JP7417095B2 - Sintered ore structure learning device, structure learning method, and structure learning program - Google Patents

Description

The present invention relates to a structure learning device, a structure learning method, and a structure learning program for learning the structure of sintered ore.

Sintered ore is a raw material that accounts for 80% of what is charged into a blast furnace, and is the main raw material that is charged into a blast furnace. The quality of sintered ore is one of the factors that determines the operational status of a blast furnace, and is determined by the pore structure of sintered ore and the composition of minerals forming the matrix. In self-fluxing sinter, which is currently widely used as a raw material for blast furnaces, the substrates are hematite (Fe ₂ O ₃ ), magnetite (Fe ₃ O ₄ ), calcium ferrite (CaO・Fe ₂ O ₃ ), and silicate slag (CaO・It has four mineral phases: SiO ₂ ). Data obtained by quantitatively measuring the structural composition of sintered ore, such as the proportion of mineral phases and pores in sintered ore, diameter distribution, and shape, is useful for gaining knowledge about quality control of sintered ore. Useful.

A technology for quantifying the structure of sintered ore is developed by analyzing the cross-sectional images of the sintered ore observed with an optical microscope, scanning electron microscope (SEM), etc. using an imaging device. known (for example, see Patent Documents 1 and 2). In the techniques described in Patent Documents 1 and 2, the structure of the sintered ore is estimated based on the brightness of pixels forming an image of the sintered ore. However, since the brightness of images differs depending on the imaging conditions and the imaging location of the sample, in the techniques described in Patent Documents 1 and 2, the estimated structure of the sintered ore differs from image to image, and the sintered ore It is not easy to estimate the texture of ores with high accuracy.

Furthermore, a technique is known for quantifying the positional relationship of mineral phases from a cross-sectional image of a cross-section of sintered ore (see, for example, Non-Patent Document 1 and Patent Document 3). Non-Patent Document 1 describes the correlation between the contact area between magnetite and silicate slag, the contact area between calcium ferrite and silicate slag, and the reduction pulverizability of sintered ore. Patent Document 3 discloses that an adjacency ratio indicating the proportion of mineral layers contained in the sintered ore adjacent to each other is calculated from a cross-sectional image of the cross section of the sintered ore taken by an imaging device, and the adjacency ratio is calculated based on the calculated adjacency ratio. Techniques for estimating the properties of ores are described.

On the other hand, there is a known technology that uses image data representing an image as training data and uses a learning model that is machine learned using deep learning to identify various features included in an image with high accuracy ( For example, see Patent Document 4). However, there is no known technique for estimating the structure of sintered ore using a machine-learned learning model.

Japanese Patent Application Publication No. 2013-122403 Japanese Patent Application Publication No. 2014-137344 Japanese Patent Application Publication No. 2014-215987 JP 2019-200769 Publication

“Relation between quality and morphology of sinter -texture analysis contribution”, D. Jeulin: Ironmaking & Steelmaking, 10(1983), p.145.

The inventors of the present invention have discovered that when a learning model uses cross-sectional images of sintered ore to machine learn the structure of sintered ore, by appropriately selecting cross-sectional images to be used as training data, We have found that we can generate a learning model that provides information that allows highly accurate estimation of ore properties. More specifically, the inventors of the present invention have developed a learning model that is capable of estimating not the structure of the mineral phase of sintered ore contained in sintered ore, but the structure of a composite structure formed by the mineral phase. It was discovered that by generating sintered ore, information from which the properties of sintered ore can be estimated can be provided.

An object of the present invention is to provide a structure learning device capable of generating a learning model capable of estimating the structure of sintered ore with high accuracy.

The present invention, which solves these problems, is based on a sintered ore microstructure learning device, a microstructure learning method, and a microstructure learning program shown below.
(1) an image data acquisition unit that acquires a plurality of image data each showing a plurality of cross-sectional images taken of a cross-section of sintered ore;
a learning data generation unit that generates a plurality of learning data by associating each of the plurality of regions of the cross-sectional image with the structure of a composite structure formed by mineral phases contained in the sintered ore;
a learning processing unit that causes a learning model to learn the relationship between the plurality of cross-sectional images and the tissue configuration of the composite tissue using the plurality of learning data;
a learning model output unit that outputs a learning model that estimates the tissue structure of the composite tissue from a cross-sectional image of a cross section of the sintered ore based on the relationship between the learned plurality of cross-sectional images and the structure of the composite tissue;
A device for learning the structure of sintered ore, comprising:
(2) The composite structure includes a first composite structure in which either hematite or magnetite is surrounded by calcium ferrite, a second complex structure in which calcium ferrite and silicate slag are combined, and either hematite or magnetite. The structure of the sintered ore according to (1), which includes a third complex structure that is a complex structure of either hematite and magnetite and a silicate slag; Configuration learning device.
(3) The plurality of learning data are generated by further associating the first mineral phase which is either hematite or magnetite, the second mineral phase which is the primary phase of calcium ferrite, and the plurality of regions. The sintered ore structure learning device according to (2).
(4) The first mineral phase includes a 10th mineral phase which is a residual source ore, an 11th mineral phase which is primary hematite or primary magnetite, and a 12th mineral phase which is secondary hematite or secondary magnetite. The device for learning the structure structure of the described sintered ore.
(5) the second mineral phase includes a 2a mineral phase that is acicular calcium ferrite, and a 2b mineral phase that is columnar calcium ferrite;
(3) or the second composite structure includes a 2a composite structure that is a composite structure of acicular calcium ferrite and silicate slag, and a 2b composite structure that is a composite structure of columnar calcium ferrite and silicate slag; or The sintered ore structure learning device according to (4).
(6) Obtaining a plurality of image data each showing a plurality of cross-sectional images of the cross-section of the sintered ore,
Generate multiple pieces of learning data by associating each of the multiple regions of the cross-sectional image with the structure of the composite structure formed by the mineral phase contained in the sintered ore,
Using multiple learning data, a learning model learns the relationship between multiple cross-sectional images and the tissue structure of the composite tissue,
Outputs a learning model that estimates the structure of a composite tissue from a cross-sectional image of a cross section of sintered ore, based on the relationship between the learned cross-sectional images and the structure of the complex structure.
A sintered ore structure learning program characterized by causing a computer to execute processing.
(7) Obtaining a plurality of image data each showing a plurality of cross-sectional images of the cross-section of the sintered ore,
Generate multiple pieces of learning data by associating each of the multiple regions of the cross-sectional image with the structure of the composite structure formed by the mineral phase contained in the sintered ore,
Using multiple learning data, a learning model learns the relationship between multiple cross-sectional images and the tissue structure of the composite tissue,
Outputs a learning model that estimates the structure of a composite tissue from a cross-sectional image of a cross section of sintered ore, based on the relationship between the learned cross-sectional images and the structure of the complex structure.
A method for learning the structure of sintered ore, comprising:

The learning model generated by the tissue structure learning device according to one embodiment can provide information that allows highly accurate estimation of the characteristics of sintered ore.

(a) is the first cross-sectional image, and (b) is the second cross-sectional image. (a) is the third cross-sectional image, (b) is the fourth cross-sectional image, (c) is the fifth cross-sectional image, and (d) is the sixth cross-sectional image. FIG. 1 is a diagram showing an organizational structure learning system including an organizational structure learning device according to an embodiment. 4 is a block diagram showing the organizational structure learning device shown in FIG. 3. FIG. 4 is a flowchart of an organizational structure learning process executed by the organizational structure learning device shown in FIG. 3. FIG. Figure 4 shows the process of associating each of a plurality of regions of a cross-sectional image with a mineral phase contained in sintered ore using a GUI displayed by the learning data generation unit, and (a) shows the first state. (b) shows the second state, (c) shows the third state, and (d) shows the fourth state. (a) is a diagram showing the source of a cross-sectional image used as learning data, and (b) is a diagram showing a cross-sectional image used as learning data. (a) is a cross-sectional image for evaluation (part 1), (b) is a cross-sectional image for evaluation (part 2), (c) is a cross-sectional image for evaluation (part 3), and (d ) is the cross-sectional image for evaluation (part 4), (e) is the output image of the learning model corresponding to the cross-sectional image shown in (a), and (f) corresponds to the cross-sectional image shown in (b). (g) is an output image of the learning model corresponding to the cross-sectional image shown in (c), and (h) is an output image of the learning model corresponding to the cross-sectional image shown in (d). be. (a) is a cross-sectional image for evaluation (part 5), (b) is a cross-sectional image for evaluation (part 6), (c) is a cross-sectional image for evaluation (part 7), and (d ) is the cross-sectional image for evaluation (part 8), (e) is the output image of the learning model corresponding to the cross-sectional image shown in (a), and (f) corresponds to the cross-sectional image shown in (b). (g) is an output image of the learning model corresponding to the cross-sectional image shown in (c), and (h) is an output image of the learning model corresponding to the cross-sectional image shown in (d). be.

DESCRIPTION OF THE PREFERRED EMBODIMENTS An organizational structure learning device, an organizational structure learning method, and an organizational structure learning program according to the present invention will be described below with reference to the drawings. However, the technical scope of the present invention is not limited to these embodiments.

(Overview of organizational structure learning device according to embodiment)
The inventors of the present invention discovered that when a learning model is used to machine learn the structure of sintered ore, a cross-sectional image of sintered ore used as training data and a machine-learned learning model are used to learn the sintered ore structure. We investigated the relationship with estimation accuracy when estimating the configuration. Specifically, the inventors of the present invention studied the classification of composite structures according to the reactions of self-fusing sintered ore based on the Fe ₂ O ₃ --SiO ₂ --CaO ternary phase diagram.

1 and 2 are diagrams showing examples of cross-sectional images used as learning data. FIG. 1(a) is the first cross-sectional image, FIG. 1(b) is the second cross-sectional image, FIG. 2(a) is the third cross-sectional image, and FIG. 2(b) is the fourth cross-sectional image. These are cross-sectional images, and FIG. 2(c) is a fifth cross-sectional image, and FIG. 2(d) is a sixth cross-sectional image. The first cross-sectional image shown in FIG. 1(a) includes a composite structure generally referred to as a high-temperature type, and the second cross-sectional image shown in FIG. 1(b) includes a composite structure generally referred to as a low-temperature type. Including organizations.

When the maximum temperature is reached in the sintered ore manufacturing process, the raw material of the sintered ore consists of about 30% to 50% hematite and about 50% to 70% melt. In the sintered ore manufacturing process, approximately 30% to 50% of the iron ore remains unmolten even when the maximum temperature is reached. Hematite that remains in the form of the original iron ore is called residual ore (also simply called original ore). The remaining source ore is the area indicated by "A" in FIG. 1(a). Hematite that is fragmented by the melt and remains in a patchy manner is called patchy hematite, recrystallized hematite, primary hematite, etc. Primary hematite will be described later.

Note that hematite transforms into magnetite at a temperature of 1350°C or higher. When the maximum temperature is reached at 1350° C. or higher in the sinter manufacturing process, magnetite appears in the iron oxide phase of the sinter.
The iron oxide phase is a general term for the crystalline phases of hematite and magnetite. Here, it may be used when neither hematite nor magnetite is specified.

The melt at the time of reaching the maximum temperature is locally heterogeneous. Most of the primary crystals that first crystallize from the melt are hematite (at 1350°C or higher, magnetite). Calcium ferrite and silicate slags may also be present locally. Here, calcium ferrite is also academically referred to as SFCA phase, and includes C2F phase, CF phase, and CF2 phase of the Fe ₂ O ₃ --SiO ₂ --CaO ternary phase diagram. The silicate slag includes a CS phase, a C2S phase, and an amorphous phase of the Fe ₂ O ₃ --SiO ₂ --CaO ternary phase diagram.

Hematite crystallized from melt exhibits a unique morphology. It is called skeletal ridge hematite or secondary hematite.
Here, the iron oxide phase is also collectively referred to as the first mineral phase. Furthermore, the remaining original ore is subdivided into a 10th mineral phase, primary hematite and primary magnetite are referred to as an 11th mineral phase, and secondary hematite and secondary magnetite are referred to as a 12th mineral phase.

When the primary crystals of calcium ferrite remain in the sintered ore, a second mineral phase, which is the primary crystal of calcium ferrite, is formed. The second mineral phase is the region indicated by "C" in FIG. 1(a).

The second mineral phase (calcium ferrite phase) may be subdivided in its form. That is, when the content of silica (SiO ₂ ) contained in the raw material of the sintered ore is relatively low and the maximum temperature in the manufacturing process of the sintered ore is relatively low, the second mineral phase has an acicular shape. present. Therefore, in particular, acicular calcium ferrite, also referred to as acicular SFCA, may be classified as the 2a mineral phase. The second a mineral phase is the region indicated by "Ca" in FIG. 2(a).

When the content of silica contained in the raw material of the sintered ore is relatively low and the maximum temperature in the manufacturing process of the sintered ore is relatively high, the second mineral phase takes on a columnar shape. Therefore, in particular, columnar calcium ferrite, also referred to as columnar SFCA, may be classified as the 2b mineral phase. The 2b mineral phase is the region indicated by "Cb" in FIG. 2(b).

After the primary crystals crystallize, various composite structures are formed depending on the melt composition at that time. When the composition of the melt near the primary crystal of iron oxide is in the calcium ferrite production region, a peritectic reaction in which calcium ferrite is produced from the primary crystal and the melt proceeds. As a result, a first composite structure is formed, which is a peritectic structure in which an iron oxide phase is surrounded by calcium ferrite. The first composite tissue is the area indicated by "B" in FIG. 1(a).

When hematite, calcium ferrite, etc. crystallize, the silica concentration of the remaining melt increases, and eventually it becomes a eutectic composition of calcium ferrite and silicate slag. Subsequent solidification proceeds through a eutectic reaction between calcium ferrite and silicate slag. As a result, a second composite structure, which is a coexisting structure in which calcium ferrite and silicate slag coexist, is formed. The second composite tissue is the area indicated by "D" in FIG. 1(a). Furthermore, on the ternary phase diagram, there is also a ternary eutectic reaction in which hematite, calcium ferrite, and silicate slag simultaneously crystallize at 1216°C, but this is not observed in the actual sintered ore structure. .

The third composite structure is formed by the coexistence of the residual source ore, primary crystal, or recrystallized phase of iron oxide, and any one of the second mineral phase and the second composite structure. The third composite structure is a composite structure of either hematite or magnetite and calcium ferrite, and is a region indicated by "E" in FIG. 1(b).

Further, in a region where the CaO concentration is low, the melt solidifies as a silicate slag after the formation of an iron oxide phase. In this case, a fourth composite structure is formed in which the iron oxide phase and the silicate slag in the melt coexist. The fourth composite structure is a composite structure of an iron oxide phase and a silicate slag, and is a region indicated by "F" in FIG. 1(b).

Table 1 shows the first mineral phase, the second mineral phase, and the first to fourth composite structures.

Furthermore, since calcium ferrite and silicate slag exhibit a complex structure in which they are intricately intertwined with each other, the inventors of the present invention determined that each of the second mineral phase and the second complex structure is composed of calcium ferrite. It was discovered that the data can be further classified according to shape and used as learning data.

The composite structure of calcium ferrite and silicate slag (second composite structure) is a composite structure of acicular calcium ferrite and silicate slag, or a composite structure of columnar calcium ferrite and silicate slag. It may also be classified as a 2b composite tissue. The second a composite tissue is the area indicated by "Da" in FIG. 2(c).
The 2b composite tissue is the area indicated by "Db" in FIG. 2(d).

Machine learning was performed using learning data generated by associating each of a plurality of regions of a cross-sectional image with the first to fourth composite tissues indicated by "B" and "D" to "F" in Fig. 1. The learning model can provide information that allows highly accurate estimation of the properties of sintered ore.

In addition, the learning model uses learning data generated by associating each of a plurality of regions of a cross-sectional image with the first to second mineral phases indicated by "A" and "C" in Fig. 1. It may also be machine learned.

The learning model also uses cross-sectional images of the 2a mineral phase, 2b mineral phase, 2a composite structure, and 2b composite structure indicated by "Ca", "Cb", "Da", and "Db" in FIG. Machine learning may be performed using learning data generated by associating each of a plurality of regions. A learning model machine-learned using learning data generated by associating each of a plurality of regions of a cross-sectional image with the 2a mineral phase, the 2b mineral phase, the 2a composite structure, and the 2b composite structure shown in FIG. can provide information that allows the properties of sintered ore to be estimated with higher accuracy.

(Configuration and functions of organizational structure learning device according to embodiment)
FIG. 3 is a diagram showing an organizational structure learning system including an organizational structure learning device according to an embodiment.

The tissue structure learning system 100 includes an optical microscope 101, an imaging device 102, and a tissue structure learning device 1. The optical microscope 101 magnifies and observes a desired region of the sintered ore 105, which has been cut so that its cross section is visible and is placed on the stage 110, at a desired magnification, such as 500 times. The area of the sintered ore 105 magnified by the optical microscope 101 is imaged by the imaging device 102. The imaging device 102 includes an image detection element such as a CCD (Charge Coupled Device) image sensor, and a storage element that stores an image detected by the image detection element, and has an area of the sintered ore 105 magnified by the optical microscope 101. Take an image.

Image data showing a cross-sectional image of the sintered ore captured by the imaging device 102 is transmitted to the tissue structure learning device 1 via the communication wiring 103. In the tissue structure learning system 100, the imaging device 102 and the tissue structure learning device 1 are communicably connected via communication wiring 103. However, in the tissue structure learning system according to the embodiment, the imaging device 102 and the tissue structure learning device 1 do not need to be communicably connected. When the imaging device 102 and the tissue structure learning device 1 are not connected, image data showing a cross-sectional image of the sintered ore is stored in a removable storage medium in the imaging device 102 and the tissue structure learning device 1, and the image data is stored in the imaging device 102 and the tissue structure learning device 1. It is moved from the device 102 to the organizational structure learning device 1.

(Configuration and functions of organizational structure learning device according to embodiment)
FIG. 4 is a block diagram showing the organizational structure learning device 1. As shown in FIG.

The organizational structure learning device 1 includes a communication section 11 , a storage section 12 , an input section 13 , an output section 14 , a processing section 20 , and a learning model 30 . The communication section 11, the storage section 12, the input section 13, the output section 14, the processing section 20, and the learning model 30 are connected to each other via a bus 15. The structure learning device 1 uses image data showing a cross-sectional image of the sintered ore captured by the imaging device 102 to learn the relationship between the cross-sectional image of the sintered ore and the structure of the sintered ore using a learning model 30. Let them learn. In one example, the organizational structure learning device 1 is a personal computer.

The communication unit 11 has a wired communication interface circuit such as Ethernet (registered trademark). The communication unit 11 communicates with the imaging device 102, a server (not shown), etc. via the communication wiring 103.

The storage unit 12 includes, for example, at least one of a semiconductor storage device, a magnetic tape device, a magnetic disk device, or an optical disk device. The storage unit 12 stores operating system programs, driver programs, application programs, data, etc. used in processing by the processing unit 20. For example, the storage unit 12 uses image data showing a cross-sectional image of the sintered ore to perform a structure-structure learning process that causes the learning model 30 to learn the relationship between the cross-sectional image of the sintered ore and the structure of the sintered ore. It stores an organizational structure learning program and the like for causing the processing unit 20 to execute. The organizational structure learning program may be installed in the storage unit 12 from a computer-readable portable recording medium such as a CD-ROM or DVD-ROM using a known setup program or the like. Furthermore, the storage unit 12 stores various data used in the organizational structure learning process.

The input unit 13 may be any device that can input data, such as a touch panel or a keyboard. The operator can use the input unit 13 to input characters, numbers, symbols, and the like. When the input unit 13 is operated by an operator, it generates a signal corresponding to the operation. The generated signal is then supplied to the processing unit 20 as an operator's instruction.

The output unit 14 may be any device that can display videos, images, etc., such as a liquid crystal display or an organic EL (Electro-Luminescence) display. The output unit 14 displays a video according to the video data supplied from the processing unit 20, an image according to the image data, and the like. Further, the output unit 14 may be an output device that prints images, images, characters, etc. on a display medium such as paper.

The processing unit 20 includes one or more processors and their peripheral circuits. The processing unit 20 centrally controls the overall operation of the organizational structure learning device 1, and is, for example, a CPU. The processing unit 20 executes processing based on programs (driver programs, operating system programs, application programs, etc.) stored in the storage unit 12. Further, the processing unit 20 can execute multiple programs (application programs, etc.) in parallel.

The processing section 20 includes an image data acquisition section 21 , an image data correction section 22 , a learning data generation section 23 , a learning processing section 24 , and a learning model output section 25 . Each of these units is a functional module realized by a program executed by a processor included in the processing unit 20. Alternatively, each of these parts may be implemented in the organizational structure learning device 1 as firmware.

The learning model 30 is a learning model that learns the features of an abstracted image through supervised learning. For example, using known machine learning techniques such as deep learning, a cross-sectional image of sinter Learn the relationship with organizational structure. Deep learning is machine learning using a multilayer neural network consisting of an input layer, a hidden layer, and an output layer. Each node of the input layer receives a respective feature vector of a cross-sectional image of the sintered ore. The feature vector may be, for example, a partial image obtained by cutting out a part of the image. Further, each of the partial images used as a feature vector may be cut out so that a portion thereof overlaps with another partial image. Each node in the middle layer outputs the sum of the values obtained by multiplying each feature vector output from each node in the input layer by a weight, and the output layer outputs the sum of the values obtained by multiplying each feature vector output from each node in the middle layer by a weight. Outputs the sum of the values multiplied by the weights. The learning model 30 learns while adjusting each weight so that the difference between the output value from the output layer and each cross-sectional image of the photographed sintered ore becomes small. The learning model 30 is, for example, "ZEN Intellesis" manufactured by ZEISS.

(Organizational structure learning process by the organizational structure learning device according to the embodiment)
FIG. 5 is a flowchart of the organizational structure learning process executed by the organizational structure learning device 1. The organizational structure learning process shown in FIG. 5 is mainly executed by the processing section 20 in cooperation with each element of the organizational structure learning device 1 based on a program stored in the storage section 12 in advance.

First, the image data acquisition unit 21 acquires a plurality of image data each representing a plurality of cross-sectional images obtained by capturing a cross-section of sintered ore (S101). The plurality of cross-sectional images corresponding to the plurality of image data images acquired by the image data acquisition unit 21 are cross-sectional images of the sintered ore captured by the imaging device 102.

Next, the image data correction unit 22 performs a predetermined correction process on each of the plurality of image data acquired in the process of S101 (S102). The image data correction unit 22 corrects the brightness of the cross-sectional image so that, for example, the highest brightness among the brightness of pixels included in the cross-sectional image corresponding to the image data becomes a predetermined brightness. The image data correction unit 22 also adjusts the cross-sectional image so that the difference in brightness between the highest brightness and the lowest brightness among the brightness of pixels included in the cross-sectional image corresponding to the image data becomes a desired brightness difference. Correct the contrast.

Next, the learning data generation unit 23 associates each of the plurality of regions of each cross-sectional image corresponding to the plurality of image data images with the mineral phase contained in the sintered ore and the composite structure formed by the mineral phase. A plurality of learning data are generated (S103). The learning data generating unit 23 displays a GUI including a cross-sectional image of the sintered ore on the output unit 14, thereby converting each of the plurality of regions of the cross-sectional image into a complex structure contained in the sintered ore according to the operator's instructions. to associate with. Moreover, the area|region with several learning data generation parts 23 may be linked|related with a pore.

FIG. 6 is a diagram showing a process of associating each of a plurality of regions of a cross-sectional image with a mineral phase or a composite structure using a GUI displayed by the learning data generation unit 23. 6(a) shows the first state, FIG. 6(b) shows the second state, FIG. 6(c) shows the third state, and FIG. 6(d) shows the fourth state. show.

First, the learning data generation unit 23 displays a learning data generation GUI including a cross-sectional image on the output unit 14 (see FIG. 6(a)). Next, the operator specifies the region of the composite tissue A on the cross-sectional image of the sintered ore included in the learning data generation GUI (see FIG. 6(b)). As shown by the arrow α in FIG. 6(b), the area of the composite tissue A is specified by the operator by specifying a part of the area.

Next, the learning data generation unit 23 estimates a region similar to the composite tissue A designated by the worker as a region of the composite tissue A, and adds regions of the composite tissue A other than those estimated by the worker. Generate a cross-sectional image. The learning data generation unit 23 displays a learning data generation GUI including the generated cross-sectional image on the output unit (see FIG. 6(c)). The region of the composite tissue A is displayed so as to be visually distinguishable from other regions, for example, by being colored in a predetermined color.

Thereafter, the learning data generation unit 23 repeats the process of interactively specifying the area of composite structure A with the operator until the area of composite structure A in the cross-sectional image of the sintered ore matches the area desired by the operator. .

When the region of the composite structure A in the cross-sectional image of the sintered ore matches the region desired by the worker, the learning data generation unit 23 instructs the worker to specify the region of the composite structure B that is different from the complex structure A. Run interactively. The learning data generation unit 23 specifies the mineral phase area by performing the same process as the process of specifying the area of the composite structure A for all the mineral phases included in the cross-sectional image. When the learning data generation unit 23 completes the process of specifying all the mineral phase or composite tissue regions included in the cross-sectional image, the learning data generation unit 23 performs the same process on other cross-sectional images in which no mineral phase or composite tissue region is specified. Execute.

The learning data generation unit 23 performs a process of specifying a region of a mineral phase or a composite structure for all the image data acquired in the process of S101 (see FIG. 6(d)). In the example shown in FIG. 6(d), the cross-sectional image includes three composite tissues, composite tissue A and composite tissue B, and one composite tissue C. When the learning data generation unit 23 completes the process of specifying the region of the mineral phase or composite structure for all the image data acquired in the process of S101, the process of S103 ends.

The plural pieces of learning data generated in the process of S103 are the learning data generated by associating a plurality of regions with a composite structure formed by either hematite or magnetite, calcium ferrite, or silicate slag. including. By associating multiple regions with a composite structure formed by either hematite or magnetite, calcium ferrite, or silicate slag, and generating training data, the structure of sintered ore can be determined with high precision. A learning model that can be estimated can be generated.

The composite tissue in which the learning data generation unit 23 is associated with a plurality of regions may include at least one of the first to fourth composite tissues indicated by "B", "D" to "F" in FIG. good. Further, the structure associated with the plurality of regions by the learning data generation unit 23 may further include a first mineral phase to a second mineral phase indicated by "A" and "C" in FIG. 1 in addition to the composite structure. good.

The second mineral phase may include a 2a mineral phase indicated by "Ca" in FIG. 2(a) and a 2b mineral phase indicated by "Cb" in FIG. 2(b). Further, the second tissue may include a 2a composite tissue indicated by "Ba" in FIG. 2(c) and a 2b composite tissue indicated by "Db" in FIG. 2(d).

The learning data generation unit 23 determines that the structures associated with the plurality of regions include the third mineral phase, the fourth mineral phase, the 2a composite structure, and the 2b composite structure, so that the acicular and columnar calcium ferrites are transformed into silicate slags. It is possible to determine whether or not the tissue is a composite tissue.

In addition, the learning data generation unit 23 subdivides the iron oxide phase (first mineral phase) into primary hematite (primary magnetite) (eleventh mineral phase) and secondary hematite (secondary hematite). (magnetite) (12th mineral phase).

Next, the learning processing unit 24 uses the learning data generated in the process of S103 to generate a plurality of cross-sectional images corresponding to the plurality of image data acquired in the process of S101 and the mineral phase and composite structure of the sintered ore. The learning model 30 is made to learn the relationship between the organization structure and the organization structure (S104). The learning model 30 learns the relationship between the cross-sectional image and the mineral phase and composite structure of the sintered ore, thereby determining the mineral phase and composite structure of the sintered ore from the cross-sectional image of the cross-section of the sintered ore. It becomes possible to perform a process of estimating the organizational structure of a person with high accuracy.

Then, the learning model output unit 25 outputs the learning model 30 subjected to machine learning in the process of S104 to, for example, a server (not shown) (S105). The learning model 30 is used to estimate the structure of sintered ore, for example, in a sintered ore manufacturing process at a steelworks.

(Operations and effects of the organizational structure learning device according to the embodiment)
The structure learning device 1 generates a learning model 30 using learning data generated by associating each of a plurality of regions of a cross-sectional image with a composite structure formed by a mineral phase contained in sintered ore as training data. . The learning model 30 generates the learning model 30 by associating multiple regions of the cross-sectional image not only with the mineral phase but also with the composite structure and using the learning data as training data. It can be estimated with accuracy. Specifically, the learning model 30 is generated by associating each of a plurality of regions of the cross-sectional image with the first to fourth composite tissues indicated by "B" and "D" to "F" in FIG. Machine learning is performed using the training data.

The tissue configuration learning device 1 also uses learning data generated by associating each of a plurality of regions of a cross-sectional image with the first mineral phase to the second mineral phase indicated by “A” and “C” in FIG. A learning model 30 is generated using the data as teaching data. Since the learning model 30 is generated by associating multiple regions of the cross-sectional image with the mineral phase in addition to the composite structure and using the learning data as training data, the learning model 30 can estimate the structure of the sintered ore with even higher accuracy. can.

The tissue configuration learning device 1 also performs learning that is generated by associating each of a plurality of regions of the cross-sectional image with the third mineral phase, the fourth mineral phase, the 2a composite structure, and the 2b composite tissue shown in FIG. A learning model 30 is generated using the training data as teacher data. The learning model 30 is generated by associating a plurality of regions of the cross-sectional image with the third mineral phase, fourth mineral phase, second a composite structure, and second b composite structure having similar shapes, and using the learning data as teacher data. The model 30 can estimate the structure of sintered ore with even higher accuracy.
In particular, by using the learning model of the present invention, the second mineral phase can be changed into a 2a mineral phase and a 2b mineral phase, and the second composite structure can be changed into a 2a composite structure and a 2b composite structure, compared to the conventional method. It can be easily subdivided with high precision and the composition ratio of each component can be quantified. As a result, it is expected that the relationship between the structure of sintered ore and its quality, especially its reducibility, will be more clearly related than ever before.
Further, by using the learning model of the present invention, it is possible to quantify not only the composition ratio of the first mineral phase but also the composition ratio of the twelfth mineral phase, which was previously impossible. As a result, the 12th mineral phase is said to control the reduction powdering property, which is one of the qualities of sintered ore, so the relationship between the structure of sintered ore and its reduction powdering property has been clarified more clearly than ever before. It is expected that it will be possible to relate to

Multiple regions of 12 cross-sectional images of sintered ore are associated with the mineral phase, composite structure, and pore structure to generate multiple pieces of learning data, and the generated learning data are used as training data. A learning model was used to learn the relationship between cross-sectional images and the structure of sintered ore. The learning model used was "ZEN Intellesis" manufactured by ZEISS.

The learning data used as training data was obtained by classifying multiple regions of a cross-sectional image of sintered ore taken at a magnification of 500 times into eight types of mineral phases, composite structures, and pores. The mineral phases include primary hematite, secondary hematite, original compact hematite, primary magnetite, secondary magnetite, columnar calcium ferrite, acicular calcium ferrite, and silicate slag. The composite structure includes a composite structure of acicular calcium ferrite and silicate slag.

FIG. 7(a) is a diagram showing the original image of the cross-sectional image used as the learning data. FIG. 7(b) is an image obtained by visually classifying the original image of FIG. 7(a) by an operator for each mineral to be classified, and is a diagram showing an image used as learning data. Each of FIGS. 7(a) and 7(b) includes 12 cross-sectional images arranged in 3 rows and 4 columns, and the image shown in FIG. 7(a) and the image shown in FIG. 7(b) are Corresponding images are placed at the same location. For example, the image placed at the upper left corner in FIG. 7(a) corresponds to the image placed at the upper left corner in FIG. 7(b).

In FIGS. 7(a) and 7(b), primary hematite is indicated by 40, secondary hematite is indicated by 41, original hematite is indicated by 42, primary magnetite is indicated by 43, The secondary magnetite is designated by 44. Further, columnar calcium ferrite is indicated by 45, acicular calcium ferrite is indicated by 46, silicate slag is indicated by 47, a composite structure of acicular calcium ferrite and silicate slag is indicated by 48, and pores are indicated by 48. is indicated by the reference numeral 49.

In Figure 7(b), primary hematite is colored orange, secondary hematite is colored gray, original hematite is colored red, primary magnetite is colored olive, and secondary magnetite is colored blue. be done. In addition, columnar calcium ferrite is colored yellow, acicular calcium ferrite is colored light blue, silicate slag is colored green, the composite structure of acicular calcium ferrite and silicate slag is colored yellow-green, and pores are colored purple. colored.

A learning model that was machine learned using the cross-sectional images shown in FIG. 7 as training data was evaluated using 300 cross-sectional images of sintered ore. The 300 cross-sectional images used to evaluate the learning model were prepared by capturing cross-sectional images at 10 locations from each of the 30 sintered ores.

8 and 9 are diagrams showing a cross-sectional image for evaluation and a cross-sectional image output by a learning model corresponding to the cross-sectional image for evaluation. The cross-sectional images shown in FIGS. 8(a) to 8(d) and 9(a) to 9(d) are images used as cross-sectional images for evaluation. 8(e) to 8(h) and 9(e) to 9(h) correspond to the cross-sectional images shown in FIGS. 8(a) to 8(d) and 9(a) to 9(d), respectively. The image output by the learning model is shown below.

The cross-sectional images shown in FIGS. 8(a) to 8(d) and 9(a) to 9(d) are cross-sectional images of sintered ore taken at a magnification of 500 times. The learning model, as shown in Figures 8(a) to 8(d) and 9(a) to 9(d), colors the structure of the sintered ore in the same color as the structure shown in Figure 7(b). and output it.

Table 2 shows the comparison results between the organizational structure determination based on the learning model generated by the organizational structure learning device 1 and the organizational configuration determination according to the comparative example.

In the reference example (judgment criteria), the operator visually identified and determined the mineral phase, composite structure, and pores contained in all 300 cross-sectional images for evaluation. In the reference example, all 300 cross-sectional images for evaluation are visually identified by the operator, so the determination accuracy is 100%, but the burden of determination becomes extremely large, which is not preferable.

Note that the determination result by the operator in the reference example becomes the determination standard when determining the determination accuracy of the comparative example and the example. When the determination results according to the comparison example and the example for the tissue included in the target cross-sectional image completely match the determination results by the operator, the determination result according to the comparison example and the example for the target cross-sectional image is determined to be correct. Ta. On the other hand, when the judgment results according to the comparison example and the example for the tissue included in the target cross-sectional image differ even in part from the judgment result by the operator, the judgment result according to the comparison example and the example for the target cross-sectional image differs even in part. determined to be incorrect.

A comparative example is a determination based on brightness analysis described in Patent Document 2, in which a threshold value is set in a brightness histogram to determine whether a region of a cross-sectional image corresponds to either a mineral phase or a composite structure. In brightness analysis, it is not easy to accurately determine tissues with similar brightness distributions, and the determination accuracy in the comparative example was 0%.

In the example, it was determined whether a region of the cross-sectional image corresponds to either a mineral phase or a composite structure using a machine learning learning model using the cross-sectional image shown in FIG. 7 as training data. The determination accuracy of the determination by the learning model in the example was 91%.

In the example, the learning model separates acicular calcium ferrite into those in which it exists alone and those in which it coexists with slag to form a composite structure. Table 3 shows the results of quantifying the respective composition ratios.

Further, in the embodiment, since the learning model determines the tissue included in the cross-sectional image, the operator can obtain information regarding the structure of the sintered ore with a minimum burden.

1 Tissue configuration learning device 20 Processing unit 21 Image data acquisition unit 22 Image data correction unit 23 Learning data generation unit 24 Learning processing unit 25 Learning model output unit 30 Learning model 100 Tissue configuration learning system 101 Optical microscope 102 Imaging device

Claims (7)

an image data acquisition unit that acquires a plurality of image data each showing a plurality of cross-sectional images taken of a cross-section of the sintered ore;
a learning data generation unit that generates a plurality of learning data by associating each of the plurality of regions of the cross-sectional image with a structure of a composite structure formed by a mineral phase contained in the sintered ore;
a learning processing unit that causes a learning model to learn a relationship between the plurality of cross-sectional images and the tissue configuration of the composite tissue using the plurality of learning data;
A learning model output that outputs a learning model that estimates the tissue structure of the composite tissue from a cross-sectional image of a cross section of sintered ore based on the learned relationship between the plurality of cross-sectional images and the structure of the composite tissue. Department and
A device for learning the structure of sintered ore, comprising: The composite structure includes a first composite structure in which either hematite or magnetite is surrounded by calcium ferrite, a second complex structure in which calcium ferrite and silicate slag are combined, and either hematite or magnetite. Learning the structure of the sintered ore according to claim 1, which includes a third complex structure that is a complex structure of and calcium ferrite, and a fourth complex structure that is a complex structure of either hematite or magnetite and silicate slag. Device. The plurality of learning data are generated by further associating the plurality of regions with a first mineral phase that is an iron oxide phase of hematite or magnetite, and a second mineral phase that is a primary phase of calcium ferrite. 2. The sintered ore structure learning device according to item 2. The first mineral phase includes a tenth mineral phase that is a residual source ore, an eleventh mineral phase that is primary hematite or primary magnetite, and a twelfth mineral phase that is secondary hematite or secondary magnetite. A device for learning the structure of sintered ore. The second mineral phase includes a 2a mineral phase that is acicular calcium ferrite, and a 2b mineral phase that is columnar calcium ferrite,
3. The second composite structure includes a 2a composite structure that is a composite structure of acicular calcium ferrite and a silicate slag, and a 2b composite structure that is a composite structure of columnar calcium ferrite and a silicate slag. Or the sintered ore structure learning device according to 4. Obtain multiple image data each showing multiple cross-sectional images of the cross-section of the sintered ore,
a learning data generation unit that generates a plurality of learning data by associating each of the plurality of regions of the cross-sectional image with a structure of a composite structure formed by a mineral phase contained in the sintered ore;
causing a learning model to learn the relationship between the plurality of cross-sectional images and the tissue configuration of the composite tissue using the plurality of learning data;
outputting a learning model that estimates the tissue structure of the composite tissue from a cross-sectional image obtained by capturing a cross section of sintered ore based on the learned relationship between the plurality of cross-sectional images and the structure of the composite tissue;
A sintered ore structure learning program characterized by causing a computer to execute processing. Obtain multiple image data each showing multiple cross-sectional images of the cross-section of the sintered ore,
a learning data generation unit that generates a plurality of learning data by associating each of the plurality of regions of the cross-sectional image with a structure of a composite structure formed by a mineral phase contained in the sintered ore;
causing a learning model to learn the relationship between the plurality of cross-sectional images and the tissue configuration of the composite tissue using the plurality of learning data;
outputting a learning model that estimates the tissue structure of the composite tissue from a cross-sectional image obtained by capturing a cross section of sintered ore based on the learned relationship between the plurality of cross-sectional images and the structure of the composite tissue;
A method for learning the structure of sintered ore, comprising:

Images