Information processing program, information processing apparatus, data structure, and information processing method

Abstract

To provide an information processing program capable of classifying the deformation characteristics of coke for each CSR value while suppressing the computational cost.SOLUTION: The information processing program includes the steps of: concerning three-dimensional model of coke, assigning multiple types of labels to a three-dimensional grid depending on the hardness distribution within the substance of the coke; receiving an input 3D model data in multiple batches with multiple types of labels as the number of batches of the 3D model and the 4D data 23 consisting of the 3D elements of the 3D model; performing convolution by multiplying filter elements and adding a bias on the 4D data 23; and acquiring one-dimensional fully connected data as output data 24.SELECTED DRAWING: Figure 6

Description

Technical Field

[0001] The present invention relates to an information processing program, an information processing apparatus, a data structure, and an information processing method.

Background Art

[0002] A blast furnace is a countercurrent packed bed reactor in which ore and coke are alternately charged and packed in layers. Coke serves as a spacer that ensures air permeability while supporting the load of the charged raw materials in the packed bed.

[0003] In recent years, for the purpose of reducing CO2 emissions, an operation with a high ratio of charged ore to charged coke in the blast furnace, that is, an operation with a low coke ratio, has been demanded. Thermodynamically, CO2 can be reduced by the low coke ratio operation, but the thinning of the coke layer has a great impact on the air permeability in the blast furnace. When the progress of ore reduction stagnates due to a decrease in the air permeability of the ore layer, the reaction load on the coke increases, and the coke is easily worn. When the coke with reduced strength is pulverized, narrowing and blockage of the voids in the blast furnace occur, and the air permeability in the blast furnace deteriorates extremely. Therefore, in order to perform stable operation in a blast furnace with a low coke ratio, it is necessary to accurately know the quality and behavior of the coke.

[0004] As an index for evaluating the quality of coke, the gasification amount (Coke Reaction Index; CRI) obtained by reacting coke with CO2 gas for a certain period of time, and the strength after reaction (Coke Strength after Reaction; CSR) indicating the subsequent rotational strength are used.

[0005] For example, in Patent Document 1, the reaction rate (CRI) of the blended coke is obtained based on the weighted average value of the CRI of the single-flavor carbon coke, and the obtained CRI of the blended coke is corrected based on the operating conditions consisting of the coke arrival temperature, the porosity of the coke, and the furnace width of the coke oven. The CSR of the blended coke is estimated based on the corrected CRI of the blended coke and the surface fracture strength DI 150 6.

Prior Art Documents

Patent Document

[0006]

Patent Document 1

[0007] As a method for knowing the behavior of coke, a movement simulation of coke and powder using the Discrete Element Method (DEM), which is a calculation method of solid particle movement, has been attempted. In this method, based on a kinetic model that reflects the change in coke shape due to in-furnace reaction and load, the flow phenomena of solid, gas, liquid, and powder are visually evaluated, and the clogging of the voids in the blast furnace is predicted. This method can consider powder and void geometry and can uniformly handle ventilation, powder passage, liquid passage, and moving bed by coupling with the Computational Fluid Dynamics (CFD) method. Therefore, it is expected as a new evaluation index for the relationship between raw material properties (coke pulverization), air permeability, and liquid permeability.

Summary of the Invention

Problems to be Solved by the Invention

[0008] According to the estimation method of the above Patent Document 1, since calculations are performed based on a plurality of information including operating conditions, there is a problem that it is difficult to suppress the calculation cost.

[0009] Also, in the above-described movement simulation, although the calculation load is large, the range that can be evaluated considering the actual packed bed structure up to the coke shape is limited to a cubic space of about 1 m 3 and there is a problem that it is difficult to analyze the deformation behavior that occurs probabilistically in a scale hierarchical manner.

[0010] The present invention has been devised in view of such problems, and one of its purposes is to classify the deformation characteristics of coke for each CSR value while suppressing the calculation cost.

Means for Solving the Problems

[0011] The information processing program disclosed herein causes a computer to perform the following processing on a three-dimensional model of coke: assign a plurality of types of labels to a three-dimensional lattice according to the hardness distribution within the substance of the coke, accept as input the data of a plurality of batches of the three-dimensional model to which the plurality of types of labels have been assigned, as four-dimensional data consisting of the number of batches of the three-dimensional model and the three-dimensional elements of the three-dimensional model, perform a convolution process on the four-dimensional data by multiplying by a filter element and adding a bias, and obtain one-dimensional fully-connected data as output data.

Advantages of the Invention

[0012] According to the present invention, it is possible to classify the deformation characteristics of coke for each CSR value while suppressing the calculation cost.

Brief Description of the Drawings

[0013]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Mode for Carrying Out the Invention

[0014] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. However, the embodiment described below is merely an example, and there is no intention of excluding various modifications or applications of technologies not explicitly described below. That is, the present embodiment can be variously modified and implemented without departing from its gist. In addition, each drawing does not mean that it includes only the components shown in the drawing, and can include other functions. In the drawings used in the following description, parts denoted by the same reference numerals represent the same or similar parts unless otherwise specified.

[0015] 〔A〕One Embodiment The information processing apparatus 1 shown in FIG. 1 is an example of an information processing apparatus that generates four-dimensional data from a three-dimensional model of coke, performs deep learning on the four-dimensional data by a neural network, and acquires output data and a classification model. Further, the information processing apparatus 1 is an example of an information processing apparatus that classifies new data with the output data using the classification model.

[0016] 〔A-1〕Example of Hardware Configuration of Information Processing Apparatus FIG. 1 is a block diagram schematically showing an example of the hardware (HW) configuration of the information processing apparatus (computer) 1 of the present embodiment. As an HW configuration, the computer 1 may exemplarily include a processor 1a, a memory 1b, a storage unit 1c, an IF (Interface) unit 1d, an IO (Input / Output) unit 1e, and a reading unit 1f.

[0017] Processor 1a is an example of an arithmetic processing unit that performs various controls and operations. Processor 1a may be communicably connected to each block in computer 1 via bus 1i. Examples of processor 1a include integrated circuits (ICs) such as a Central Processing Unit (CPU) and a Graphics Processing Unit (GPU).

[0018] Memory 1b is an example of HW that stores various data and information such as programs. Examples of memory 1b include one or both of a volatile memory such as a DRAM (Dynamic Random Access Memory) and a non-volatile memory such as a PM (Persistent Memory).

[0019] Storage unit 1c is an example of HW that stores various data and information such as programs. Examples of storage unit 1c include various storage devices such as a magnetic disk device such as an HDD (Hard Disk Drive), a semiconductor drive device such as an SSD (Solid State Drive), and a non-volatile memory. Examples of non-volatile memory include, for example, flash memory, SCM (Storage Class Memory), ROM (Read Only Memory), and the like.

[0020] Further, storage unit 1c may store a program (information processing program) 1g that realizes all or part of various functions of computer 1.

[0021] For example, the processor 1a of the information processing apparatus 1 can realize the functions of the information processing apparatus 1 (control unit 10) shown in FIG. 2 described later by expanding and executing the program 1g stored in the storage unit 1c in the memory 1b. Further, the memory unit 20 illustrated in FIG. 2 described later may be realized by a storage area included in at least one of the memory 1b and the storage unit 1c. Furthermore, the acquisition unit 11, generation unit 12, machine learning unit 13, and classification unit 14 illustrated in FIG. 2 described later may store the acquired (generated) information in at least one of the memory 1b and the storage unit 1c as an example of a storage device.

[0022] The IF unit 1d is an example of a communication IF that controls connections and communications with a network. For example, the IF unit 1d may include an adapter compliant with a LAN (Local Area Network) such as Ethernet (registered trademark), or an optical communication such as FC (Fibre Channel). For example, the information processing apparatus 1 may be communicably connected to a measurement device (3D scanner) (not shown) via the IF unit 1d. Also, for example, the program 1g may be downloaded from a network to the computer 1 via the communication IF and stored in the storage unit 1c.

[0023] The IO unit 1e may include one or both of an input device and an output device. Examples of the input device include a keyboard, a mouse, a touch panel, and the like. Examples of the output device include a monitor, a projector, a printer, and the like. For example, for the acquisition unit 11, generation unit 12, machine learning unit 13, and classification unit 14 shown in FIG. 2 described later, information and instructions may be input from the input device of the IO unit 1e, and the acquired (generated) information may be output and displayed on the output device.

[0024] The reading unit 1f is an example of a reader that reads information (data) and program information recorded on the recording medium 1h. The reading unit 1f may include a connection terminal or device to which the recording medium 1h can be connected or inserted. Examples of the reading unit 1f include an adapter compliant with USB (Universal Serial Bus) or the like, a drive device that accesses a recording disk, a card reader that accesses a flash memory such as an SD card, and the like. Note that a program 1g may be stored in the recording medium 1h, and the reading unit 1f may read the program 1g from the recording medium 1h and store it in the storage unit 1c.

[0025] Examples of the recording medium 1h include non-temporary computer-readable recording media such as magnetic / optical disks and flash memories. Examples of the magnetic / optical disk include a flexible disk, a CD (Compact Disc), a DVD (Digital Versatile Disc), a Blu-ray Disc, an HVD (Holographic Versatile Disc), and the like. Examples of the flash memory include semiconductor memories such as a USB memory and an SD card.

[0026] The HW configuration of the information processing apparatus 1 described above is an example. Therefore, an increase or decrease in HW (for example, addition or deletion of an arbitrary block), division, integration in an arbitrary combination, or addition or deletion of a bus, etc. within the information processing apparatus 1 may be appropriately performed. For example, in the information processing apparatus 1, at least one of the IO unit 1e and the reading unit 1f may be omitted.

[0027] 〔A-2〕Example of functional configuration of information processing apparatus FIG. 2 is a block diagram schematically showing an example of the functional configuration of the information processing apparatus 1 of the present embodiment. As shown in FIG. 2, the information processing apparatus 1 includes a control unit 10, a memory unit 20, an input unit 30, and an output unit 40. The control unit 10, the memory unit 20, the input unit 30, and the output unit 40 may be connected to be communicable with each other.

[0028] The input unit 30 is realized by the IF unit 1d, the IO unit 1e, and the reading unit 1f in FIG. 1, and receives the input of information and instructions necessary for the processing performed by the control unit 10. Further, the output unit 40 is realized by the IO unit 1e in FIG. 1, and outputs the information obtained by the processing performed by the control unit 10.

[0029] The control unit 10 is the processor 1a shown in FIG. 1, and by executing the Operating system (OS) and the information processing program 1g stored in the storage unit 1c, it functions as the acquisition unit 11, the generation unit 12, the machine learning unit 13, and the classification unit 14.

[0030] The memory unit 20 is realized by the storage area of at least one of the memory 1b and the storage unit 1c shown in FIG. 2, and stores various data. As shown in FIG. 2, the memory unit 20 stores, for example, the three-dimensional model 21, the teacher data 22, the four-dimensional data 23, the output data 24, and the classification model 25.

[0031] The acquisition unit 11 acquires the three-dimensional model 21 and the teacher data 22 input via the input unit 30. The acquired three-dimensional model 21 and teacher data 22 may be stored in the memory unit 20.

[0032] The three-dimensional model 21 is the shape of a substance created in a three-dimensional space. The three-dimensional model 21 is created using a 3D scanner, and examples of the file format include Standard Triangulated Language (STL). The three-dimensional model 21 is, for example, polygon data surrounded by a mesh composed of three points connecting the vertices of a triangle, where 300,000 points are selected as representative values from the point cloud of the surface coordinates of the scanned substance.

[0033] In this embodiment, the three-dimensional model 21 is data obtained by scanning the process of deformation of one unit of substance in a plurality of stages. In other words, the three-dimensional model 21 is time-series data (a collection) indicating the shape change of the substance.

[0034] The substance in this embodiment is coke. Coke is coal that has been carbonized at high temperature (through a coking process), from which components such as sulfur, coal tar, pitch, sulfuric acid, and ammonia have been removed. Coke has, for example, the surface shape shown in Fig. 3(a).

[0035] In this embodiment, a three-dimensional model 21 is obtained by scanning the gradual shape changes of coke during the process of examining the Coke Strength after Reaction (CSR) value of coke. Hereinafter, the method for obtaining the three-dimensional model 21 of the coke in this embodiment will be described.

[0036] In this embodiment, metallurgical coke with a certain strength was used as the sample coke. The initial particle size of the sample coke was obtained by screening approximately 10 kg of coke through a sieve with apertures of 25 - 38 mm, and the one with an average particle size of 35 mm was selected. This is because the particle size of coke in the lower part of the blast furnace usually has about 50% of particles with a size of 25 mm or more by mass ratio.

[0037] (1.1) First, the sample coke is gasified. In this embodiment, the sample coke was gasified under the conditions of a CO2 atmosphere (100%), 1100 °C, and 40 min.

[0038] (1.2) Next, a plurality of cokes are taken out from the sample coke after gasification treatment, and these are used as representative cokes with a rotation speed of 0 [rev]. As shown in Fig. 3(b), a through-hole with a diameter of φ1 mm was provided in the representative coke, and a thread was passed through it and tied to the coke to serve as a marker. This is for identifying the representative coke from among a plurality of sample cokes in (1.3) described later. The representative coke was scanned with a 3D scanner to obtain the three-dimensional model 21. The three-dimensional model 21 at this point is data of the representative coke before deformation (initial state). In this embodiment, a NextEngine 3D scanner was used.

[0039] (1.3) Subsequently, a sample coke containing representative coke is subjected to a predetermined number of rotational impacts. As a rotational strength test method for applying the rotational impact, in this embodiment, the tumbler test method of JIS K2151-1977 is adopted. The tumbler testing machine used in the tumbler test method is suitable for tracking the subtle shape changes of coke because it has a lower drop impact than a drum testing machine and can suppress volume destruction. The tumbler testing machine is filled with a sample coke containing representative coke and rotated to a predetermined number. In this embodiment, the predetermined numbers are set to 50, 100, 200, and 400 [rev]. Each time the rotation reaches the predetermined number, the tumbler testing machine is stopped, the representative coke is taken out and scanned with a 3D scanner to obtain a three-dimensional model 21. Fig. 3(c) shows the deformation states of two cokes subjected to rotational impacts.

[0040] (1.4) The above procedure is performed for all representative cokes, and for each representative coke, three-dimensional models 21 of the coke before deformation and after four stages of deformation are obtained. In this embodiment, a three-dimensional model 21 is obtained with five (0, 50, 100, 200, 400 [rev]) scan data per representative coke particle as a set.

[0041] (1.5) By the above tumbler test, the CSR value of each representative coke can be obtained. The CSR value corresponding to the three-dimensional model 21 of each representative coke is used as teacher data 22.

[0042] Here, the CSR (Coke Strength after Reaction) value is the above-mentioned strength after reaction and can evaluate the quality, particularly the hardness, of coke. The CSR value can be calculated from the weight of the coke in the initial state and the weight of the coke after the tumbler test, and represents the ratio finally remaining without pulverization from the initial state in weight %. For example, when the CSR value is 61.9, it indicates that the weight of the coke after the tumbler test is 61.9% of the weight of the coke in the initial state. Therefore, the higher the CSR value, the higher the hardness of the coke.

[0043] Figure 4 exemplifies representative cokes having a high CSR value (61.8) and a low CSR value (38.5). The destruction of the coke occurs due to volume destruction, which is a destruction that occurs volumetrically so as to largely fragment the particles in the initial state, and surface destruction, which is a destruction in which the surface of the particles in the initial state is compressed to generate fine powder. As shown in Figure 4, since the coke with a low CSR value is likely to undergo volume destruction in addition to surface destruction, the destruction progresses more easily compared to the coke with a high CSR value.

[0044] Based on the three-dimensional model 21 acquired by the acquisition unit 11, the generation unit 12 generates four-dimensional data 23. The generated four-dimensional data 23 may be stored in the memory unit 20.

[0045] The four-dimensional data 23 is data to be processed by the machine learning unit 13 and the classification unit 14. The four dimensions mean that it consists of the number of batches of the three-dimensional model 21 and the three-dimensional elements of the three-dimensional model 21. Here, the number of batches is the number of cokes selected for learning, and the three-dimensional elements of the three-dimensional model 21 are the height, width, and depth of the three-dimensional lattice in which the three-dimensional model 21 is arranged. In other words, the four-dimensional data 23 is data consisting of four-dimensional elements of (number of batches, depth, height, width).

[0046] The four-dimensional data 23 is data of a plurality of batches of the three-dimensional model 21 in which a plurality of types of labels are respectively assigned to the three-dimensional lattice according to the hardness distribution in the substance for the three-dimensional model 21. Specifically, the four-dimensional data 23 is data in which pulverization information of the substance is embedded at coordinates corresponding to each part on the three-dimensional lattice virtualized in the substance for each stage of shape change. The hardness distribution is determined by the shape change of the coke in the rotary tumbler. In other words, the hardness distribution is what appears as a difference in hardness for each part in the substance of the coke as a change in shape by applying rotational impact in the tumbler test.

[0047] FIG. 5(a) is a cross-sectional view of the coke during the deformation process. Ω indicates the pulverization region, and as the rotation speed increases, the pulverization region advances toward the inside of the coke. When the surface shape of a substance is regarded as the boundary between a predetermined three-dimensional lattice and the surface of the substance arranged within the lattice, the deformation process of the substance can be considered to be digitized on the surface coordinates of the substance as the moving distance of the detected boundary. Hereinafter, a method for generating the four-dimensional data 23 in the present embodiment will be described.

[0048] In the present embodiment, the four-dimensional data 23 is generated using the three-dimensional model 21 of the representative coke corresponding to any one of four types of CSR values (A: 61.9, B: 59.2, C: 40.5, D: 38.5).

[0049] (2.1) First, the generation unit 12 calculates the displacement of the surface coordinates of the coke according to the increase in the predetermined rotation speed. The generation unit 12 arranges the entire three-dimensional model 21 in a cubic space (cubic region) with a length L, and divides the cubic space into blocks of a square lattice with an interval dx. By dividing into blocks, the spatial characteristics of the three-dimensional shape can be easily captured. In the present embodiment, an interval dx of 64 is adopted, and thus the square lattice has a size of 64×64×64. Further, the generation unit 12 calculates the distance from the center coordinates of each lattice of the square lattice (block) with a side length of dx to the triangular mesh surface of the coke. By using a Level set function (a function that is positive inside the solid phase and negative outside) for the calculation, the solid shape can be digitized by arranging only positive particles. In the present embodiment, Insight (registered trademark) Meshman_ParticleGen_HPC Ver. 2.0.1 was used for the above data processing.

[0050] (2.2) Next, the generation unit 12 assigns a plurality of types of labels on a three-dimensional lattice according to the hardness distribution of the coke. Specifically, the generation unit 12 assigns a label indicating the hardness distribution to each cubic lattice that constitutes the voxel data of the cubic space in which the three-dimensional model 21 obtained from the above (2.1) is arranged. Here, voxel data is a representation of a three-dimensional shape as a set of cubes. With these voxels, it becomes easy to specify the coordinates of the surface shape part of the substance. Note that the data may be binary, text, or described in other formats.

[0051] Furthermore, the process in which the generation unit 12 embeds the quantified hardness information into each voxel will be described with reference to FIG. 5(b). FIG. 5(b) is a diagram for explaining the process of assigning labels on a three-dimensional lattice in the information processing apparatus 1 according to the embodiment in a two-dimensional table. The table shown in FIG. 5(b) numerically represents the pulverization information of the part of the surface shape of the coke, here a certain region surrounded by a square. Although this process is a three-dimensional process, for convenience of explanation, in FIG. 5, the voxels are replaced with pixels and represented as a two-dimensional process. In the example of FIG. 5, the height and width of the table respectively correspond to the height and depth of the block, and the height and width of one cell respectively correspond to the height and depth of the voxel.

[0052] The numbers 0 to 5 entered in the cells of the table are labels indicating the presence (pulverization) of the coke and are assigned to each part (voxel) on the three-dimensional lattice. The part where the coke does not exist is indicated by "0", and the part that remained without being pulverized due to the increase in the number of rotations in the tumbler test is indicated by the numbers 1 to 5 according to the increase in the number of rotations. That is, the numbers 1 to 5 correspond to the level of hardness. Furthermore, in FIG. 5(b), the difference in hardness is shown by the shade of color together with the numbers.

[0053] The boundary between the region marked with "0" and the region marked with "1" indicates the surface shape of the three-dimensional model 21 obtained at the rotation speed of 0. Similarly, the boundary between the region marked with "1" and the region marked with "2" indicates the surface shape of the three-dimensional model 21 obtained after receiving the rotational impact at the rotation speed of 50, and the boundary between the region marked with "2" and the region marked with "3" indicates the surface shape of the three-dimensional model 21 obtained after receiving the rotational impact at the rotation speed of 100. Further, the boundary between the region marked with "3" and the region marked with "4" indicates the surface shape of the three-dimensional model 21 obtained after receiving the rotational impact at the rotation speed of 200, and the boundary between the region marked with "4" and the region marked with "5" indicates the surface shape of the three-dimensional model 21 obtained after receiving the rotational impact at the rotation speed of 400.

[0054] The region of "1" is the part that remained without being pulverized by the rotational impact at the rotation speed of 0. Similarly, the region of "2" is the part that remained without being pulverized by the rotational impact at the rotation speed of 50, and the region of "3" is the part that remained without being pulverized by the rotational impact at the rotation speed of 100. Further, the region of "4" is the part that remained without being pulverized by the rotational impact at the rotation speed of 200, and the region of "5" is the part that remained without being pulverized by the rotational impact at the rotation speed of 400.

[0055] In this way, the sites on the three-dimensional lattice are information indicating the deformation process of the coke per particle itself. In other words, the sites on the three-dimensional lattice are information indicating the hardness distribution within the coke material.

[0056] (2.3) The generation unit 12 randomly selects a plurality of data of the three-dimensional model 21 with labels assigned on the three-dimensional lattice obtained by the above procedure, and sets this as one batch. This one batch of data is generated as four-dimensional data 23. FIG. 6 shows an example of deep learning by a neural network in the information processing apparatus 1 according to the embodiment. In this embodiment, 100 pieces of data are randomly selected from 8192 pieces of four-dimensional data 23, and this is set as one batch. The four-dimensional data 23 of this embodiment is four-dimensional data with (number of batches, depth, height, width) = (100, 64, 64, 64). Note that the image of the coke shown at the left end of FIG. 6 is an image.

[0057] Since the coke is isotropic, by rotating the three-dimensional model 21 in the three-dimensional lattice and performing the above processes (2.1) to (2.3) on the rotated three-dimensional model 21, a plurality of four-dimensional data 23 can be generated. Therefore, if there is a set of three-dimensional models 21 for each representative coke, four-dimensional data 23 of any number of patterns can be obtained. In the present embodiment, eight patterns of four-dimensional data 23 were generated from a set of three-dimensional models 21 for each representative coke. Since the above-described four-dimensional data 23 is machine-learned by the machine learning unit 13, it is also referred to as a data structure for machine learning.

[0058] When an instruction for machine learning of the four-dimensional data 23 is input via the input unit 30, the machine learning unit 13 inputs the four-dimensional data 23 generated by the generation unit 12 into a neural network to obtain output data 24. The machine learning unit 13 updates the weights and biases so that the error of the output data 24 becomes small.

[0059] When teacher data 22 is added to the four-dimensional data 23, the machine learning unit 13 updates the weights and biases so that the error becomes small based on the error between the output data 24 of the neural network and the teacher data 22.

[0060] In the present embodiment, the four-dimensional data 23 to which the teacher data 22 is given is learned. The four-dimensional data 23 was learned for 20 epochs with one batch of learning composed of 100 randomly selected data as one epoch.

[0061] The neural network used here has, for example, three layers: an input layer, an intermediate layer, and an output layer as shown in FIG. 6. It receives the input of the four-dimensional data 23 in the input layer, performs a convolution process on the four-dimensional data 23 in the intermediate layer, and outputs one-dimensional fully connected data as the output data 24 in the output layer.

[0062] In the input layer, the four-dimensional data 23 is converted into a matrix. In this embodiment, the im2col (image to column) function is used. This enables the data input to be arranged conveniently for the filters in the intermediate layer. That is, the three-dimensional grid can be unfolded into a single row to execute matrix calculations more rapidly.

[0063] In the intermediate layer, the filter elements are multiplied by the four-dimensional data 23 and a bias is added. In the intermediate layer, the characteristics of the deformation process of the coke contained in the four-dimensional data 23 are learned step by step. The intermediate layer of this embodiment is composed of any combination of filter elements, namely, a convolutional layer (T), an activation function (K), a pooling layer (P), and an affine layer (A). In this embodiment, in the convolutional layer, feature extraction is performed with a depth of 6 layers, and the ReLU function is used as the activation function. Further, the data is compressed in the pooling layer, and a fully connected operation of the data is performed in the affine layer. #1 to #6 in FIG. 6 indicate the 1st to 6th layers of each element, and the alphabets T, K, P, and A indicate each layer. For example, #1-T indicates the first-layer convolutional layer.

[0064] In the output layer, the one-dimensional fully connected data is output as output data 24. In this embodiment, the output layer outputs the probabilities of four types of CSR values by means of the Softmax function through a fully connected layer having 50 hidden layer neurons. The machine learning unit 13 acquires the probabilities of the four types of output CSR values as output data 24.

[0065] The sizes of the filter elements (batch (number of batches), channel (depth), height, width) of each layer in the intermediate layer can be arbitrarily set. In this embodiment, the sizes of the filter elements are set such that they gradually decrease from #1-T at the left end to #1-A at the right end of the intermediate layer.

[0066] The machine learning unit 13 compares the output data 24 with the teacher data 22 and updates the weight and bias parameters. In this embodiment, after calculating the loss function E, the error gradient is obtained by the backpropagation method and optimized by Adaptive Movement Estimation (Adam), thereby updating the weight and bias parameters. The learning rate is set to 0.001, and the Dropout algorithm is adopted to avoid overfitting.

[0067] When the machine learning unit 13 executes machine learning a predetermined number of times or when the error becomes smaller than a predetermined value, the machine learning may be terminated, and various parameters such as weights and biases may be stored in the memory unit 20 as the classification model 25. Further, the machine learning unit 13 may display the output data 24 via the output unit 40 on an output device such as a monitor, for example.

[0068] After the machine learning of the four-dimensional data 23 by the machine learning unit 13, the classification unit 14 outputs the classification result by the learned classification model 25 with data regarding a new substance as input. In this embodiment, the classification unit 14 classifies coke according to the CSR value according to the hardness distribution of the coke using the classification model 25 generated based on the output data 24. In other words, the classification unit 14 performs classification of substances according to the hardness distribution using the output data 24. The classification result is an evaluation prediction value of the substance and may be indicated by a numerical value or other expression. The classification result of this embodiment is the CSR prediction value and is indicated by, for example, a probability.

[0069] For example, when the acquisition unit 11 receives a new three-dimensional model 21 (to be classified) via the input unit 30, it may be stored in the memory unit 3 as the three-dimensional model 21 to be classified. The generation unit 42 may generate the four-dimensional data 23 to be classified using the three-dimensional model 21 to be classified by the above-described methods (1.1) to (1.4) and (2.1) to (2.3) and store it in the memory unit 20.

[0070] The classification unit 14 refers to the memory unit 20 and classifies the generated four-dimensional data 23 to be classified using the learned classification model 25. For example, the classification unit 14 may construct a neural network with various parameters such as the weights and biases of the learned classification model 25, input the four-dimensional data 23 to be classified into the neural network, and obtain a classification result. Further, the obtained classification result may be displayed (output) as output data 24 to an output device such as a monitor via the output unit 40.

[0071] 〔A-3〕Operation of the information processing apparatus FIG. 7 is an example of a flowchart of the processing procedure in the training stage according to the information processing method of the embodiment.

[0072] In the control unit 10 of the information processing apparatus 1, the acquisition unit 11 acquires the three-dimensional model 21 via the input unit 30 (step S1) and stores it in the memory unit 20. Further, the acquisition unit 11 may acquire the teacher data 22 associated with the three-dimensional model 21. Note that when the acquisition unit 11 stores the three-dimensional model 21 in the memory unit 20, it may output a generation instruction to the generation unit 12.

[0073] When a generation instruction is input from, for example, the acquisition unit 11 or the input unit 30, the generation unit 12 generates four-dimensional data 23 based on the three-dimensional model 21 (step S2) and stores the generated four-dimensional data 23 in the memory unit 3. The four-dimensional data 23 may be generated using the methods described in (2.1) to (2.3) above.

[0074] When the machine learning unit 13 receives a machine learning instruction from the generation unit 12, for example, it refers to the memory unit 3 and performs machine learning on the four-dimensional data 23 (step S3). The machine learning may be performed by the neural network described above. The machine learning unit 13 ends the machine learning when, for example, machine learning has been executed a predetermined number of times, or when the error between the output data 24 of the neural network and the teacher data 22 becomes smaller than a predetermined value. Then, the machine learning unit 13 stores various parameters such as weights and biases in the memory unit 20 as the classification model 25 (step S4), and the process ends. Note that the output data 24 may be output via the output unit 40 after step S3.

[0075] FIG. 8 is an example of a flowchart of the processing procedure of the classification stage according to the information processing method of the embodiment.

[0076] In the control unit 10 of the information processing apparatus 1, the acquisition unit 11 acquires a new (classification target) three-dimensional model 21 via the input unit 30 (step S11) and stores it in the memory unit 3. Note that when the acquisition unit 11 stores the three-dimensional model 21 in the memory unit 3, it may output a generation instruction to the generation unit 12.

[0077] When the generation unit 12 receives a generation instruction from the acquisition unit 11, for example, it generates four-dimensional data 23 based on the three-dimensional model 21 (step S12) and stores the generated four-dimensional data 23 in the memory unit 20. The four-dimensional data 23 may be generated using the methods described in (2.1) to (2.3) above.

[0078] When the classification unit 14 receives a classification instruction from the generation unit 12, for example, it refers to the memory unit 20 and classifies the four-dimensional data 23 to be classified using the classification model 25 generated based on the output data 24 (step S13). The classification unit 14 outputs the classification result of the classification model 25 via the output unit 40 (step S14), and the process ends.

[0079] 〔A-4〕Effect (1) The acquisition unit 11 acquires a three-dimensional model 21 of coke (substance), and the generation unit 12 assigns a plurality of types of labels to a three-dimensional lattice according to the hardness distribution within the substance of the coke for the three-dimensional model 21, and generates four-dimensional data 23 composed of the number of batches of the three-dimensional model and the three-dimensional elements of the three-dimensional model from the data of the plurality of batches of the three-dimensional model to which the plurality of types of labels are assigned. Further, the machine learning unit 13 receives the input of the four-dimensional data 23, performs a convolution process on the four-dimensional data 23 by multiplying by filter elements and adding a bias, and acquires one-dimensional fully connected data as output data 24.

[0080] Thus, according to the information processing program 1g, the information processing apparatus 1, and the information processing method of the present embodiment, powdering information of the substance can be embedded in coordinates corresponding to the surface shape part of the substance in the three-dimensional lattice in which the substance is arranged for each stage of shape change. Thereby, since data of a large capacity such as the hardness distribution of the shape of one particle of coke can be obtained as three-dimensional (four-dimensional including the number of batches) data, an increase in dimension can be prevented, and the calculation cost in machine learning can be suppressed. Further, the deformation process of coke for each CSR value can be learned.

[0081] (2) The classification unit 14 uses the output data 24 to perform classification of coke according to the hardness distribution.

[0082] Parameters such as the weights and biases of the neural network can be updated using the output data 24, and the deformation characteristics of coke for each CSR can be classified using the obtained classification model 25.

[0083] (3) The hardness distribution of the three-dimensional model 21 is determined by the shape change of coke in the rotary tumbler.

[0084] As a result, a three-dimensional model 21, which is time-series information of the shape of a single particle of coke that continuously deforms moment by moment under an external force, can be obtained. Further, teacher data 22, which is the CSR value corresponding to each three-dimensional model 21, can be obtained.

[0085] (4) The generation unit 12 assigns a plurality of types of labels to each site on the lattice that remains without pulverization due to an increase in the rotation speed of the rotary tumbler.

[0086] As a result, the location where the shape of a single particle of coke has changed can be identified, and the progress of pulverization can be confirmed, so that the deformation process of the coke can be accurately grasped.

[0087] (5) The input unit 30 inputs information and instructions necessary for the processing of the control unit 10 to the control unit 10, and the output unit 40 outputs the information obtained by the processing of the control unit 10.

[0088] As a result, the information processing program 1g, the information processing apparatus 1, and the information processing method of the present embodiment can obtain information necessary for processing (three-dimensional data of coke (substance)) from the outside and show the processing results (probability of the CSR value of coke (substance), CSR prediction value (evaluation prediction value)) to the outside.

[0089] (6) Further, the four-dimensional data 23 generated from the three-dimensional model 21 of coke described above is a data structure for machine learning that is processed by a computer.

[0090] Since the data structure is used for a computer to perform machine learning processing, it conforms to a program and has the same effects as those described in the information processing program 1g above.

[0091] With the above configuration, it was possible to classify the deformation characteristics of coke for each CSR value. Therefore, it is expected that the above classification model 25 will be further developed to inversely calculate the deformation process of coke by inputting the CSR value. If the deformation process of coke calculated from the CSR value is incorporated into the motion simulation using the above kinetic model, etc., the calculation of the mechanical response of the deformation within the coke material can be omitted, so that the calculation cost can be significantly reduced when analyzing the behavior of coke in the entire blast furnace.

[0092] 〔A-5〕Evaluation Figure 9 is a graph comparing the classification accuracies when the above information processing method is learned by changing the number of layers. The number of layers of the neural network in the machine learning unit 13 was set to 1 layer or 6 layers. In Figure 9, the training result (train) obtained by performing machine learning using four-dimensional data to be trained is shown by a broken line, and the test result (test) obtained by classifying the four-dimensional data to be classified with the classification model is shown by a solid line.

[0093] Looking at the training result (train), a correct answer rate of 99% or more was obtained for the four-dimensional data to be trained. Also, looking at the test result (test), a correct answer rate of 97% or more was obtained for the four-dimensional data to be classified. From this, it was confirmed that according to the information processing method according to the present embodiment, the deformation of coke for each CSR can be accurately classified.

[0094] Focusing on the difference in the number of layers, the increase in the correct answer rate with the increase in epochs is faster for 1 layer than for 6 layers. That is, 1 layer can recognize the deformation of coke faster than 6 layers. The reason why 6 layers are slower in recognition speed than 1 layer is that an algorithm for reducing overfitting works during machine learning and classification, and neurons are randomly eliminated (dropout). Therefore, although the recognition speed of the deformation is inferior, it is considered that performing machine learning and classification of the deformation characteristics of coke with 6 layers can suppress overfitting, so that a highly accurate classification result can be obtained.

[0095] FIG. 10 shows the transition of the loss function E when the above-described information processing method is executed by a six-layer neural network, and it was confirmed that the loss function E gradually decreases.

[0096] 〔B〕Others The technology according to the above-described embodiment can be implemented with the following modifications and changes.

[0097] For example, the acquisition unit 11, generation unit 12, machine learning unit 13, and classification unit 14 included in the information processing apparatus 1 shown in FIG. 1 may be merged in any combination or divided respectively.

[0098] In this embodiment, four-dimensional data 23 is generated from the three-dimensional model 21. However, the three-dimensional model 21 may be scan (point cloud) data at the previous stage, surface data at the subsequent stage, or voxel data, in addition to polygon data, and may be acquired as voxel data, but is not limited thereto.

[0099] In this embodiment, coke is taken as an example for explanation. However, this embodiment is also applicable to substances other than coke. That is, according to the information processing program 1g, information processing apparatus 1, and information processing method in this embodiment, for a three-dimensional model of a substance other than coke, a plurality of types of labels are respectively assigned on a three-dimensional lattice according to the hardness distribution in the substance, and data of the three-dimensional model of a plurality of batches to which the plurality of types of labels are assigned is received as input as four-dimensional data composed of the number of batches of the three-dimensional model and the three-dimensional elements of the three-dimensional model. For the four-dimensional data, convolution processing is performed by multiplying by filter elements and adding biases, and one-dimensional fully connected data can be obtained as output data.

Explanation of Signs

[0100] 1 Information processing apparatus (computer) 1a Processor 1b Memory 1c Storage unit 1d IF unit 1e IO unit 1f Reading unit 1g Program (information processing program) 1h Recording medium 10 Control unit 11 Acquisition unit 12 Generation unit 13 Machine learning unit 14 Classification unit 20 Memory unit 21 Three-dimensional model 22 Teacher data 23 Four-dimensional data (data structure) 24 Output data 25 Classification model 30 Input unit 40 Output unit

Claims

1. For a three - dimensional model of coke, a plurality of types of labels are respectively assigned to a three - dimensional lattice according to the hardness distribution within the substance of the coke, and data of the three - dimensional model of a plurality of batches to which the plurality of types of labels are assigned is received as input as four - dimensional data consisting of the number of batches of the three - dimensional model and the three - dimensional elements of the three - dimensional model. Convolution processing is performed on the four - dimensional data by multiplying by filter elements and adding a bias, and one - dimensional fully - connected data is obtained as output data. An information processing program for causing a computer to execute the processing.

2. Using the output data, classification of the coke according to the hardness distribution is performed. The information processing program according to claim 1, for causing the computer to execute the processing.

3. The hardness distribution of the three - dimensional model is determined by the shape change of the coke in a rotary tumbler. The information processing program according to claim 1 or 2.

4. For each site on the lattice that remains without pulverization due to an increase in the rotation speed of the rotary tumbler, the plurality of types of labels are assigned. The information processing program according to claim 3, for causing the computer to execute the processing.

5. A three - dimensional model of coke is obtained. For the three - dimensional model, a plurality of types of labels are respectively assigned to a three - dimensional lattice according to the hardness distribution within the substance of the coke, and four - dimensional data consisting of the number of batches of the three - dimensional model and the three - dimensional elements of the three - dimensional model is generated from the data of the three - dimensional model of a plurality of batches to which the plurality of types of labels are assigned. The input of the four - dimensional data is received, convolution processing is performed on the four - dimensional data by multiplying by filter elements and adding a bias, and one - dimensional fully - connected data is obtained as output data. An information processing apparatus including a processor.

6. An input unit that inputs the three-dimensional model to the processor, and an output unit that outputs one-dimensional fully-connected data acquired by the processor. The information processing apparatus according to claim 5, comprising:

7. Acquire a three-dimensional model of coke, For the three-dimensional model, respectively assign a plurality of types of labels on a three-dimensional lattice according to the hardness distribution within the coke substance, Generate four-dimensional data consisting of the number of batches of the three-dimensional model and the three-dimensional elements of the three-dimensional model from the data of the three-dimensional model of a plurality of batches to which the plurality of types of labels are assigned, Receive an input of the four-dimensional data, perform a convolution process on the four-dimensional data by multiplying by filter elements and adding a bias, and acquire one-dimensional fully-connected data as output data. An information processing method in which a computer executes the process.

8. For a three-dimensional model of a substance, respectively assign a plurality of types of labels on a three-dimensional lattice according to the hardness distribution within the substance, and receive an input of the data of the three-dimensional model of a plurality of batches to which the plurality of types of labels are assigned as four-dimensional data consisting of the number of batches of the three-dimensional model and the three-dimensional elements of the three-dimensional model, Perform a convolution process on the four-dimensional data by multiplying by filter elements and adding a bias, and acquire one-dimensional fully-connected data as output data. An information processing program for causing a computer to execute the process.

Images