Prediction method, prediction device, and prediction program

Abstract

This prediction method is executed by a computer and comprises: a step in which, using feature data of mineral particles, a prediction model is created by executing machine learning in which the feature data is used as training data; and a step in which the feature data is input into the prediction model, and on the basis of a value output by the prediction model, the prediction model predicts the recovery probability of the mineral particles by ore selection. The feature data of the mineral particles includes individual particle feature data indicating the characteristics of the mineral particles, and particle group feature data indicating the characteristics of a particle group composed of the plurality of mineral particles.

Description

Prediction method, prediction device, and prediction program

[0001] The present invention relates to a prediction method, a prediction device, and a prediction program for predicting the results of mineral processing.

[0002] Conventionally, mineral dressing techniques have been known that separate mined ore into useful minerals and waste minerals using techniques such as flotation, gravity separation, or magnetic separation. Because only a limited number of ore samples can be collected for testing during the exploration stage, it is necessary to predict the recovery rate, grade, and other properties of useful minerals with a limited number of tests. Non-Patent Document 1 discloses a method for predicting the recovery rate, grade, and other properties of useful minerals using a prediction model created by machine learning using training data on characteristic data for each particle in the mineral identified by a mineral particle analyzer (MLA). Non-Patent Document 2 also discloses a method for predicting the recovery rate, grade, and other properties of useful minerals using a machine-learned prediction model.

[0003] Shibuya, M., Kobayashi, S., Sakakibara, T., Ono, T., and Kawasaki, K., "Basic study on building a flotation prediction model from MLA data using machine learning," Journal of Mineral Resources and Materials Science, Vol. 9 (2022), No. 2, Japan Society of Mineral Resources and Materials Science, September 7, 2022. Pereira L, Frenzel M, Khodadadzadeh M, Tolosana-Delgado R, Gutzmer J. A self-adaptive particle-tracking method for minerals processing. J Clean Prod 2021;279:123711.

[0004] The behavior of mineral particles is thought to be influenced by other surrounding mineral particles. Therefore, when particle recovery probability is predicted using a machine-learned prediction model using only the characteristic data of each particle as training data, the recovery rate and grade can differ significantly from the actual recovery rate and grade.

[0005] Therefore, the present invention has been made in consideration of these points, and aims to provide a prediction method, a prediction device, and a prediction program that improve the prediction accuracy of the probability of recovery of mineral particles through ore dressing.

[0006] A first aspect of the present invention is a prediction method executed by a computer, comprising the steps of: using characteristic data of mineral particles to create a prediction model by machine learning using the characteristic data as training data; and inputting the characteristic data into the prediction model and predicting the probability of recovering mineral particles through mineral dressing based on the value output by the prediction model, wherein the characteristic data of the mineral particles includes individual particle characteristic data that indicates the characteristics of the mineral particle, and particle group characteristic data that indicates the characteristics of a particle group composed of a plurality of the mineral particles.

[0007] In the prediction method, the individual particle characteristic data and the particle group characteristic data used as the training data in the step of creating the prediction model may include individual characteristic data and particle group characteristic data of mineral particles in the recovered concentrate and tailings.

[0008] In the prediction method, the individual particle characteristic data and the particle group characteristic data used in the step of predicting the recovery probability of the mineral particles may include individual particle characteristic data and particle group characteristic data of mineral particles whose recovery probability is unknown.

[0009] The prediction method further includes the steps of: acquiring first condition data indicating the conditions under which the ore dressing is performed; acquiring second condition data indicating the conditions under which the training data was obtained; acquiring prior probability data indicating a prior probability which is the ratio of the number of concentrate particles or the ratio of the number of tailing particles to the number of feed particles corresponding to each of a plurality of conditions under which the ore dressing is performed; and identifying, in the prior probability data, an optimal prior probability corresponding to the condition closest to the condition indicated by the first condition data among the plurality of conditions indicated by the second condition data. In the step of predicting the recovery probability, the recovery probability may be predicted by correcting the value output by the prediction model using the optimal prior probability.

[0010] The prediction method may further include creating the prediction model by machine learning selected from random forest, neural network, and logistic regression analysis.

[0011] In the step of predicting the recovery probability, a portion of the feature quantities selected by dimension reduction from a plurality of feature quantities included in the acquired individual particle feature data and a portion of the feature quantities selected by dimension reduction from a plurality of feature quantities included in the acquired particle group feature data may be input to the prediction model.

[0012] A second aspect of the present invention is a prediction program for causing a computer to execute the steps of: using characteristic data of mineral particles to perform machine learning on the characteristic data as training data to create a prediction model; and inputting the characteristic data into the prediction model and predicting the probability of recovering mineral particles through mineral dressing based on the value output by the prediction model, wherein the characteristic data of the mineral particles includes individual particle characteristic data that indicates the characteristics of the mineral particle, and particle group characteristic data that indicates the characteristics of a particle group composed of a plurality of the mineral particles.

[0013] A prediction device of a third aspect of the present invention is a prediction device having a feature data acquisition unit that acquires feature data of mineral particles, and a prediction unit that inputs the feature data into a prediction model created by machine learning using the feature data as training data, and predicts the probability of recovering mineral particles by ore dressing based on the value output by the prediction model, wherein the feature data of the mineral particles includes individual particle feature data that indicates the characteristics of the mineral particle, and particle group feature data that indicates the characteristics of a particle group composed of a plurality of the mineral particles.

[0014] The present invention has the effect of improving the accuracy of prediction of the probability of recovery of mineral particles through ore dressing.

[0015] 1 is a diagram for explaining an overview of a prediction system according to an embodiment of the present invention. FIG. 1 is a diagram for explaining an overview of feature data. FIG. 2 is a diagram showing an example of training data. FIG. 3 is a diagram showing an overview of prediction data used when predicting recovery probability using a prediction model M. FIG. 4 is a diagram showing the configuration of a prediction device 1. FIG. 5 is a flowchart showing the processing flow of the learning phase in the prediction device 1. FIG. 6 is a flowchart showing the processing flow of the prediction phase in the prediction device 1. FIG. 7 is a table showing a list of prediction models used in comparative experiments. FIG. 8 is a diagram showing the recovery rate and grade of copper obtained by flotation of particles pulverized for a 15-minute crushing time. FIG. 9 is a diagram showing the predicted results of the grades of multiple types of minerals obtained by flotation of particles pulverized for a 15-minute crushing time. FIG. 10 is a diagram showing the recovery rate and grade of copper obtained by flotation of particles pulverized for a 35-minute crushing time. FIG. 11 is a diagram showing the predicted results of the grades of multiple types of minerals obtained by flotation of particles pulverized for a 35-minute crushing time.

[0016] [Outline of Prediction System] Fig. 1 is a diagram for explaining an outline of a prediction system according to this embodiment. The prediction system is a system for predicting at least one of the probability of recovery of mineral particles by ore dressing, and the recovery rate and grade of the mineral particles obtained by ore dressing.

[0017] Mineral dressing is the separation of useful minerals (concentrate) from useless minerals (tailings) in ore, and can be achieved by methods such as flotation, gravity separation, or magnetic separation. Useful minerals are minerals that correspond to the elements targeted for recovery. While the present invention can be applied to any mineral dressing method, this specification will mainly describe the case where the mineral dressing method is flotation.

[0018] Flotation is one of the most widely used ore separation methods in the mineral processing industry. It separates useful and unusable mineral particles based on the hydrophobicity of the ore surface. For example, as shown in FIG. 1 , the ore feed before separation is fed into a crusher 100 for crushing, and the crushed ore particles are then fed into a flotation machine 200. Specifically, the finely crushed ore particles are fed into the flotation machine 200 containing a solvent such as water along with a foaming agent. The air bubbles generated by stirring cause the hydrophobic useful mineral particles to adhere to the bubbles, which float the useful mineral particles and allow them to be collected. The floating mineral particles (mineral particles) resulting from flotation are called concentrate, and the precipitated mineral particles (mineral particles) are called tailings.

[0019] The recovery rate of concentrate from mined ore and the grade, which is expressed as the mass percentage of target elements contained in the recovered concentrate, vary depending on the mine. In order to increase the economic profit of mining, it is necessary to increase the recovery rate and grade. Flotation involves operational variables such as crushing time, flotation time, and pH. By determining the operational variables that lead to the optimal grade and recovery rate, the recovery rate and grade can be increased.

[0020] Conventionally, these manipulated variables have been determined by repeatedly conducting flotation tests and correcting the manipulated variables empirically. When the manipulated variables are determined empirically, the manipulated variable values ​​must be set, the test results (i.e., the grade and recovery rate of the target element) must be confirmed, and the manipulated variable values ​​must be reset repeatedly, resulting in a large cost in terms of time and money.

[0021] Therefore, in recent years, research has been conducted to reduce the number of tests required to determine the manipulated variables by using machine learning-based predictive models. Machine learning is a computational method that uses past information such as experience and data to improve accuracy in a given prediction task, and in recent years, predictive models have been developed using machine learning to predict the results of flotation tests.

[0022] Conventional prediction models output the recovery probability of a particle when feature data indicating the particle's characteristics are input. When the ore-dressing method is flotation, the recovery probability is the probability that the mineral particle will float (i.e., the probability that it will become a concentrate). Conventional prediction models were created under the assumption that interactions between particles do not affect the recovery probability.

[0023] However, it is believed that inter-particle interactions during flotation have a significant impact on the ore dressing results. Therefore, in the prediction method according to this embodiment, processing is performed that takes into account the impact of inter-particle interactions during flotation, thereby improving the prediction accuracy of the recovery probability. Specifically, in the prediction method according to this embodiment, when creating a prediction model, in addition to the characteristic data of individual mineral particles, characteristic data indicating the characteristic quantities of particle groups containing the mineral particles are also used as training data. Furthermore, when predicting the recovery probability using this prediction model, in addition to the characteristic data of individual mineral particles, characteristic data indicating the characteristic quantities of particle groups containing the mineral particles are input into the prediction model. The prediction model is created by machine learning selected from, for example, random forests, neural networks, and logistic regression analysis.

[0024] [Summary of Feature Data] Figure 2 is a diagram for explaining an overview of feature data. Figure 2(a) is an image of a cross section of a mineral for which feature data is to be created, taken with an electron microscope. Each of the various shaped regions shown in the image shown in Figure 2(a) corresponds to a mineral particle (e.g., particle P1, particle P2). The multiple particles included in the region surrounded by the outer quadrangle in Figure 2(a) constitute a particle group PG.

[0025] Fig. 2(b) is a diagram showing an outline of training data for creating a prediction model. The training data shown in Fig. 2(b) corresponds to the particle group PG shown in Fig. 2(a) and is composed of feature data of each particle from particle P1 to particle Pm (m is an integer) included in the particle group PG.

[0026] 2(b) includes individual particle characteristic data indicating the characteristic amounts of individual particles P corresponding to each of the multiple particles P included in the particle group PG, particle group characteristic data indicating the characteristic amounts of the particle group PG, and a label indicating whether the particle P is a concentrate or a tailing. Since concentrates and tailings contain a large number of particle groups PG, multiple sets of teacher data corresponding to the multiple particle groups PG are used as the teacher data. As a result, multiple combinations of individual particle characteristic data, particle group characteristic data, and labels are used as the teacher data.

[0027] As the feature data, for example, MLA data showing multiple features of particles identified by a mineral particle analyzer is used, but the feature data may be created by other means other than MLA data. When creating the training data, first, random sampling is performed from multiple MLA data corresponding to multiple particles so that the particle number ratio of concentrate to tailings is 1:1. Next, bootstrap is performed multiple times on the randomly sampled particles to create multiple data sets, and individual particle feature data for each of the multiple particles included in each of the multiple data sets obtained by bootstrap is identified. Next, particle group feature data corresponding to each of the multiple particles is created by calculating the feature amount of the particle group containing each of the multiple particles. Then, the training data is created by adding a label indicating whether each of the multiple particles is concentrate or tailings to the set of individual feature data and particle group feature data corresponding to each of the multiple particles.

[0028] FIG. 3 is a diagram showing an example of training data. In the training data shown in FIG. 3, a combination of a particle group number and a particle number is associated with a feature amount related to an individual particle, a feature amount related to a particle group, and a label. The feature amounts related to individual particles include a feature amount related to particle shape and a feature amount related to particle composition. The feature amount related to particle shape of an individual particle is, for example, a particle diameter and a cross-sectional area of ​​the particle. The feature amount related to particle composition is, for example, the content ratio of a mineral such as pyrite or chalcopyrite. As the feature amount related to individual particles, feature amounts corresponding to, for example, 20 types of features are used, and as the feature amount related to particle composition, feature amounts corresponding to, for example, 90 types of features, which is greater than the feature amounts related to individual particles, are used. Note that the number of feature amounts is arbitrary, and some feature amounts may be selected from a plurality of available feature amounts to reduce the number of each feature amount.

[0029] The feature quantity of a particle group is a value indicating a collective characteristic that is highly related to inter-particle interactions, such as the average diameter of the particles in the particle group or the density of the particles in the particle group. The label is a value indicating whether the particle group is a concentrate or a tailing.

[0030] FIG. 4 is a diagram showing an overview of prediction data used when predicting recovery probability using the prediction model M. The prediction data is data in which individual particle characteristic data and particle group characteristic data are associated, but differs from the training data shown in FIG. 2( b) in that it does not include labels. The prediction data is created, for example, based on MLA data of the ore supply. The prediction data shown in FIG. 4 is composed of characteristic data corresponding to particles different from those corresponding to the characteristic data included in the training data, so in FIG. 4, the particle names are R1 to Rn (n is an integer). When such prediction data is input into the prediction model, the prediction model outputs a recovery probability.

[0031] Generally, there is more tailings than concentrate, and the ratio of the ore distributed from the feed to the concentrate varies depending on the conditions of the ore dressing (e.g., the conditions of flotation), the specifications of the equipment used in the ore dressing (e.g., the flotation machine 200), the implementation conditions defined by the distribution of feature quantities of the ore feed, etc. Therefore, as described above, when training data is created by performing bootstrap, which samples MLA data so that the particle number ratio of the concentrate to the tailings is 1:1, an error may occur between the recovery probability output by the prediction model and the actual recovery probability under the implementation conditions.

[0032] Therefore, in the prediction method according to this embodiment, a prior probability indicating the ratio of concentrate to tailings corresponding to the conditions for performing ore dressing is specified, and the recovery probability output by the prediction model is corrected using the prior probability corresponding to the conditions for performing ore dressing. This makes it possible to predict the recovery probability with high accuracy for each of the performing conditions. The details of the correction process using the prior probability will be described later.

[0033] [Configuration of Prediction Device 1] Fig. 5 is a diagram showing the configuration of the prediction device 1. The prediction device 1 includes an external IF unit 11, a storage unit 12, and a control unit 13. The control unit 13 is, for example, a CPU (Central Processing Unit), and functions as a condition data acquisition unit 131, a prior probability data acquisition unit 132, an identification unit 133, a feature data acquisition unit 134, a learning unit 135, a prediction unit 136, and a calculation unit 137 by executing a program stored in the storage unit 12. The control unit 13 also functions as a prediction model M by executing a program stored in the storage unit 12. The prediction model M may be provided outside the control unit 13 (e.g., in another computer).

[0034] The external IF unit 11 has at least one of a user interface such as a keyboard and mouse for a user of the prediction device 1 to input data, a display for displaying data, and a communication interface for transmitting and receiving data to and from an external device. The prediction device 1 acquires data from a user or an external device via the external IF unit 11, and outputs the results of calculations of the recovery probability of concentrate particles, the recovery rate and grade of concentrate.

[0035] The storage unit 12 has storage media such as a read-only memory (ROM), a random access memory (RAM), and a solid-state drive (SSD). The storage unit 12 stores programs executed by the control unit 13. The storage unit 12 also stores various data used by the control unit 13 to calculate the recovery probability, recovery rate, and quality. The storage unit 12 may also store weights of the neural network included in the prediction model M.

[0036] The condition data acquisition unit 131 acquires condition data indicating the conditions for performing ore dressing. The condition data includes, for example, operation variables when performing ore dressing, the model name of the flotation machine 200, or the distribution of feature quantities of the ore feed. The condition data acquisition unit 131 acquires the condition data, for example, by displaying a screen for inputting the condition data on a display. The condition data acquisition unit 131 may acquire, from an external device, condition data input by a user in the external device. The condition data acquisition unit 131 acquires, for example, first condition data indicating the conditions for performing ore dressing and second condition data indicating the conditions under which the training data was obtained. The condition data acquisition unit 131 inputs the acquired condition data to the identification unit 133.

[0037] The condition data acquisition unit 131 may acquire data indicating target elements to be subjected to mineral dressing. The data indicating the target elements is, for example, text data indicating elements such as Cu, Fe, etc. The condition data acquisition unit 131 inputs the data indicating the target elements to the feature data acquisition unit 134. The data indicating the target elements may be stored in advance in the storage unit 12.

[0038] The prior probability data acquisition unit 132 acquires prior probability data indicating the prior probability of the concentrate ratio or the tailing ratio corresponding to each of a plurality of conditions for performing ore dressing. That is, the prior probability data acquisition unit 132 acquires a plurality of prior probability data associated with each of the conditions.

[0039] Like the condition data acquiring unit 131, the prior probability data acquiring unit 132 may acquire the prior probability data by displaying a screen for inputting the prior probability data on a display, or may acquire prior probability data input by a user in an external device from the external device. The prior probability data acquiring unit 132 inputs the acquired prior probability data to the identifying unit 133. The prior probability data acquiring unit 132 may input the acquired prior probability data to the identifying unit 133 by storing the acquired prior probability data in the storage unit 12.

[0040] The identification unit 133 identifies an optimal prior probability corresponding to the prior probability data corresponding to the condition closest to the condition indicated by the first condition data, among the multiple prior probability data input from the prior probability data acquisition unit 132. The identification unit 133 selects the prior probability data corresponding to the condition closest to the condition for ore dressing indicated by the first condition data, for example, based on the distance between multiple vectors each consisting of multiple elements representing the conditions corresponding to each of the multiple prior probability data and a vector each consisting of multiple elements representing the conditions for ore dressing. That is, the identification unit 133 determines the optimal prior probability to be the prior probability indicated by the prior probability data corresponding to the vector closest to the vector each consisting of multiple elements representing the conditions indicated by the first condition data. The identification unit 133 notifies the prediction unit 136 of the identified optimal prior probability.

[0041] The vectors each consisting of a plurality of elements representing the conditions corresponding to each of the plurality of prior probability data are also called "conducted mineral dressing condition data" and correspond to the second condition data. The vectors each consisting of a plurality of elements representing the conditions for mineral dressing are also called "unconducted mineral dressing condition data" and correspond to the first condition data.

[0042] [Learning Phase] Next, we will explain the processing in the learning phase for creating the prediction model M. The feature data acquisition unit 134 acquires learning individual feature data indicating the features of mineral particles and learning particle group feature data indicating the features of particle groups.

[0043] Specifically, the feature data acquisition unit 134 acquires first training individual particle feature data indicating the features of training particles contained in the concentrate. The feature data acquisition unit 134 acquires second training individual particle feature data indicating the features of training particles contained in the tailings. The feature data acquisition unit 134 acquires first training particle group feature data indicating the features of a training particle group composed of a plurality of training particles including training particles, which is contained in the concentrate. The feature data acquisition unit 134 acquires second training particle group feature data indicating the features of a training particle group composed of a plurality of training particles including training particles, which is contained in the tailings.

[0044] The characteristic data acquisition unit 134 creates first teacher data in which a set of first learning individual particle characteristic data and first learning particle group characteristic data is labeled to indicate that the data is concentrate, and inputs the first teacher data to the learning unit 135. In addition, the characteristic data acquisition unit 134 inputs second teacher data in which a set of second learning individual particle characteristic data and second learning particle group characteristic data is labeled to indicate that the data is tailings to the learning unit 135.

[0045] In order to create a plurality of prediction models M corresponding to a plurality of conditions under which ore dressing may be performed, the feature data acquisition unit 134 may acquire training particle feature data and training particle group feature data in association with the conditions. That is, in the step of acquiring first training individual particle feature data, the feature data acquisition unit 134 may acquire a plurality of first training individual particle feature data corresponding to each of a plurality of conditions under which ore dressing is performed. In the step of acquiring second training individual particle feature data, the feature data acquisition unit 134 may acquire a plurality of second training individual particle feature data corresponding to each of a plurality of conditions under which ore dressing is performed.

[0046] In the step of acquiring first training particle group characteristic data, the characteristic data acquisition unit 134 may acquire a plurality of first training particle group characteristic data corresponding to each of a plurality of conditions for performing ore dressing. In the step of acquiring second training particle group characteristic data, the characteristic data acquisition unit 134 may acquire a plurality of second training particle group characteristic data corresponding to each of a plurality of conditions for performing ore dressing. The characteristic data acquisition unit 134 creates first teacher data and second teacher data for each condition and inputs the created first teacher data and second teacher data to the learning unit 135.

[0047] The learning unit 135 creates a prediction model M by performing machine learning using first teacher data in which a set of first learning individual particle characteristic data and first learning particle group characteristic data is labeled to indicate concentrate, and second teacher data in which a set of second learning individual particle characteristic data and second learning particle group characteristic data is labeled to indicate tailings. The learning unit 135 creates a prediction model M by inputting the first teacher data or the second teacher data into the prediction model M and updating the weights of the neural network included in the prediction model M.

[0048] The learning unit 135 may create a prediction model corresponding to each of a plurality of conditions by performing machine learning using first training data in which a set of first training individual particle characteristic data and first training particle group characteristic data is labeled to indicate concentrate, and second training data in which a set of second training individual particle characteristic data and second training particle group characteristic data is labeled to indicate tailings, for each condition under which ore dressing is performed. In this case, the learning unit 135 stores in the memory unit 12 neural network weights included in the plurality of prediction models M that differ for each condition, for example, in association with the conditions under which ore dressing is performed.

[0049] [Prediction Phase] After the learning phase, the prediction phase is executed. In the prediction phase, the prediction unit 136 predicts the recovery probability of mineral particles contained in the ore (feed) to be beneficiated, and the calculation unit 137 outputs information indicating the recovery rate and grade of the mineral particles after beneficiation. The recovery probability of the mineral particles to be beneficiated is assumed to be unknown.

[0050] In the prediction phase, the characteristic data acquisition unit 134 acquires individual particle characteristic data that indicate the characteristics of mineral particles to be beneficiated. The characteristic data acquisition unit 134 also acquires particle group characteristic data that indicate the characteristics of particle groups composed of multiple particles contained in the ore feed. The characteristic data acquisition unit 134 notifies the prediction unit 136 of the acquired individual particle characteristic data and particle group characteristic data.

[0051] The prediction unit 136 predicts the recovery probability for each of the multiple particles contained in the ore feed. The prediction unit 136 predicts the recovery probability of each mineral particle based on the value output by the prediction model M in response to inputting the individual particle characteristic data and particle group characteristic data into the prediction model M, which is created by machine learning using a set of training individual particle characteristic data and training particle group characteristic data labeled as concentrate or tailings as training data. The prediction unit 136 predicts the recovery probability of each particle by inputting the individual particle characteristic data and particle group characteristic data corresponding to the particle into the prediction model M for each particle type.

[0052] The prediction unit 136 may, for example, input some feature quantities selected by dimension reduction from the multiple feature quantities included in the individual particle feature data acquired by the feature data acquisition unit 134, and some feature quantities selected by dimension reduction from the multiple feature quantities included in the particle group feature data acquired by the feature data acquisition unit 134 into the prediction model M.

[0053] As described above, if individual particle characteristic data and particle group characteristic data for learning in which the ratio of concentrate to tailings is 1:1 are used when creating the prediction model M, the prediction model M will output a recovery probability that is higher than the recovery probability under the conditions when actually sorting the ore.

[0054] Therefore, in the step of predicting the recovery probability, the prediction unit 136 predicts the recovery probability of each mineral particle by correcting the value output by the prediction model M with the optimal prior probability calculated by the identification unit 133. Specifically, the prediction unit 136 corrects the recovery probability of each target particle by normalizing the recovery probability of one particle output by the prediction model M by a value obtained by dividing the optimal prior probability by the proportion of particles labeled as concentrate in the training data. The prediction unit 136 inputs the corrected recovery probability of each mineral particle to the calculation unit 137.

[0055] The prediction unit 136 may predict the recovery probability using a prediction model M corresponding to the conditions indicated by the first condition data. The prediction unit 136, for example, selects a prediction model M having the greatest similarity between the conditions indicated by the first condition data and the conditions indicated by the second condition data corresponding to the training individual particle characteristic data and training particle group characteristic data used for training by each of the multiple prediction models M, and inputs the individual particle characteristic data and particle group characteristic data into the selected prediction model M, thereby increasing the recovery probability output by the prediction model M.

[0056] The calculation unit 137 calculates the recovery rate of the element whose recovery rate is to be calculated and the grade of the element whose grade in the concentrate is to be calculated, based on the recovery probability of each particle notified by the prediction unit 136. Specifically, the calculation unit 137 calculates the recovery rate of the element whose recovery rate is to be calculated by ore dressing, by dividing the sum of the product of the recovery probability of each particle predicted by the prediction unit 136, the mass of each particle obtained by the MLA analysis, and the grade of the element in each particle whose recovery rate is to be calculated by the calculation unit 136, by the mass of each particle obtained by the MLA analysis and the grade of the element in each particle whose recovery rate is to be calculated by the calculation unit 136.

[0057] More specifically, when the number of particles of the target element for which the prediction unit 136 predicted the recovery probability is m and the corrected recovery probability for the j-th particle Xj is y(Xj), the calculation unit 137 calculates the recovery rate R by the following formula using the mass of Xj and the purity wt % of the target element in Xj. Note that the target element in the following description is the element for which the recovery rate is to be calculated.

[0058] Furthermore, in order to calculate the grade indicating the proportion of the target element in the concentrate obtained by ore dressing, the calculation unit 137 first multiplies the recovery probability of each of the multiple target particles predicted by the prediction unit 136 by the mass of each of the multiple target particles, and then adds them together.The calculation unit 137 then calculates the grade indicating the proportion of the target element in the concentrate obtained by ore dressing based on the ratio of the sum of the multiple recovery probability of each of the multiple target particles, the mass of each of the multiple target particles, and the mass percentage of the target particles contained in the feed ore to the sum.

[0059] Specifically, the calculation unit 137 calculates X ｊ The mass of ｊ When the mass % of the target element is wt, the quality Q is calculated by the following formula.

[0060] The calculation unit 137 outputs the calculated values ​​indicating the recovery rate and grade via the external IF unit 11. The calculation unit 137 may display the values ​​indicating the recovery rate and grade on a display, or may transmit them to an external computer.

[0061] [Processing Flow in Prediction Device 1] Figure 6 is a flowchart showing the processing flow of the learning phase in the prediction device 1. The feature data acquisition unit 134 acquires first training individual particle feature data corresponding to the concentrate (S11). The feature data acquisition unit 134 acquires first training particle group feature data corresponding to the concentrate (S12). The feature data acquisition unit 134 acquires second training individual particle feature data corresponding to the tailings (S13). The feature data acquisition unit 134 acquires second training particle group feature data corresponding to the tailings (S14). The order of the processing from S11 to S14 is arbitrary.

[0062] The feature data acquisition unit 134 creates first training data by labeling the first training individual particle feature data and the first training particle group feature data as being concentrates, and creates second training data by labeling the second training individual particle feature data and the second training particle group feature data as being tailings (S15). The learning unit 135 creates a prediction model M using the created first training data and second training data (S16).

[0063] 7 is a flowchart showing the flow of processing in the prediction phase in the prediction device 1. The condition data acquisition unit 131 acquires first condition data (S21) indicating conditions for performing ore dressing of a mineral containing mineral particles of a target element for which the recovery rate and grade are to be predicted. The identification unit 133 identifies a prior probability corresponding to the first condition data acquired by the condition data acquisition unit 131 (S22) and notifies the prediction unit 136 of the identified prior probability.

[0064] The characteristic data acquisition unit 134 acquires individual particle characteristic data for each mineral particle of the target element for which the recovery rate and grade are to be predicted (S23). The characteristic data acquisition unit 134 also acquires particle group characteristic data corresponding to the particles (S24). The characteristic data acquisition unit 134 inputs the individual particle characteristic data and the particle group characteristic data into the prediction model M (S25). The characteristic data acquisition unit 134 sequentially inputs the individual particle characteristic data and the particle group characteristic data corresponding to each of the multiple particles into the prediction model M.

[0065] Next, the prediction unit 136 corrects the recovery probability corresponding to each of the multiple mineral particles output by the prediction model M based on the prior probability notified by the identification unit 133 (S26). The calculation unit 137 calculates the recovery rate and grade of the mineral particles corresponding to the target element based on the corrected recovery probability for each mineral particle (S27). Note that the processes from S23 to S25 may be performed before performing the processes from S21 to S22.

[0066] <Experimental Results> [Ore Sample Collection] Twelve samples, each weighing 1 kg, were collected from copper ore extracted from the mine. All conditions except for the crushing time were the same, with four copper ore samples assigned to each of three crushing time conditions (15 minutes, 25 minutes, and 35 minutes). Twelve resin-embedded samples (three concentrates and nine tailings) were prepared from the concentrate and tailings, and MLA analysis was performed. The MLA analysis mode was set to XBSE, and data on more than 50,000 particles from each sample was obtained.

[0067] [Creating a Prediction Model] To verify the effectiveness of the prediction method according to this embodiment, we created multiple prediction models corresponding to conventional prediction methods and a prediction model corresponding to the prediction method according to this embodiment. 80,000 items (40,000 items from the concentrate and 40,000 items from the tailings) were randomly sampled from the MLA data of the concentrate and tailings. The collected MLA data contained 108 features, and 10 features were selected using the minimum redundant maximum ratio (MRMR) algorithm to minimize correlation between features and shorten training time. The training particle group feature data used to create a prediction model corresponding to the prediction method according to this embodiment were D50, D80, diameter standard deviation, average density, density standard deviation, average copper mass fraction, average iron mass fraction, and average sulfur mass fraction.

[0068] 8 is a table showing a list of prediction models used in the comparative experiments. Model name M15 is a prediction model created using training data of individual particle characteristic data of particles obtained by crushing a mineral for 15 minutes. Model name M35 is a prediction model created using training data of individual particle characteristic data of particles obtained by crushing a mineral for 35 minutes. Model name M-Mix is ​​a prediction model created using training data of individual particle characteristic data of multiple particles obtained by crushing a mineral for 15 minutes, 25 minutes, and 35 minutes.

[0069] The model named M-New is a prediction model M according to this embodiment, which was created using training data including particle group characteristic data as well as individual particle characteristic data of multiple particles obtained by crushing a mineral for 15 minutes, 25 minutes, and 35 minutes.

[0070] [Comparison Results] Figure 9 is a diagram showing the recovery rate and grade of copper obtained by flotation of mineral particles pulverized under conditions of a 15-minute pulverization time. Circles in Figure 9 represent true values ​​indicating the actually measured recovery rate and grade. Squares in Figure 9 represent the recovery rate and grade calculated using the M-New prediction model M according to this embodiment. The recovery rate and grade were calculated by correcting the values ​​output by the prediction model M using a prior probability (e.g., optimal prior probability). Because the pulverization time of the particles to be predicted was 15 minutes, the recovery rate and grade when the prediction model named M35 was used deviated significantly from the true value. The recovery rate and grade when the prediction models named M15 and M-Mix were used were closer to the true value than the values ​​corresponding to model M35, but the values ​​corresponding to model M-New were closer to the true value. In this way, the effectiveness of predicting recovery probability using particle group characteristic data was confirmed.

[0071] 10 is a diagram showing the results of predicting the grades of several types of minerals obtained by flotation of particles crushed for a crushing time of 15 minutes. The values ​​on the far left corresponding to each mineral are the true values, and the values ​​on the far right are the grade values ​​predicted using prediction model M with the model name M-New. Looking at the results for chalcopyrite and pyrite, the grade values ​​predicted using prediction model M with the model name M-New are almost equal to the true values, confirming the effectiveness of predicting recovery probability using particle group characteristic data.

[0072] FIG. 11 is a diagram showing the recovery rate and grade of copper obtained by flotation of particles pulverized under conditions of a 35-minute pulverization time. The circles in FIG. 11 represent true values ​​indicating the actually measured recovery rate and grade. The squares in FIG. 11 represent the recovery rate and grade calculated using the M-New prediction model M according to this embodiment. Because the pulverization time of the particles to be predicted is 35 minutes, the recovery rate and grade when the prediction model named M15 is used deviate significantly from the true value. The recovery rate and grade when the prediction models named M35 and M-Mix are used are closer to the true value than the value corresponding to model M15, but the value corresponding to model M-New is closer to the true value. This example also confirms the effectiveness of predicting recovery probability using particle group characteristic data.

[0073] 12 is a diagram showing the results of predicting the grades of several types of minerals obtained by flotation of particles crushed for a crushing time of 35 minutes. The value on the far left corresponding to each mineral is the true value, and the value on the far right is the grade value predicted using prediction model M with the model name M-New. Looking at the results for chalcopyrite and pyrite, the grade values ​​predicted using prediction model M with the model name M-New are almost equal to the true value, confirming the effectiveness of predicting recovery probability using particle group characteristic data.

[0074] [Effects of the Prediction Method According to the Present Embodiment] As described above, in the prediction method according to the present embodiment, a computer inputs individual particle characteristic data and particle group characteristic data into a prediction model created by machine learning using a set of training individual particle characteristic data and training particle group characteristic data labeled as concentrate or tailings as training data, and predicts the particle recovery probability based on the value output by the prediction model. By predicting the particle recovery probability using the particle group characteristic data in this way, the prediction accuracy of the recovery rate and grade of mineral particles obtained by ore dressing is improved.

[0075] The present invention has been described above using embodiments, but the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes are possible within the scope of the gist of the present invention. For example, all or part of the device can be configured by functionally or physically distributing or integrating in any unit. Furthermore, new embodiments resulting from any combination of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiments resulting from the combination also have the effects of the original embodiments.

[0076] For example, in the above description, the prediction device 1 creates a prediction model M and predicts the recovery rate and grade of the target mineral, but the prediction device 1 may not have the learning unit 135 and may use a pre-created prediction model M. Also, although an example has been given of the case where the ratio of the first training data corresponding to concentrate to the second training data corresponding to tailings is 1:1 in the training data, this ratio may also be arbitrary.

[0077] Furthermore, in the above description, an example has been given in which MLA data is used to create individual particle characteristic data and particle group characteristic data, but other data may also be used. In particular, a particle size distribution measuring device may be used to create particle group characteristic data.

[0078] DESCRIPTION OF SYMBOLS 1 Prediction device 11 External IF unit 12 Memory unit 13 Control unit 100 Crusher 131 Condition data acquisition unit 132 Prior probability data acquisition unit 133 Identification unit 134 Feature data acquisition unit 135 Learning unit 136 Prediction unit 137 Calculation unit 200 Flotation machine

Claims

1. A computer-implemented prediction method comprising the steps of: using characteristic data of mineral particles to create a prediction model by machine learning using the characteristic data as training data; and inputting the characteristic data into the prediction model and predicting the probability of recovering mineral particles through mineral dressing based on the value output by the prediction model, wherein the characteristic data of the mineral particles includes individual particle characteristic data that indicates the characteristics of the mineral particle, and particle group characteristic data that indicates the characteristics of a particle group made up of a plurality of the mineral particles.

2. The prediction method of claim 1, wherein the individual particle characteristic data and the particle group characteristic data used as the training data in the step of creating the prediction model include individual particle characteristic data and particle group characteristic data of mineral particles in the recovered concentrate and tailings.

3. The prediction method according to claim 1 or 2, wherein the individual particle characteristic data and the particle group characteristic data used in the step of predicting the recovery probability of the mineral particles include individual particle characteristic data and particle group characteristic data of mineral particles whose recovery probability is unknown.

4. The prediction method according to claim 1, further comprising the steps of: acquiring first condition data indicating the conditions under which ore dressing is performed; acquiring second condition data indicating the conditions under which the training data was obtained; acquiring prior probability data indicating a prior probability which is the ratio of the number of concentrate particles or the ratio of the number of tailing particles to the number of feed ore particles corresponding to each of a plurality of conditions under which ore dressing is performed; and identifying, in the prior probability data, an optimal prior probability corresponding to the condition closest to the condition indicated by the first condition data among the plurality of conditions indicated by the second condition data; wherein, in the step of predicting the recovery probability, the recovery probability is predicted by correcting the value output by the prediction model by the optimal prior probability.

5. The prediction method according to claim 1, further comprising a step of creating the prediction model by machine learning selected from random forest, neural network, and logistic regression analysis.

6. The prediction method according to claim 1, wherein in the step of predicting the recovery probability, a portion of feature quantities selected by dimension reduction from a plurality of feature quantities included in the acquired individual particle feature data and a portion of feature quantities selected by dimension reduction from a plurality of feature quantities included in the acquired particle group feature data are input to the prediction model.

7. A prediction program for causing a computer to execute the steps of: using characteristic data of mineral particles to create a prediction model by machine learning using the characteristic data as training data; and inputting the characteristic data into the prediction model and predicting the probability of recovering mineral particles through ore dressing based on a value output by the prediction model, wherein the characteristic data of the mineral particles includes individual particle characteristic data that indicates the characteristics of the mineral particle, and particle group characteristic data that indicates the characteristics of a particle group made up of a plurality of the mineral particles.

8. A prediction device having: a feature data acquisition unit that acquires feature data of mineral particles; and a prediction unit that inputs the feature data into a prediction model created by machine learning using the feature data as training data, and predicts the probability of recovering mineral particles through ore dressing based on a value output by the prediction model, wherein the feature data of the mineral particles includes individual particle feature data that indicates the features of the mineral particles, and particle group feature data that indicates the features of particle groups composed of a plurality of the mineral particles.

Images