fuel and calcium ferrite, as described in non-patent literature (Tetsu-to-Hagane Vol. 109 (2023), No. 4, pp. 235-244). In the production of sintered ore 3, air, i.e., gaseous fuel with an oxygen concentration of 21 vol%, was used as the gaseous fuel, but the condition was changed to using gaseous fuel with an oxygen concentration of 30 vol%. Then, sintered ore 4 was produced based on the changed manufacturing conditions (manufacturing process). No other conditions, such as the mixing ratio of raw materials, the component ratio in the mixed raw materials, the particle size of the raw materials, the granulation time, the firing temperature, or the firing time, were changed.
[0083] For the obtained sintered ore 4, microstructure images of four fields were acquired, similar to sintered ore 3, and phase extraction was performed. The ratio S calculated from the obtained phase fractions is shown in Table 4. From Table 4, it can be seen that sintered ore 4 satisfied the conditions in the above determination process. Therefore, it is considered that the reduction rate was improved. From the above, it was found that the present invention can also contribute to improving the quality of agglomerated ore.
[0084] Table 4
Claims
1. The acquisition process involves obtaining microstructural image data of iron-containing ore, A labeling step in which a label indicating which phase of the iron-containing ore corresponds to each pixel constituting the tissue image data, A channel selection step of selecting at least one channel from the channels that constitute the color space of the aforementioned tissue image data, A feature calculation step of calculating feature quantities from the grayscale values of the selected channel, A method for generating a phase extraction model for iron-containing ore, comprising: a model generation step of generating a phase extraction model for iron-containing ore that takes the aforementioned features as input and the aforementioned labels as output.
2. A method for generating a phase extraction model of an iron-containing ore according to claim 1, further comprising a tone adjustment step for adjusting the tone of the tissue image data prior to the feature calculation step, In the acquisition step described above, multiple tissue image data are acquired, A method for generating an extraction model of an iron-containing ore phase, wherein the tonal range of a plurality of tissue image data is aligned in the aforementioned tone adjustment step.
3. In the aforementioned model generation process, A candidate model is generated that takes the aforementioned features as input and the aforementioned labels as output. A method for generating an extraction model of the phase of an iron-containing ore according to claim 1, wherein an extraction model of the phase of the iron-containing ore is determined from the candidate model based on the accuracy of the candidate model.
4. In the aforementioned model generation process, A candidate model is generated that takes the aforementioned features as input and the aforementioned labels as output. A method for generating an extraction model of the phase of an iron-containing ore according to claim 2, wherein an extraction model of the phase of the iron-containing ore is determined from the candidate model based on the accuracy of the candidate model.
5. A method for extracting phases of iron-containing ore, comprising an extraction step of extracting phases of iron-containing ore using an extraction model of iron-containing ore generated by the method for generating an extraction model of iron-containing ore according to any one of claims 1 to 4.
6. A determination step to determine whether the manufactured agglomerate ore satisfies predetermined conditions, If it is determined in the determination step that the conditions are not met, a manufacturing condition change step is performed to change the manufacturing conditions of the agglomerated ore. A method for producing agglomerated ore, comprising a manufacturing step for producing agglomerated ore using the manufacturing conditions changed in the manufacturing condition modification step, A method for producing agglomerated ore, wherein the determination in the determination step is performed based on the result of extracting the phase of the agglomerated ore by the method for extracting the phase of the iron-containing ore described in claim 5.