Classification result prediction method, particle manufacturing method, classification result prediction device, and classification system

Abstract

To provide a method for accurately predicting a relationship between a particle diameter and a recovery rate of particles in particle classification using a classifier.SOLUTION: A classification result prediction method includes the steps of: acquiring a first learned model having machine-learned on the basis of first device conditions of a classifier and a result of fitting, with a function, a relationship between a particle diameter and a recovery rate of particles in classification of particles under the first device conditions using the classifier; and inputting second device conditions of the classifier to the first learned model and estimating, from an obtained output result, the relationship between the particle diameter and the recovery rate of particles under the second device conditions.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The disclosure of the present specification relates to a classification result prediction method, a particle manufacturing method, a classification result prediction device, and a classification system. [Background technology]

[0002] Particle classification using a classifier is an important process that is widely used in a variety of fields, and attempts have been made to accurately and quickly predict the particle classification performance of a classifier. In Non-Patent Document 1, in a cyclone classifier, the "particle diameter at which centrifugal force and fluid drag are balanced" is considered to be equivalent to the "particle diameter at which the recovery rate is 50%," and this is calculated by computer simulation. The results of this computer simulation are learned by a machine learning model, and the "particle diameter at which the recovery rate is 50%" according to the conditions is instantly output. In Non-Patent Document 2, a machine learning model is constructed that learns the difference between the results of computer simulation and experimental results for the "particle size at which the recovery rate is 50%". By using the trained machine learning model thus constructed, the accuracy of the prediction of the "particle size at which the recovery rate is 50%" by numerical calculation is improved with respect to the actual experimental results. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] D. Park and JS Go, “Design of Cyclone Separator Critica Diameter Model Based on Machine Learning and CFD,” Processes, vol. 8, no. 11, p. 1521, Nov. 2020, doi: 10.3390 / pr8111521. [Non-Patent Document 2] DK Le, M. Guo, and JY Yoon, “A hybrid CFD - Deep Learning methodology to improve the accuracy of cut-off diameter prediction in coarse-grid simulations for cyclone separators,” Journal of Aerosol Science, vol. 170, p. 106143, May 2023, doi: 10.1016 / j.jaerosci.2023.106143. Summary of the Invention [Problem to be solved by the invention]

[0004] In a particle manufacturing process using a classification device, it is important to predict in advance how many particles of a desired particle size can be recovered under certain device conditions. However, while the methods described in Non-Patent Documents 1 and 2 can quickly and accurately predict the threshold particle size that is intermediate between "recovered and not recovered," it is difficult to predict the recovery rate of particles having particle sizes larger or smaller than that threshold particle size.

[0005] The present invention has been made in view of the above problems, and an object of the present invention is to provide a method for accurately predicting the relationship between particle diameter and particle recovery rate in particle classification using a classification device. [Means for solving the problem]

[0006] A classification result prediction method according to one embodiment of the present invention includes a step of acquiring a first trained model through machine learning based on first apparatus conditions of a classification device and a result of fitting a function to the relationship between particle diameter and particle recovery rate in particle classification under the first apparatus conditions using the classification device, and a step of inputting second apparatus conditions of the classification device to the first trained model and estimating the relationship between particle diameter and particle recovery rate under the second apparatus conditions from the obtained output results.

[0007] In addition, a classification result prediction method according to another aspect of the present invention includes the steps of: acquiring a first trained model obtained by machine learning based on a correction coefficient and the first apparatus conditions, where the correction coefficient is a coefficient indicating the difference between a result of fitting a prediction result obtained using an information processing device with a function regarding the relationship between particle diameter and particle recovery rate in particle classification under first apparatus conditions using a classification device and a result of fitting an experimental result obtained by classifying particles using the classifier with the function; acquiring a result of fitting a prediction result obtained using the information processing device with the function regarding the relationship between particle diameter and particle recovery rate in classification under second apparatus conditions using the classifier; and correcting the prediction result obtained for classification under the second apparatus conditions using an output result obtained by inputting the second apparatus conditions to the first trained model.

[0008] Furthermore, a particle manufacturing method according to another aspect of the present invention includes a condition determination step of determining device conditions for the classification device based on a classification result predicted using the above-mentioned classification result prediction method, and a classification step of classifying particles using the classification device under the device conditions determined in the condition determination step.

[0009] In addition, a classification result prediction device according to another aspect of the present invention has a trained model acquisition unit that acquires a first trained model that has been machine-learned based on the device conditions of a classification device and the results of fitting a function to the relationship between particle diameter and particle recovery rate in particle classification under first device conditions using the classification device, and an estimation unit that inputs second device conditions of the classification device to the first trained model and estimates the relationship between particle diameter and particle recovery rate under the second device conditions from the obtained output results.

[0010] In addition, a classification result prediction device according to another aspect of the present invention has: a trained model acquisition unit that acquires a first trained model that has been machine-learned based on a correction coefficient, where the correction coefficient is a coefficient indicating the difference between a result of fitting a prediction result obtained using an information processing device with a function regarding the relationship between particle diameter and particle recovery rate in particle classification under first device conditions using a classifier, and a result of fitting an experimental result obtained by classifying particles using the classifier with the function; a fitting result acquisition unit that acquires the result of fitting the prediction result obtained using the information processing device with the function regarding the relationship between particle diameter and particle recovery rate in classification under second device conditions using the classifier; and a correction unit that corrects the fitting result obtained for classification under the second device conditions using an output result obtained by inputting the second device conditions to the first trained model.

[0011] A classification system according to another aspect of the present invention includes the above classification result prediction device and the above classification device. Effect of the Invention

[0012] According to the present invention, there is provided a method for accurately predicting the relationship between particle size and particle recovery rate in particle classification using a classifier. [Brief description of the drawings]

[0013] [Figure 1]1 is a functional block diagram of a classification result prediction device according to a first embodiment of the present invention. [Diagram 2] 3 is a flowchart showing an outline of a process performed by the classification result prediction device according to the first embodiment of the present invention. [Diagram 3] 1 is a block diagram showing a hardware configuration of a classification result prediction device according to a first embodiment of the present invention. [Figure 4] FIG. 11 is a functional block diagram of a classification result prediction device according to a second embodiment of the present invention. [Diagram 5] 10 is a flowchart showing an outline of a process performed by a classification result prediction device according to a second embodiment of the present invention. [Figure 6] FIG. 11 is a functional block diagram of a classification system according to a third embodiment of the present invention. [Figure 7] FIG. 2 is a schematic cross-sectional view of a classification device. [Figure 8] This figure shows the calculated recovery rate curve by computer simulation and the predicted recovery rate curve by trained model A. [Figure 9] This figure shows the experimental values, predicted values ​​by trained model A (uncorrected), and predicted values ​​by trained model A corrected by trained model B for the recovery rate curve. [Figure 10] 1A shows the particle size distribution after classification, where (a) shows the particle size distribution for the experimental values ​​and the particle size distribution for the predicted values ​​(uncorrected) by trained model A, and (b) shows the particle size distribution for the experimental values ​​and the particle size distribution for the predicted values ​​by trained model A corrected using trained model B. [Figure 11] FIG. 13 is a diagram showing the results of predicting a recovery rate curve using a trained model C relating to a comparative example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, the embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, similar or corresponding elements are denoted by the same reference numerals, and their description may be omitted or simplified. Note that the embodiments exemplified below do not limit the present invention according to the claims, and all of the combinations of features described in each embodiment are not necessarily essential to the solution of the present invention.

[0015] First Embodiment FIG. 1 shows an example of the configuration of an apparatus for carrying out the classification result prediction method according to this embodiment. As shown in FIG. 1, a classification result prediction device 1 according to this embodiment includes a trained model acquisition unit 11, an estimation unit 12, an input unit 13, an output unit 14, and a storage unit 15. An example of a classification result prediction method using the classification result prediction device 1 according to this embodiment will be described below with reference to a flowchart of FIG.

[0016] First, in step S11, the trained model acquisition unit 11 acquires a first trained model based on instructions input by a user from the input unit 13. Here, the first trained model is a trained model that is machine-learned based on first device conditions of a classifier and results of fitting a function to the relationship between particle diameter and particle recovery rate in particle classification under the first device conditions using the classifier.

[0017] The trained model acquisition unit 11 may acquire the first trained model via a network or the like from a server or the like external to the classification result prediction device 1. In addition, the first trained model may be stored in the memory unit 15 in advance, and the trained model acquisition unit 11 may acquire the first trained model from the memory unit 15.

[0018] The first device conditions of the classification device used by the first trained model for machine learning are conditions that affect the classification characteristics of particles, and may include, for example, setting conditions related to the structure of the classification device, operating conditions of the classifier, and environmental conditions in which the classifier is placed during classification.

[0019] The first trained model is also a model in which the result of fitting the relationship between particle size and particle recovery rate with a function is used for machine learning. In this embodiment, the classification device to be used for predicting the classification result is configured to classify a plurality of particles according to particle size. The relationship between particle size and particle recovery rate indicates the recovery rate for each particle size when the particles are classified under the first device conditions using the classification device. The relationship between particle size and particle recovery rate can be a relationship for particles of a predetermined particle size division classified by the classification device. That is, for example, when particles are classified into three divisions of large particle size, medium particle size, and small particle size according to the size of the particle size, the relationship between particle size and particle recovery rate can be a relationship for particles classified into the medium particle size division.

[0020] As the function used for fitting, any function can be used without particular limitation as long as it can express the relationship between the particle size and the particle recovery rate. Specific functions that can be used include, for example, an error function, a complementary error function, a sigmoid function, a hyperbolic tangent function, and an inverse hyperbolic sine function.

[0021] The relationship between particle size and particle recovery rate used for learning by the first trained model may include results obtained by prediction through computer simulation. This makes it possible to easily prepare a large number of relationships between particle size and particle recovery rate corresponding to each of a plurality of different first device conditions, and the first trained model can be trained based on a large number of training data.

[0022] The first trained model is machine-trained using the results of fitting a function to the relationship between particle size and particle recovery rate, where the results of fitting the function can include fitting parameters of the function.

[0023] For example, the first trained model is prepared as follows. First, the results of classifying particles using a classifier under certain first equipment conditions are calculated using a computer simulation.

[0024] Next, a curve is created by plotting the relationship between particle size and particle recovery rate obtained by computer simulation, with particle size on the horizontal axis and the recovery rate of that particle size on the vertical axis. Hereinafter, this curve is called the recovery rate curve.

[0025] Next, curve fitting is performed on the recovery curve to prepare a functional form f(d) that can be expressed by a few fitting parameters without losing information.

[0026] Similarly, the classification results when the first apparatus conditions are changed are calculated by computer simulation, and many recovery rate curves corresponding to the changed first apparatus conditions are prepared. Then, by performing the above curve fitting for each recovery rate curve, the fitting parameters of f(d) obtained for each first apparatus condition are obtained.

[0027] Next, a machine learning model is prepared, and the machine learning model is trained by using a plurality of first apparatus conditions and the fitting parameters of f(d) obtained for each of the plurality of first apparatus conditions, thereby preparing a first trained model.

[0028] In addition, any machine learning model can be used as the machine learning model for machine learning, as long as it is an appropriate machine learning model that can learn the relationship between the first apparatus conditions and the relationship between particle size and particle recovery rate.

[0029] Next, in step S12, the estimation unit 12 inputs the second device conditions of the classification device to the first trained model acquired by the trained model acquisition unit 11, and estimates the relationship between the particle diameter and the particle recovery rate under the second device conditions from the obtained output result. Here, the second device conditions are device conditions that are the basis for obtaining the result to be predicted by the classification result prediction method according to this embodiment, and include the same condition elements as the first device conditions. That is, for example, if the first trained model is trained using the cross-sectional area of ​​the flow path through which the particles pass in the classification device and the flow rate of air supplied to the device for classification as the first device conditions, the second device conditions also include the above cross-sectional area and the above flow rate.

[0030] By inputting the second device conditions of the classification device into the first trained model, the output result is obtained as a "function fitting result for the relationship between particle size and particle recovery rate" corresponding to the second device conditions. That is, for example, consider the case where the first trained model is machine-learned using the fitting parameters of f(d) as described in the above example. In this case, by inputting the second device conditions into the first trained model, the fitting parameters of f(d) for the classification result under the second device conditions are obtained as an output result. Next, the obtained fitting parameters are substituted for f(d), and a graph of f(d) is drawn with the particle size on the horizontal axis and the recovery rate of that particle size on the vertical axis. By viewing the graph obtained in this way as a recovery rate curve corresponding to the second device conditions, the relationship between the particle size and the recovery rate of the particles under the second device conditions can be estimated. The result of estimation by the estimation unit 12 in step S12 may be presented to the user via the output unit 14, or may be stored in the storage unit 15.

[0031] The classification device for performing classification, the results of which are predicted by applying the classification result prediction method according to this embodiment, can be any device that can be used for classification, without any particular limitations. For example, the classification device may be configured to classify particles by utilizing a gas flow. In this case, the first device condition and the second device condition may each include a flow rate of gas introduced into the classification device. In addition, the classification device may be configured to classify particles by sucking gas from a plurality of flow paths. In this case, the first device condition and the second device condition may each include a flow rate of gas sucked from each of the plurality of flow paths. In addition, for example, the classification device may include a plurality of flow paths through which the classified particles flow, an edge that branches the plurality of flow paths and defines the openings of each of the plurality of flow paths, and a movable mechanism for the edge to change the size of the opening. In this case, the first device condition and the second device condition may each include a setting condition for the movable mechanism. Furthermore, the movable mechanism may be configured to rotate the edge about an axis, and in this case, the setting condition for the movable mechanism may include an angle at which the edge is rotated. In addition, the movable mechanism may be configured to translate the edge, and in this case, the setting condition for the movable mechanism may include a displacement amount for translating the edge.

[0032] In this embodiment, the first trained model is not trained using the relationship between particle size and particle recovery rate as it is, but is trained using the obtained result after fitting the relationship with a function. This allows the dimension of the output vector of the first trained model to be lowered, and it is possible to perform highly accurate machine learning in a short time in learning using a large amount of teacher data. In addition, since the first trained model has a lower dimension of the output vector than a trained model trained using the relationship between particle size and particle recovery rate itself, when the second device conditions are input in step S12, it is possible to output a highly accurate result in a short time. In other words, by using the classification result prediction method according to this embodiment, it is possible to accurately and quickly predict how many particles of a desired particle size can be recovered.

[0033] 3 is a block diagram showing a hardware configuration of the classification result prediction device 1 according to this embodiment. The classification result prediction device 1 has a CPU 101, a RAM 102, a ROM 103, a HDD (Hard Disk Drive) 104, a communication I / F 105, an input device 106, and an output device 107. These components are connected to each other via a bus or the like.

[0034] The CPU 101 is a processor that reads out to the RAM 102 programs stored in the ROM 103 and the HDD 104, executes them, and performs calculation processing and controls each part of the classification result prediction device 1. The processing performed by the CPU 101 may include obtaining a first trained model and estimating the relationship between particle diameter and particle recovery rate.

[0035] The RAM 102 is a volatile storage medium and functions as a work memory when the CPU 101 executes a program. The ROM 103 is a non-volatile storage medium and stores firmware and the like necessary for the operation of the classification result prediction device 1. The HDD 104 is a non-volatile storage medium and stores a program used for processing such as obtaining the first trained model and estimating the relationship between particle diameter and particle recovery rate in this embodiment.

[0036] The communication I / F 105 is a communication device based on standards such as Wi-Fi (registered trademark), Ethernet (registered trademark), Bluetooth (registered trademark), etc. The communication I / F 105 is used for communication with a classification device, etc.

[0037] The input device 106 is a device for inputting information to the classification result prediction device 1, and is typically a user interface for a user to operate the classification result prediction device 1. Examples of the input device 106 include a keyboard, a button, a mouse, a touch panel, and the like.

[0038] The output device 107 is a device that outputs information to the outside from the classification result prediction device 1, and is typically a user interface for presenting information to a user. Examples of the output device 107 include a display and a speaker.

[0039] The hardware configuration of the classification result prediction device 1 described above is an example, and can be changed as appropriate. For example, examples of processors that can be mounted on the classification result prediction device 1 include GPU, ASIC, FPGA, etc., in addition to the above-mentioned CPU 101. Furthermore, a plurality of these processors may be provided, and the plurality of processors may perform processing in a distributed manner. Furthermore, the HDD 104 may be a storage medium such as an optical disk, a magneto-optical disk, or an SSD (Solid State Drive).

[0040] Second Embodiment Next, in this embodiment, a method for predicting classification results will be described, which includes quickly and accurately correcting the results predicted using an information processing device regarding the relationship between particle diameter and particle recovery rate in particle classification using a classifier.

[0041] FIG. 4 shows an example of the configuration of an apparatus for carrying out the classification result prediction method according to this embodiment. As shown in FIG. 4, the classification result prediction device 2 according to this embodiment includes a trained model acquisition unit 21, a fitting result acquisition unit 22, a correction unit 23, an input unit 13, an output unit 14, and a storage unit 15.

[0042] An example of a classification result prediction method using the classification result prediction device 2 according to this embodiment will be described below with reference to the flowchart of FIG. First, in step S21, the trained model acquisition unit 21 acquires a first trained model based on an instruction input by a user from the input unit 13. The trained model acquisition unit 21 may acquire the first trained model from a server or the like external to the classification result prediction device 2 via a network or the like. Alternatively, the first trained model may be stored in the memory unit 15 in advance, and the trained model acquisition unit 21 may acquire the first trained model from the memory unit 15.

[0043] In this embodiment, the first trained model is a trained model that is machine-learned based on the correction coefficient and the first device conditions. Here, the correction coefficient is defined as a coefficient indicating the difference between the following two results. One is a result of fitting a predicted result obtained using an information processing device with a function for the relationship between the particle diameter and the particle recovery rate in the classification of particles under the first device conditions using a classifier. Hereinafter, this result is called a calculated value fitting result. The other is a result of fitting an experimental result obtained by classifying particles using a classifier with the above function, and hereinafter, this result is called an experimental value fitting result. That is, the correction coefficient is a coefficient indicating the difference between the calculated value fitting result and the experimental value fitting result, and the first trained model according to this embodiment is a trained model obtained by using this correction coefficient in machine learning.

[0044] In the above-mentioned calculation value fitting result and the above-mentioned experimental value fitting result, as a function used for fitting, as in the first embodiment, any function can be used without particular limitation as long as it can express the relationship between the particle diameter and the particle recovery rate. In addition, as the first trained model according to this embodiment, any machine learning model can be used without particular limitation as long as it is an appropriate machine learning model that can learn the relationship between the correction coefficient and the first device condition.

[0045] The above prediction results obtained using an information processing device are, for example, results obtained by computer simulation.

[0046] Furthermore, for example, the result of fitting the prediction result obtained using an information processing device with the function may be an output result obtained by inputting the first device condition to the second trained model. Here, the first trained model described in the first embodiment above can be suitably used as the second trained model. That is, the second trained model may be a trained model that is machine-learned based on the third device condition of the classifier and the result of fitting the relationship between the particle diameter and the particle recovery rate in the classification of particles under the third device condition using the classifier with the function. Here, the third device condition corresponds to the first device condition described in the first embodiment above.

[0047] In this embodiment, the classification device for predicting the classification result is configured to classify a plurality of particles according to their particle diameters, as in the first embodiment, and the relationship between the particle diameter and the particle recovery rate can be the relationship for particles of a predetermined particle diameter range classified by the classification device.

[0048] In the above-mentioned calculation value fitting result and the above-mentioned experimental value fitting result, the result of fitting with a function may include the fitting parameters of the function, and in this case, the correction coefficient may include a ratio between the fitting parameters obtained for the prediction result obtained using an information processing device and the fitting parameters obtained for the experimental result obtained by classifying particles using a classifier.

[0049] Next, in step S22, the fitting result acquisition unit 22 acquires the result of fitting the prediction result obtained by the information processing device with the above function for the relationship between the particle diameter and the particle recovery rate when the particle is classified under the second device conditions using the classification device. Here, the second device conditions are device conditions that are the basis for obtaining the result to be predicted by the classification result prediction method according to the present embodiment.

[0050] Next, in step S23, the correction unit 23 corrects the prediction result obtained for classification under the second device conditions obtained in step S22, using the output result obtained by inputting the second device conditions to the first trained model.

[0051] By inputting the second device conditions to the first trained model, a correction coefficient corresponding to the second device conditions can be obtained as an output. Furthermore, the result acquired by the fitting result acquisition unit 22 in step S22 can be regarded as a calculated value fitting result corresponding to the second device conditions. Therefore, by correcting the calculated value fitting result acquired in step S22 using the correction coefficient obtained in step S23, a result corresponding to the experimental value fitting result corresponding to the second device conditions can be obtained.

[0052] In other words, in this embodiment, the result predicted by the information processing device (obtained in step S22) is corrected (step S23) using the correction coefficient obtained as the output from the first trained model (obtained in step S21). This makes it possible to obtain a more accurate result than would be expected to be obtained as an experimental value.

[0053] In this embodiment, the correction coefficient used by the first trained model for learning is not the relationship between particle size and particle recovery rate itself, but is determined based on the result obtained by fitting the relationship with a function. This allows the dimension of the output vector of the first trained model to be lowered, making it possible to perform highly accurate machine learning in a short time in learning using a large amount of teacher data. In addition, since the dimension of the output vector of the first trained model is lower than that in the case where the relationship between particle size and particle recovery rate itself is used, when the second device conditions are input in step S23, highly accurate results can be output in a short time. Therefore, by using the classification result prediction method according to this embodiment, it is possible to accurately and quickly predict how many particles of a desired particle size can be recovered.

[0054] The classification device for performing classification to predict the result by applying the classification result prediction method according to this embodiment is similar to the classification device described in the first embodiment, and can be any device used for classification without any particular restrictions. In this embodiment, the first device conditions, the second device conditions, and the third device conditions can include the same conditions as the first device conditions and the second device conditions in the first embodiment, depending on the classification device used. Moreover, the hardware configuration of the classification result prediction device 2 according to this embodiment is similar to that described in the first embodiment.

[0055] <Third embodiment> The classification result prediction device according to the first embodiment and the classification result prediction device according to the second embodiment described above can be used as a classification system in combination with a classification device.

[0056] 6 is a functional block diagram showing an example of a classification system according to a third embodiment of the present invention. The classification system 3 according to this embodiment has a classification result prediction device 1 and a classification device 31. Here, an example is shown in which the classification result prediction device 1 is the classification result prediction device 1 according to the first embodiment, but the classification result prediction device 2 according to the second embodiment may be used instead of the classification result prediction device 1. The classification device 31 is similar to that described in the first and second embodiments.

[0057] The classification system according to this embodiment can be suitably used in the production of particles. That is, the method for producing particles using the classification system according to this embodiment can include a condition determination step and a classification step. Here, the condition determination step is a step of determining the device conditions of a classification device based on the classification result predicted using the classification result prediction method described above. Also, the classification step is a step of classifying particles by the classification device under the device conditions determined in the condition determination step.

[0058] In the classification system 3, for example, the classification result prediction device 1 may be configured to output device conditions corresponding to the predicted result from the output unit 14 to the classifier 31, and the classifier 31 may be configured to classify under the device conditions acquired from the output unit 14. Also, for example, the classification result prediction device 1 may be configured to perform classification in the classifier and acquire the result from the input unit 13. EXAMPLES

[0059] The present invention will be described in more detail below using examples and comparative examples. The present invention is not limited to the following examples in any way as long as the gist of the present invention is not exceeded.

[0060] <Example 1> The classification result prediction method according to the present invention was applied to a wind classifier that classifies input particles into three levels (large, medium, and small) according to the particle diameter. This classifier is often used to remove unwanted fine and coarse particles from the particles being fed in. Therefore, it was designed to collect particles with a "medium" particle size. A schematic cross-sectional view showing the particle flow path of this classification device is shown in Figure 7. Particles fed from an inlet 61 are branched according to particle diameter by edges 62 and 63. As a result, small, medium and large particles head towards outlets 64, 65 and 66, respectively.

[0061] The air classifier is operated under the following equipment conditions: Flow rate of air injected into inlet 61 The flow rate of air taken in from each of the outlets 64, 65, and 66 The angle by which edges 62 and 63 are rotated about axes 67 and 68, respectively. Displacement amount by which the edges 62 and 63 are translated in the vertical direction of the paper surface of FIG. By adjusting these apparatus conditions, particles having a desired particle size distribution are collected from outlet 65 .

[0062] First, the airflow movement inside the classifier was calculated and reproduced by three-dimensional fluid computer simulation. Next, the particle movement inside the classifier was calculated and reproduced by computer simulation. For the particle computer simulation, a discrete phase model (DPM) was used, in which the particles are moved by the force of the airflow calculated by the fluid computer simulation, and the interaction between the particles and the airflow is ignored. Particles were prepared with a total of 161 particle sizes between 1 μm and 57 μm, and 300 particles of each particle size were introduced from the inlet 61. This computer simulation was carried out for multiple levels of the classification device's device conditions (air flow rate, edge angle, and displacement). As a result, the percentage of the number of particles recovered from the outlet 65 out of 300 particles of each particle size was obtained for multiple device conditions. This percentage was taken as the recovery rate.

[0063] The calculated recovery curves for each device condition obtained by computer simulation were fitted using the following equation 1, which is a modified error function. Formula 1:

number

[0064] Next, we created a trained model that inputs the instrument conditions and outputs the fitting parameters of the recovery rate curve. This trained model is called trained model A. Trained model A was trained using the instrument conditions and fitting parameters obtained by fitting the calculated recovery rate curve obtained by computer simulation with Equation 1. Trained model A was created using a fully connected multilayer perceptron as a machine learning model.

[0065] Figure 8 shows the calculated recovery rate curve obtained by computer simulation under certain equipment conditions and the predicted recovery rate curve obtained by trained model A. As shown in Figure 8, the computer simulation results were reproduced well by using trained model A. While a computer simulation requires at least a dozen minutes, the output time of trained model A is less than one second, and the recovery rate curve was predicted more quickly than the computer simulation.

[0066] <Example 2> In Example 2, a trained model was created that outputs a correction coefficient for a predicted value obtained by the trained model for the recovery rate curve by inputting the device conditions. Here, the trained model that outputs a predicted value for the recovery rate curve based on the device conditions is trained model A2, and the trained model that outputs a correction value for the predicted value output by trained model A2 based on the device conditions is trained model B.

[0067] First, particles with a particle size distribution centered around 5 to 7 μm were prepared as particles to be subjected to classification. Next, an actual classification device was used to classify the prepared particles under certain device conditions, and the ratio of the particle size distribution before classification to the particle size distribution after classification was calculated to obtain an experimental value of the recovery rate curve.

[0068] Next, using the trained model A2, a predicted value of the recovery rate curve was obtained under the same equipment conditions as when the experimental value was obtained. The experimental and predicted values ​​of the recovery rate curve were obtained for multiple levels by changing the equipment conditions of the classification device. The experimental and predicted values ​​of the obtained recovery curve were fitted using the above formula 1 to obtain fitting parameters that represent those curves. Then, the ratio of the fitting parameters obtained for the experimental values ​​to the fitting parameters obtained for the predicted values ​​was calculated. In other words, r1 = a1(experimental value) / a1(predicted value) r2= b1(experimental value) / b1(predicted value) r3= a2(experimental value) / a2(predicted value) r4= b2(experimental value) / b2(predicted value) Thus, four ratio values ​​were obtained.

[0069] Finally, the machine learning model was trained using the instrument conditions and the ratios r1 to r4 between the fitting parameters obtained for the experimental and predicted values ​​of the fitting parameters, respectively, to obtain trained model B. Note that random forest regression was used as the machine learning model to obtain trained model B.

[0070] Figure 9 shows the experimental values, predicted values ​​by trained model A (without correction), and predicted values ​​by trained model A corrected using trained model B for a recovery rate curve under certain equipment conditions. Compared to the uncorrected predicted values, the correction by trained model B showed improved agreement with the experimental values, especially for the recovery rate of particles with a particle size of 5 μm or less. Note that the phenomenon in which the recovery rate is high for experimental values ​​of 3 μm or less is due to the phenomenon in which a small number of particles are always accidentally recovered, and is therefore not subject to prediction. In addition, for particles with a particle size of about 10 to 15 μm, the experimental values ​​of the recovery rate fluctuate greatly due to the small number of particles before classification.

[0071] The particle size distribution after classification can be obtained by multiplying the predicted value of the recovery rate curve and the particle size distribution before classification. Figure 10(a) shows a comparison between the particle size distribution after classification estimated using this method based on the predicted value (without correction) by trained model A and the actual particle size distribution after classification obtained in an experiment. Figure 10(b) shows a comparison between the particle size distribution after classification estimated using the predicted value by trained model A corrected using trained model B and the actual particle size distribution after classification obtained in an experiment.

[0072] When using the uncorrected predicted value of the recovery curve (Figure 10(a)), it can be seen that the estimated particle size distribution is overestimated in the 5 μm or less region of the distribution. On the other hand, when using the predicted value of the recovery curve corrected using trained model B (Figure 10(b)), this difference is eliminated and good agreement is observed. In this way, by using the trained model B, it was possible to predict the classification results under certain equipment conditions with high accuracy.

[0073] <Comparative Example> As a comparative example to Example 1, an example of a simple extension of the prior art will be described. First, a trained model C was created by directly learning all of the results obtained from the computer simulation, including "recovery rate of particle diameter d_1 (μm)," "recovery rate of particle diameter d_2 (μm)," ... "recovery rate of particle diameter d_n (μm)." Using the trained model C obtained, the results of predicting a recovery rate curve under certain equipment conditions are shown in Figure 11, together with the calculation results obtained based on the results of a computer simulation under the same equipment conditions. In this case, since the dimension of the output vector in the trained model C is too large, the prediction accuracy of the calculated value of the recovery rate curve is lower than that in FIG. 8 of Example 1. [Industrial Applicability]

[0074] By adjusting the device conditions of the classification device using the method according to the present invention, the labor required for trial and error at the manufacturing site can be reduced, and the production loss caused by trial and error can also be reduced. Classification devices are used in a variety of fields, including food, cosmetics, and pharmaceuticals, making the present invention highly valuable.

[0075] Disclosure according to an embodiment of the present invention includes the following configurations and methods. (Method 1) A step of acquiring a first trained model by machine learning based on first device conditions of a classification device and a result of fitting a relationship between particle diameter and particle recovery rate in particle classification under the first device conditions using the classification device with a function; inputting second device conditions of the classification device into the first trained model, and estimating a relationship between particle diameter and particle recovery rate under the second device conditions from the obtained output results; The classification result prediction method includes: (Method 2) The method for predicting a classification result according to Method 1, wherein the result of fitting with the function includes fitting parameters of the function. (Method 3) A method for predicting a classification result according to method 1 or 2, wherein the relationship between particle size and particle recovery rate under the first apparatus conditions includes a result obtained by predicting using a computer simulation. (Method 4) The method for predicting a classification result according to any one of Methods 1 to 3, wherein the classifying device is configured to classify a plurality of particles according to particle diameter. (Method 5) The method for predicting a classification result according to any one of Methods 1 to 4, wherein the relationship between the particle size and the particle recovery rate is a relationship for particles of a predetermined particle size range classified by the classifier. (Method 6) The classifier is configured to classify particles using a gas flow; The classification result prediction method according to any one of Methods 1 to 5, wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas introduced into the classification apparatus. (Method 7) The classifying device is configured to classify particles by sucking gas through a plurality of flow paths, The classification result prediction method according to any one of Methods 1 to 6, wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas taken in from each of the plurality of flow paths. (Method 8) the classification device includes a plurality of flow paths through which classified particles flow, an edge that branches the plurality of flow paths and defines an opening of each of the plurality of flow paths, and a movable mechanism for the edge that changes the size of the opening; The classification result prediction method according to any one of Methods 1 to 7, wherein the first device conditions and the second device conditions each include setting conditions of the movable mechanism. (Method 9) The movable mechanism is configured to rotate the edge about an axis; The classification result prediction method according to Method 8, wherein the setting conditions of the movable mechanism include an angle at which the edge is rotated. (Method 10) The movable mechanism is configured to translate the edge, The classification result prediction method according to Method 8, wherein the setting conditions of the movable mechanism include a displacement amount by which the edge is translated. (Method 11) a step of acquiring a first trained model by machine learning based on a correction coefficient and the first apparatus conditions, the correction coefficient being a coefficient indicating the difference between a result of fitting a prediction result obtained using an information processing device with a function regarding the relationship between particle diameter and particle recovery rate in particle classification under first apparatus conditions using a classification device and a result of fitting an experimental result obtained by classifying particles using the classification device with the function; a step of acquiring a result of fitting a prediction result obtained using the information processing device with the function with respect to a relationship between a particle size and a particle recovery rate when the particle is classified using the classification device under second device conditions; correcting a prediction result obtained about classification under the second apparatus conditions using an output result obtained by inputting the second apparatus conditions to the first trained model; The classification result prediction method includes: (Method 12) 12. The method for predicting a classification result according to method 11, wherein the result of fitting with the function includes fitting parameters of the function. (Method 13) The classification result prediction method described in Method 12, wherein the correction coefficient includes a ratio between a fitting parameter obtained for the predicted result obtained using the information processing device and a fitting parameter obtained for the experimental result obtained by classifying particles using the classification device. (Method 14) A result of fitting the prediction result obtained using the information processing device with the function is an output result obtained by inputting the first device condition to a second trained model, The method for predicting a classification result described in any one of Methods 11 to 13, wherein the second trained model is a trained model machine-learned based on third apparatus conditions of the classification device and the result of fitting the relationship between particle diameter and particle recovery rate in particle classification under the third apparatus conditions using the classification device with the function. (Method 15) 15. The method for predicting a classification result according to any one of Methods 11 to 14, wherein the classification device is configured to classify a plurality of particles according to particle diameter. (Method 16) 16. The method for predicting a classification result according to any one of Methods 11 to 15, wherein the relationship between the particle size and the particle recovery rate is a relationship for particles of a predetermined particle size range classified by the classifier. (Method 17) The classifier is configured to classify particles using a gas flow; 17. The method for predicting a classification result according to any one of Methods 11 to 16, wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas introduced into the classification apparatus. (Method 18) The classifying device is configured to classify particles by sucking gas through a plurality of flow paths, 18. The classification result prediction method according to any one of Methods 11 to 17, wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas taken in from each of the plurality of flow paths. (Method 19) the classification device includes a plurality of flow paths through which classified particles flow, an edge that branches the plurality of flow paths and defines an opening of each of the plurality of flow paths, and a movable mechanism for the edge that changes the size of the opening; 19. The classification result prediction method according to any one of Methods 11 to 18, wherein the first device conditions and the second device conditions each include setting conditions of the movable mechanism. (Method 20) A condition determination step of determining device conditions of the classification device based on a classification result predicted by using the classification result prediction method according to any one of methods 1 to 19; a classification step of classifying particles by the classifier under the device conditions determined in the condition determination step; A method for producing particles comprising the steps of: (Configuration 1) a trained model acquisition unit that acquires a first trained model that is machine-learned based on the device conditions of a classifier and a result of fitting a relationship between particle diameter and particle recovery rate in particle classification under first device conditions using the classifier with a function; an estimation unit that inputs second device conditions of the classification device to the first trained model and estimates a relationship between a particle diameter and a particle recovery rate under the second device conditions from an obtained output result; A classification result prediction device having the above structure. (Configuration 2) a trained model acquisition unit that acquires a first trained model machine-learned based on a correction coefficient and the first apparatus conditions, where the correction coefficient indicates a difference between a result of fitting a prediction result obtained using an information processing device with a function regarding a relationship between a particle diameter and a particle recovery rate in particle classification under first apparatus conditions using a classification device and a result of fitting an experimental result obtained by classifying particles using the classification device with the function; a fitting result acquisition unit that acquires a result of fitting a prediction result obtained using the information processing device with the function regarding a relationship between a particle size and a particle recovery rate when the particle is classified under second device conditions using the classification device; a correction unit that corrects the fitting result obtained for classification under the second device conditions by using an output result obtained by inputting the second device conditions to the first trained model; A classification result prediction device having the above structure. (Configuration 3) A classification system comprising the classification result prediction device according to configuration 1 or 2 and the classification device. [Explanation of symbols]

[0076] 1, 2 Classification result prediction device 3. Classification system 11, 21 Trained model acquisition section 12 Estimation part 22 Fitting result acquisition unit 23 Correction section

Claims

1. A step of acquiring a first trained model by machine learning based on first device conditions of a classification device and a result of fitting a relationship between particle diameter and particle recovery rate in particle classification under the first device conditions using the classification device with a function; inputting second device conditions of the classification device into the first trained model, and estimating a relationship between particle diameter and particle recovery rate under the second device conditions from the obtained output results; The classification result prediction method includes:

2. The method of claim 1 , wherein the result of fitting with the function includes fitting parameters of the function.

3. 2. The method for predicting a classification result according to claim 1, wherein the relationship between the particle size and the particle recovery rate under the first apparatus conditions includes a result obtained by predicting using a computer simulation.

4. The method of claim 1 , wherein the classifier is configured to classify a plurality of particles according to particle diameter.

5. 2. The method for predicting a classification result according to claim 1, wherein the relationship between the particle size and the particle recovery rate is a relationship for particles of a predetermined particle size range classified by the classifier.

6. The classifier is configured to classify particles using a gas flow; The classification result prediction method according to claim 1 , wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas introduced into the classification apparatus.

7. The classifying device is configured to classify particles by sucking gas through a plurality of flow paths, The classification result prediction method according to claim 1 , wherein the first device condition and the second device condition each include a flow rate of the gas taken from each of the plurality of flow paths.

8. the classification device includes a plurality of flow paths through which classified particles flow, an edge that branches the plurality of flow paths and defines an opening of each of the plurality of flow paths, and a movable mechanism for the edge that changes the size of the opening; The classification result prediction method according to claim 1 , wherein the first device condition and the second device condition each include a setting condition of the movable mechanism.

9. The movable mechanism is configured to rotate the edge about an axis; The classification result prediction method according to claim 8 , wherein the setting conditions of the movable mechanism include an angle at which the edge is rotated.

10. The movable mechanism is configured to translate the edge, The classification result prediction method according to claim 8 , wherein the setting conditions of the movable mechanism include a displacement amount by which the edge is translated.

11. a step of acquiring a first trained model by machine learning based on a correction coefficient and the first apparatus conditions, the correction coefficient being a coefficient indicating the difference between a result of fitting a prediction result obtained using an information processing device with a function regarding the relationship between particle diameter and particle recovery rate in particle classification under first apparatus conditions using a classification device and a result of fitting an experimental result obtained by classifying particles using the classification device with the function; a step of acquiring a result of fitting a prediction result obtained using the information processing device with the function with respect to a relationship between a particle size and a particle recovery rate when the particle is classified using the classification device under second device conditions; correcting a prediction result obtained about classification under the second apparatus conditions by using an output result obtained by inputting the second apparatus conditions to the first trained model; The classification result prediction method includes:

12. The method of claim 11 , wherein the result of fitting the function includes fitting parameters of the function.

13. 13. The classification result prediction method according to claim 12, wherein the correction coefficient includes a ratio between a fitting parameter obtained for the prediction result obtained using the information processing device and a fitting parameter obtained for the experimental result obtained by classifying particles using the classification device.

14. A result of fitting the prediction result obtained using the information processing device with the function is an output result obtained by inputting the first device condition to a second trained model, The classification result prediction method according to claim 11, wherein the second trained model is a trained model machine-learned based on third apparatus conditions of the classification device and a result of fitting the relationship between particle diameter and particle recovery rate in particle classification under the third apparatus conditions using the classification device with the function.

15. The method of claim 11 , wherein the classifier is configured to classify a plurality of particles according to particle size.

16. The method for predicting a classification result according to claim 11 , wherein the relationship between the particle size and the particle recovery rate is a relationship for particles of a predetermined particle size range classified by the classifier.

17. The classifier is configured to classify particles using a gas flow; The classification result prediction method according to claim 11 , wherein the first apparatus condition and the second apparatus condition each include a flow rate of the gas introduced into the classification apparatus.

18. The classifying device is configured to classify particles by sucking gas through a plurality of flow paths, The classification result prediction method according to claim 11 , wherein the first device condition and the second device condition each include a flow rate of the gas taken in from each of the plurality of flow paths.

19. the classification device includes a plurality of flow paths through which classified particles flow, an edge that branches the plurality of flow paths and defines an opening of each of the plurality of flow paths, and a movable mechanism for the edge that changes the size of the opening; The classification result prediction method according to claim 11 , wherein the first device condition and the second device condition each include a setting condition of the movable mechanism.

20. A condition determination step of determining device conditions of the classification device based on a classification result predicted by using the classification result prediction method according to any one of claims 1 to 19; a classification step of classifying particles by the classifier under the device conditions determined in the condition determination step; A method for producing particles comprising the steps of:

21. a trained model acquisition unit that acquires a first trained model that is machine-learned based on the device conditions of a classifier and a result of fitting a relationship between particle diameter and particle recovery rate in particle classification under first device conditions using the classifier with a function; an estimation unit that inputs second device conditions of the classification device to the first trained model and estimates a relationship between a particle diameter and a particle recovery rate under the second device conditions from an obtained output result; A classification result prediction device having the above structure.

22. a trained model acquisition unit that acquires a first trained model machine-learned based on a correction coefficient and the first apparatus conditions, where the correction coefficient indicates a difference between a result of fitting a prediction result obtained using an information processing device with a function regarding a relationship between a particle diameter and a particle recovery rate in particle classification under first apparatus conditions using a classification device and a result of fitting an experimental result obtained by classifying particles using the classification device with the function; a fitting result acquisition unit that acquires a result of fitting a prediction result obtained using the information processing device with the function regarding a relationship between a particle diameter and a particle recovery rate when the particle is classified under second device conditions using the classification device; a correction unit that corrects the fitting result obtained for classification under the second device conditions by using an output result obtained by inputting the second device conditions to the first trained model; A classification result prediction device having the above structure.

23. A classification system comprising the classification result prediction device according to claim 21 or 22 and the classifier.

Images