Cement quality prediction method

Abstract

To provide a method with which quality of cement can be predicted with high accuracy in a short time.SOLUTION: A cement quality prediction method is for predicting quality of cement by using a second prediction model created by using Lasso regression, and in the method, when the second prediction model is created, a tuning parameter α is defined by (A) a step of changing the numerical values of the tuning parameter α to create a plurality of first prediction models by using Lasso regression for every changed numerical value of the tuning parameter α, (B) a step of calculating a determination coefficient for each of the plurality of first prediction models by using a predicted value obtained by using the first prediction model and an actual measured value of output date for a test, and (C) a step of defining a tuning parameter α used for the creation of the first prediction model having the largest numerical value of the determination coefficient as a tuning parameter α used for the creation of the second prediction model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for predicting cement quality. [Background technology]

[0002] An important quality parameter for cement is, for example, the compressive strength of mortar at 28 days of age, as determined in accordance with JIS R 5201 (Physical Testing Methods for Cement). However, quality parameters that depend on age, such as the compressive strength of mortar, require a long time to determine the quality test results, which makes it difficult to ship cement after the quality test results have been confirmed. For this reason, cement manufacturing sites set quality control items for the manufacturing process, such as the composition of cement clinker (chemical composition and mineral composition) and the fineness of cement, and set empirical control standard values ​​for these quality control items in the manufacturing process so that quality items that are factors related to material age meet the specified control standard values.

[0003] As a method for predicting cement quality or production conditions with high accuracy, for example, Patent Document 1 discloses a method for predicting cement quality or production conditions using a neural network having an input layer for inputting actual measured values ​​of monitoring data in cement production and an output layer for outputting estimated values ​​of evaluation data related to the evaluation of cement quality or production conditions, wherein the neural network is a trained neural network that has been trained in advance using a plurality of training data that are combinations of actual measured values ​​of the monitoring data and actual measured values ​​of the evaluation data, and while inputting the actual measured values ​​of the monitoring data into the input layer using the trained neural network and outputting estimated values ​​of the evaluation data from the output layer, the absolute value of the difference between the estimated value and the actual measured value in the evaluation data is calculated at regular intervals, and the absolute value is calculated as the mean square error (σ) between the estimated value of evaluation data obtained by inputting the actual measured values ​​of the monitoring data that constitute the training data into the input layer of the trained neural network and the actual measured values ​​of the evaluation data that constitute the training data. L) multiplied by 1.3 is equal to or greater than the value obtained by successively performing the above-mentioned periodic intervals, and the mean square error (σ L ) and the ratio (absolute value / mean square error (σ L) The present invention describes a method for predicting the quality or manufacturing conditions of cement, characterized in that if the value obtained by multiplying the sum of (a) and (b) by the periodic interval (days) is 12 (days) or more, the trained neural network is retrained and the resulting neural network is used as a new trained neural network. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-168021 Summary of the Invention [Problem to be solved by the invention]

[0005] When predicting cement quality using a predictive model created using machine learning such as a neural network, if there are many explanatory variables that have little effect on the target variable when creating the predictive model (e.g., a trained neural network), there is a problem that the accuracy of the prediction decreases and the analysis time becomes longer. Therefore, it is necessary to select appropriate explanatory variables according to the set objective variables, but there is a problem that it is difficult to select appropriate explanatory variables and that the selection takes time. An object of the present invention is to provide a method for predicting cement quality in a short time with high accuracy. [Means for solving the problem]

[0006] As a result of extensive research into solving the above-mentioned problems, the present inventors have discovered a method for predicting cement quality using a first prediction model and a second prediction model created by Lasso regression, wherein a tuning parameter α used in Lasso regression when creating the second prediction model is: (A) varying the value of the tuning parameter α and creating a plurality of first prediction models by Lasso regression for each of the varied values ​​of the tuning parameter α; and (B) calculating a coefficient of determination (R) for each of the plurality of first prediction models using the predicted value obtained using the first prediction model and the actual measured value of test output data. 2 ) and (C) the step of calculating the coefficient of determination (R 2 The present inventors have found that the above object can be achieved by a cement quality prediction method in which the tuning parameter α used in creating a first prediction model having the largest value of α ( ) is determined as the tuning parameter α used in creating a second prediction model, and have completed the present invention. That is, the present invention provides the following [1] to [5].

[0007] [1] A method for predicting cement quality using a first prediction model and a second prediction model, wherein the second prediction model is created by Lasso regression using a plurality of second learning data, which is a combination of actual measured values ​​of second learning input data and actual measured values ​​of second learning output data related to cement quality, and the tuning parameter α used in the Lasso regression is determined by the following steps (A) to (C): the actual measured values ​​of the prediction input data are input into the second prediction model, and predicted values ​​of prediction output data related to cement quality are output from the second prediction model, and the cement quality is predicted using the predicted values ​​of the prediction output data. (A) a first prediction model creation step of varying the value of the tuning parameter α within an arbitrarily determined range and interval, and creating the first prediction model by Lasso regression using a plurality of first learning data that are combinations of actual measured values ​​of first learning input data and actual measured values ​​of first learning output data related to cement quality for each of the varied values ​​of the tuning parameter α, thereby obtaining the first prediction models the same in number as the values ​​of the tuning parameter α; (B) For each of the plurality of first prediction models obtained in the step (A), a plurality of test data are used, which are combinations of actual measured values ​​of test input data and actual measured values ​​of test output data related to cement quality, and (b1) a predicted value of test output data related to cement quality obtained by inputting the actual measured values ​​of the test input data into the first prediction model, and (b2) a coefficient of determination (R 2 ) is calculated, the coefficient of determination calculation process (C) The multiple coefficients of determination (R 2 ) of the above coefficient of determination (R 2 a tuning parameter determination step of determining the value of the tuning parameter α used in creating the first prediction model having the largest value of (a) above as the value of the tuning parameter α used in creating the second prediction model;

[0008] [2] The first learning input data consists of multiple types of data, and the coefficient of determination (R 2 ) for the first prediction model having the largest value, calculates a standard partial regression coefficient for each type of data constituting the first learning input data used to create the first prediction model, and determines the type of data constituting the second learning input data based on the values ​​of the standard partial regression coefficients. [3] A cement quality prediction method according to [2], in which the types of data constituting the first learning input data, excluding the types of data for which the standard partial regression coefficient has a value of 0, are defined as the types of data constituting the second learning input data. [4] A method for predicting cement quality according to any one of [1] to [3], wherein in step (A), the arbitrarily determined numerical range is a numerical range that is arbitrarily determined within a numerical range of more than 0 and not more than 10, and the arbitrarily determined numerical interval is a numerical interval that is arbitrarily determined within a numerical range of 0.0001 to 0.01. [5] The cement quality prediction method according to any one of [1] to [4], wherein the first learning input data, the second learning input data, the test input data, and the prediction input data are one or more types of data selected from the group consisting of data on cement, data on materials other than cement that constitute a cement composition obtained by kneading cement and water, data on the blending conditions of the cement composition, data on the means and method for mixing the cement composition, data on the environment when mixing the cement composition, and data on the transportation of the cement composition, and the first learning output data, the second learning output data, the test output data, and the prediction output data are data on the quality of the cement composition. [Effects of the Invention]

[0009] According to the present invention, the quality of cement can be predicted in a short time and with high accuracy. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 10 is a diagram showing the relationship between the tuning parameter α and the coefficient of determination (R2) in the first embodiment. [Figure 2] FIG. 1 is a diagram showing the relationship between the actually measured and predicted compressive strength values ​​at an age of 28 days in Example 1. [Figure 3] FIG. 10 is a diagram showing the relationship between the tuning parameter α and the coefficient of determination (R2) in the second embodiment. [Figure 4] FIG. 10 is a diagram showing the relationship between the actually measured value and the predicted value of the mortar flow value in Example 2. DETAILED DESCRIPTION OF THE INVENTION

[0011] The cement quality prediction method of the present invention is a method for predicting cement quality using a first prediction model and a second prediction model, wherein the second prediction model is created by Lasso regression using a plurality of second learning data that is a combination of actual measured values ​​of second learning input data and actual measured values ​​of second learning output data related to cement quality, and the tuning parameter α used in the Lasso regression is determined by the following steps (A) to (C), in which the actual measured values ​​of the prediction input data are input into the second prediction model, and predicted values ​​of prediction output data related to cement quality are output from the second prediction model, and the cement quality is predicted using the predicted value of the prediction output data. (A) a first prediction model creation step of varying the value of the tuning parameter α within an arbitrarily determined range and interval, and creating a first prediction model by Lasso regression using a plurality of first learning data that are combinations of actual measured values ​​of the first learning input data and actual measured values ​​of the first learning output data related to the quality of cement, for each value of the varied tuning parameter α, to obtain a first prediction model the same in number as the value of the tuning parameter α; (B) For each of the plurality of first prediction models obtained in step (A), a plurality of test data are used, which are combinations of actual measured values ​​of test input data and actual measured values ​​of test output data related to cement quality. (b1) A coefficient of determination (R 2 ) is calculated, the coefficient of determination calculation process (C) Multiple coefficients of determination (R 2 ), the coefficient of determination (R 2 a tuning parameter determination step of determining the value of the tuning parameter α used in creating the first prediction model having the largest value of (a) as the value of the tuning parameter α used in creating the second prediction model;

[0012] In the present invention, the second prediction model is created by Lasso regression using multiple second training data, which are combinations of actual measured values ​​of second training input data and actual measured values ​​of second training output data related to cement quality. Lasso regression is a regression method that incorporates a regularization term (the sum of the absolute values ​​of the standard partial regression coefficients) into the loss function of linear regression, and is expressed by the following formula (1) using a tuning parameter α (0<α). In the following formula (1), N is the number of output data for learning (objective variables), y 1i is the actual measurement value of the training output data, y 2i is the predicted value of the training output data, α is the tuning parameter, K is the number of training input data (explanatory variables), W j indicates the standard partial regression coefficient.

[0013]

number

[0014] Examples of the second learning input data include data on cement, data on materials other than cement that constitute a cement composition obtained by kneading cement and water, data on the blending conditions of the cement composition, data on the means and method for mixing the cement composition, data on the environment during mixing of the cement composition, and data on transportation of the cement composition. These may be used alone or in combination of two or more. Examples of data related to cement include data related to cement as a whole, data related to raw materials for cement clinker, data related to burning conditions for cement clinker, data related to grinding conditions for cement, and data related to cement clinker.

[0015] Examples of data on the entire cement include: (i) general cement data such as type, chemical composition (weight loss, percentages of SiO2, Al2O3, Fe2O3, CaO, MgO, SO3, Na2O, K2O, Na2Oeq, TiO2, P2O5, MnO, Cl, and Sr (mass%)), cement modulus (silicic acid percentage (SM), iron percentage (IM), hydraulic hardness percentage (HM)), cement fineness (Blaine specific surface area, sieve residue), X-ray diffraction (XRD) ) or calculated using the Bogue formula (the proportions (by mass) of C3S, C2S, C3A, C4AF, f.MgO, f.CaO, gypsum dihydrate, gypsum hemihydrate, and CaCO3, etc.), f.CaO (amount of free lime), particle size distribution, color (a, b, L), (ii) mineralogical properties and crystallographic properties of each mineral contained in the cement, (iii) hemihydration rate of gypsum contained in the cement, shipping temperature of the cement, etc.

[0016] Examples of data on raw materials for cement clinker include: (i) data on the raw materials for cement clinker, such as chemical composition, hydraulic hardness calculated from the chemical composition, sieve residue, Blaine specific surface area (fineness), ignition loss, supply amount, supply amount of auxiliary materials (special raw materials such as waste), amount stored in the blending silo (remaining amount), amount stored in the storage silo (remaining amount), (ii) current value of the cyclone located between the raw material mill and the blending silo for the mixed raw materials (representing the rotation speed of the cyclone, which is correlated with the speed of the raw materials passing through the cyclone), (iii) the time from when the raw materials were fed into the kiln, (iv) Data on raw materials for cement clinker, such as the chemical composition, hydraulic hardness calculated from the chemical composition, of the raw materials for cement clinker (mixed raw materials for cement clinker from which fine particles and the like have been removed by a countercurrent airflow during transportation; hereinafter referred to as the raw materials for cement clinker to be kilned) at a point in time a predetermined time ago (for example, one point in time 5 hours ago, or multiple points in time such as four points in time 3 hours, 4 hours, 5 hours, and 6 hours ago), and the like.

[0017] Examples of data on cement burning conditions include: (i) data on the burning of cement clinker, such as the amount of cement clinker raw material inserted into the kiln, kiln rotation speed, outlet temperature, burning zone temperature, cement clinker temperature, kiln average torque, O2 concentration, NO X (ii) clinker cooler temperature; and (iii) preheater gas flow rate (which is correlated with the preheater temperature). Examples of data related to cement grinding conditions include grinding temperature, amount of water sprayed in the finishing mill, separator air volume, type of gypsum, amount of gypsum added, type of grinding aid, amount of grinding aid added, amount of cement clinker added, finishing mill rotation speed, temperature of powder discharged from the finishing mill, amount of powder discharged from the finishing mill, amount of powder not discharged from the finishing mill, etc. Examples of data on cement clinker include (i) mineral composition, chemical composition, f.CaO (free lime), and volumetric weight, (ii) crystallographic properties (lattice constant, crystallite size, etc.) of each mineral contained in the cement clinker, and (iii) ratios of two or more minerals contained in the cement clinker. These may be used alone or in combination of two or more.

[0018] Examples of data on materials other than cement that constitute a cement composition (e.g., mortar, concrete, etc.) obtained by kneading cement and water include (i) aggregate (fine aggregate or coarse aggregate) data such as type, density, water absorption rate, water content, surface water content, particle size distribution, maximum size, and particle shape, (ii) type of admixture, and (iii) type of admixture. These may be used alone or in combination of two or more. Examples of data on the blending conditions of the cement composition include blending ratios of cement, fine aggregate, coarse aggregate, water, various admixtures (AE agents, water-reducing agents, AE water-reducing agents, high-performance water-reducing agents, high-performance AE water-reducing agents, superplasticizers, setting retarders, etc.), and various admixtures (ground granulated blast furnace slag, silica fume, fly ash, etc.) blended in the cement composition (for example, the amount (mass%) of admixture relative to 100 mass% of cement), and items in the design blending table, such as the water-cement ratio, air content, fine aggregate ratio, and unit water content (per 1 m of concrete). 3 These may be used alone or in combination of two or more.

[0019] Examples of data relating to the means and method for mixing the cement composition include the type, model, and capacity of the mixer, the amount of materials to be mixed, the order in which the materials are mixed, and the mixing time. These may be used alone or in combination of two or more. Examples of data relating to the environment during mixing of the cement composition include temperature (outside air temperature, temperature inside the mixer, temperature of the concrete), temperature of the mixing water, humidity, production date, production time, etc. These may be used alone or in combination of two or more. Examples of the data related to the transportation of the cement composition include the power load value inside the drum of the truck agitator, the drum capacity of the truck agitator, the transported capacity, transported mass and temperature of the cement composition, the outside air temperature during transportation, the transportation time (the time from the end of mixing to the end of transportation (unloading)), the transportation distance, the date of transportation, the time of transportation, etc. These may be used alone or in combination of two or more.

[0020] An example of the second learning output data is data on the quality of a composition obtained by kneading cement and water. Examples of data relating to the quality of cement compositions obtained by kneading cement and water include strength (compressive strength of mortar, compressive strength of concrete, flexural strength, etc.), fluidity (slump, mortar flow), heat of hydration, setting time, drying shrinkage, stability, expansion in water, sulfate resistance, carbonation, ASR resistance, air content, chloride content, crack resistance, dynamic modulus of elasticity, dynamic shear modulus of elasticity, dynamic Poisson's ratio, void volume in the hardened body, pore size distribution, durability, and color tone. These may be used alone or in combination of two or more.

[0021] The "actually measured values ​​of second learning input data" and "actually measured values ​​of second learning output data" used as second learning data are data obtained when a cement composition is actually produced by kneading cement such as mortar with water as a learning data sample. Furthermore, the "actually measured values" include not only the numerical values ​​of specific data that are actually measured, but also non-numeric data (e.g., the type of cement, etc.). The number of samples for the second learning data varies depending on the types of the second learning input data and the second learning output data. However, from the viewpoint of being able to predict quality with higher accuracy, it is preferably 50 or more, more preferably 80 or more, and particularly preferably The upper limit of the number of samples is not particularly limited, but is preferably 1,000, and more preferably 800, from the viewpoint of ease of preparing data. The various types of data used in the second training data may be the same types of data as those used in the first training data described below, or may be data selected from the data used in the first training data (details will be described later).

[0022] A second prediction model (final prediction model) can be created by Lasso regression using multiple second learning data, which are combinations of actual measured values ​​of second learning input data and actual measured values ​​of second learning output data related to cement quality. The tuning parameter α used in the Lasso regression when creating the second prediction model is determined by the following steps (A) to (C). Each step will be explained in detail below.

[0023] [(A): First prediction model creation process] This process involves varying the value of the tuning parameter α within an arbitrarily determined range and interval, and for each value of the tuning parameter α, creating a first prediction model by Lasso regression using multiple first learning data, which are combinations of actual measured values ​​of the first learning input data and actual measured values ​​of the first learning output data related to the quality of cement, to obtain a number of first prediction models equal to the number of values ​​of the tuning parameter α. The arbitrarily determined numerical range is the upper and lower limits of the tuning parameter α when the value of α is changed. By determining the above numerical range, it is possible to prevent excessive generation of prediction models. The above-mentioned numerical range may be arbitrarily determined. The numerical range may vary depending on the type of training data, but may be, for example, a numerical range that is arbitrarily determined within a range exceeding 0 and not exceeding 10.

[0024] The arbitrarily determined numerical interval is preferably a numerical interval determined arbitrarily within a numerical range of 0.0001 to 0.01 (more preferably 0.0005 to 0.005, and particularly preferably 0.001). If the numerical interval is 0.0001 or more, it is possible to prevent an excessive number of first prediction models from being created. If the numerical interval is 0.01 or less, it is possible to increase the number of first prediction models to be created and obtain a more appropriate value for the tuning parameter α.

[0025] In this process, for each value of the tuning parameter α that is changed, a first prediction model is created by Lasso regression using multiple first learning data that are combinations of actual measured values ​​of the first learning input data and actual measured values ​​of the first learning output data related to the quality of cement. Examples of the first training input data and the first training output data related to cement quality are the same as the examples of the second training input data and the second training output data related to cement quality described above, respectively. The number of samples of the first training data is the same as the number of samples for the second training data described above. The various data used as the first learning data may be the same as or different from the second learning data described above.

[0026] For example, if the value of the tuning parameter α to be changed is set within a range of 0.001 to 1.000 and the tuning parameter α is changed in numerical intervals of 0.001 (i.e., if the value of α is changed from 0.001, 0.002, 0.003, and so on to 1.000), the number of first prediction models created for each changed value of the tuning parameter α will be 1,000, which is the same as the number of changes in the value of the tuning parameter α (1,000).

[0027] [(B): Determination coefficient calculation process] In this step, for each of the plurality of first prediction models obtained in step (A), a coefficient of determination (R 2 ) is calculated. The test input data and test output data are the same as the first training input data and the first training output data related to cement quality that were used when creating the first prediction model, respectively. The number of test data samples is not particularly limited, but is preferably 10 to 50%, more preferably 20 to 40%, of the number of samples for the second training data. The test data may be the same data as that used as the training data (first training data and second training data), or may be data obtained from a separately prepared sample.

[0028] [(C): Tuning parameter determination process] This step is carried out by using the multiple coefficients of determination (R 2 ), the coefficient of determination (R 2 The step of determining the value of the tuning parameter α used to create the first prediction model with the largest value of (α=0.01) as the value of the tuning parameter α used to create the second prediction model. In this step, by setting the value of the tuning parameter α used to create the second prediction model to a specific value, the prediction accuracy of the second prediction model (final prediction model) can be further improved.

[0029] In addition, the coefficient of determination (R 2 For the first prediction model having the largest value of (i.e., the standardized partial regression coefficient for each type of data constituting the first learning input data, which is made up of multiple types of data and was used to create the first prediction model, may be calculated, and the type of data constituting the second learning input data may be determined based on the values ​​of the standardized partial regression coefficients. Data for which the calculated standard partial regression coefficient is close to 0 is considered to be data with a low correlation with the prediction output data (objective variable) related to cement quality. By excluding such data from the data that constitutes the second learning input data, the prediction accuracy of the second prediction model can be further improved. For example, by excluding data constituting the first learning input data whose standard partial regression coefficient is, for example, -0.3 to 0.3 (preferably -0.1 to 0.1, more preferably 0) from the data constituting the second learning input data, a second prediction model with higher prediction accuracy can be created.

[0030] After determining the tuning parameter α used in the Lasso regression through steps (A) to (C), a second prediction model is created by Lasso regression using multiple second learning data, which are combinations of actual measured values ​​of the second learning input data and actual measured values ​​of the second learning output data related to cement quality. Next, the actual measured values ​​of the prediction input data (actual measured values ​​obtained from the cement, etc., whose quality is to be predicted) are input into a second prediction model, and the predicted values ​​of the prediction output data regarding the quality of the cement are output from the second prediction model, and the quality of the cement can be predicted using the predicted values ​​of the prediction output data. The types of the prediction input data and the prediction output data related to cement quality are the same as the types of the second learning input data and the learning output data related to cement quality, respectively. Furthermore, from the viewpoint of being able to predict quality with higher accuracy, it is preferable that the first learning data, the second learning data, and the prediction input data are obtained in the same factory (preferably the same production line). Furthermore, from the viewpoint of being able to predict quality with higher accuracy, it is preferable that the first training data and the second training data are data obtained within one to three years from the time the input data for prediction is obtained. [Example]

[0031] The present invention will be specifically described below with reference to examples, but the present invention is not limited to these examples. [Example 1] The input data for training and testing included the chemical composition of cement measured using X-ray fluorescence analysis (XRF) (i.g., SiO2, Al2O3, Fe2O3, CaO, MgO, SO3, Na2O, K2O, Na2Oeq, TiO2, P2O5, MnO, and Sr proportions (mass%)), the chemical composition of cement measured using potentiometric titration (Cl proportion (mass%)), the cement modulus (silicic acid percentage (SM), iron percentage (IM), hydraulic hardness percentage (HM)), and the fineness of cement (Blaine specific surface area (cm 2A total of 39 types of data were used: the percentage of 850 μm sieve residue (mass%), the mineral composition of the cement measured using the Rietveld method by X-ray diffraction (XRD) (the percentages (mass%) of C3S, C2S, C3A, C4AF, f.MgO, f.CaO, gypsum dihydrate, gypsum hemihydrate, and CaCO3, and the gypsum hemihydration rate), the mineral composition calculated using the Bogue formula (the percentages (mass%) of C3S, C2S, C3A, and C4AF), the percentage (mass%) of wet f.CaO, color (a, b, L), and shipping temperature (°C). In addition, the actual measured values ​​of compressive strength of mortar containing the above cement at an age of 28 days were used as learning output data regarding cement quality and test output data regarding cement quality. The measured values ​​for the various data above were obtained from ordinary Portland cement manufactured at the same factory between January 1, 2013 and June 1, 2020. In addition, the number of pieces of data used as learning data (combinations of actual measured values ​​of the learning input data and the learning output data) and test data (combinations of actual measured values ​​of the test input data and the test output data) was 151.

[0032] Of the 151 data, 113 data, equivalent to 75%, were used as training data (first training data) to create multiple first prediction models by Lasso regression. The tuning parameter α used in the Lasso regression was within the numerical range of 0.001 to 1.000, and was changed in numerical intervals of 0.001 (i.e., the numerical value of the above α was changed from 0.001, 0.002, 0.003, ... 1.000). A first prediction model was created for each changed value of the tuning parameter α, resulting in a total of 1,000 first prediction models.

[0033] Of the 151 data, 38 data, equivalent to 25%, were used as test data, and the coefficient of determination (R 2 ) was calculated. Specifically, the measured values ​​of 38 test input data were input into the first prediction model, and the predicted values ​​of test output data related to cement quality were output from the first prediction model. Next, the coefficient of determination (R 2 ) was calculated. The value of the tuning parameter α used in creating the first prediction model and the coefficient of determination (R 2 ) is shown in Figure 1. 1,000 coefficients of determination (R 2 ) with the largest coefficient of determination (R 2 The value of the tuning parameter α used to create the first prediction model from which the calculated value of α was calculated was 0.057. This value (0.057) was set as the value of the tuning parameter α used to create the second prediction model.

[0034] For the first prediction model obtained when the tuning parameter α was set to 0.057, the standard partial regression coefficient was calculated for each of the training input data (36 types) used to create the first prediction model, and 24 types of the training data whose standard partial regression coefficients were not 0 were designated as the training input data (second training input data) to be used to create the second prediction model. The above 24 types of data are the chemical composition of cement measured using X-ray fluorescence analysis (XRF) (i.e., the proportions (mass%) of SiO2, Fe2O3, SO3, K2O, Na2Oeq, TiO2, P2O5, MnO, and Sr), the chemical composition of cement measured using potentiometric titration (i.e., the proportion (mass%) of Cl), the modulus of cement (i.e., the silicate content (SM), iron content (IM), and hydraulic content (HM)), the fineness of cement (i.e., the Blaine specific surface area (cm)), and the chemical composition of cement measured using potentiometric titration (i.e., the proportion (mass%) of Cl). 2 / g), percentage of 850 μm sieve residue (mass %), mineral composition of cement measured using the Rietveld method by X-ray diffraction (XRD) (percentages of C3S, C3A, f.MgO, f.CaO, gypsum dihydrate, gypsum hemihydrate, and CaCO3 (mass %)), mineral composition calculated using the Bogue formula (percentage of C2S (mass %)), percentage of wet f.CaO (mass %), color (b), and shipping temperature (°C).

[0035] The tuning parameter α was set to 0.057, and a second prediction model was created by Lasso regression using the above 24 types of data as learning input data. The second predictive model took less than a minute to develop. The measured values ​​of the 115 learning input data were input to the second prediction model, and the predicted values ​​of the test output data related to the quality of 115 cements were output from the second prediction model. Next, the coefficient of determination (R 2 ) was calculated to be 0.57. In addition to the above-mentioned training data and test data, 38 prediction data (combinations of actual values ​​of prediction input data and actual values ​​of prediction output data) were prepared, which were the actual measured values ​​of the above 39 types of data (actual measured values ​​of prediction input data) obtained from cement manufactured between June 5, 2020 and January 7, 2021, and the actual measured values ​​of the compressive strength of mortar containing the above cement at an age of 28 days (actual measured values ​​of prediction output data). The 38 actual values ​​of the prediction input data were input into the second prediction model, and the second prediction model output the predicted values ​​of the prediction output data related to the quality of the 38 cements. Next, the coefficient of determination (R 2 ) was calculated to be 0.60. The coefficient of determination when using prediction data (0.60) is similar to the coefficient of determination when using training data (0.57), indicating that mortar flow values ​​can be predicted with excellent accuracy even when using data that was not used to train the prediction model. Figure 2 shows the relationship between the measured and predicted strength of mortar at 28 days of age.

[0036] [Example 2] As the learning input data and the test input data, the actual measured values ​​of the 39 types of data used in Example 1 were used. Furthermore, the actual measured values ​​of the mortar flow value of the mortar containing the above cement were used as learning output data regarding the quality of cement and test output data regarding the quality of cement. The measured values ​​for the various data above were obtained from ordinary Portland cement manufactured between January 1, 2013 and June 1, 2020. The number of data used as training data and test data was 151.

[0037] Of the 151 data, 113 data, equivalent to 75%, were used as training data (first training data) to create multiple first prediction models by Lasso regression. The tuning parameter α used in the Lasso regression was within the numerical range of 0.001 to 6.000, and was changed in numerical intervals of 0.001 (i.e., the numerical value of the above α was changed from 0.001, 0.002, 0.003, to 6.000), and a first prediction model was created for each changed value of the tuning parameter α, resulting in a total of 6,000 first prediction models.

[0038] Of the 151 data, 38 data, equivalent to 25%, were used as test data, and the coefficient of determination (R 2 ) was calculated. The value of the tuning parameter α used to create the first prediction model and the coefficient of determination (R 2 ) is shown in Figure 3. 6,000 coefficients of determination (R 2 ) with the largest coefficient of determination (R 2 The value of the tuning parameter α used to create the first prediction model from which the calculated value of α was calculated was 2.468. This value (2.468) was set as the value of the tuning parameter α used to create the second prediction model.

[0039] For the first prediction model obtained when the tuning parameter α was set to 2.468, the standard partial regression coefficient was calculated for each of the training input data (36 types) used to create the first prediction model, and 19 types of the training data whose standard partial regression coefficients were not 0 were designated as the training input data (second training input data) to be used to create the second prediction model. The above 19 types of data are the chemical composition of cement measured using X-ray fluorescence analysis (XRF) (the proportions (mass%) of CaO, KO, NaOeq, TiO2, PO5, and Sr), the chemical composition of cement measured using potentiometric titration (the proportion (mass%) of Cl), the modulus of cement (iron fraction (IM)), the fineness of cement (Blaine specific surface area (cm)), and the like. 2 / g), the mineral composition of the cement measured using the Rietveld method by X-ray diffraction (XRD) (the proportions (mass%) of C4AF, f.MgO, f.CaO, hemihydrate gypsum, and CaCO3), the mineral composition calculated using the Bogue formula (the proportion (mass%) of C2S), the proportion (mass%) of wet f.CaO, color (a, b), and shipping temperature (°C).

[0040] The tuning parameter α was set to 2.468, and a second prediction model was created by Lasso regression using the above 19 types of data as learning input data. The second predictive model took less than a minute to develop. The measured values ​​of the 115 learning input data were input to the second prediction model, and the predicted values ​​of the test output data related to the quality of 115 cements were output from the second prediction model. Next, the coefficient of determination (R 2 ) was calculated to be 0.45. In the same manner as in Example 1, the coefficient of determination (R 2 ) was calculated to be 0.39. The coefficient of determination when using prediction data (0.39) is similar to the coefficient of determination when using training data (0.45), indicating that mortar flow values ​​can be predicted with excellent accuracy even when using data that was not used to train the prediction model. Figure 4 shows the relationship between the measured and predicted strength of mortar at 28 days of age.

Claims

1. A method for predicting cement quality using a first predictive model created using machine learning and a second predictive model created using machine learning, comprising: The second prediction model is created by Lasso regression using a plurality of second learning data that are combinations of actual measured values ​​of second learning input data and actual measured values ​​of second learning output data related to cement quality, The tuning parameter α used in the Lasso regression is determined by the following steps (A) to (C): A method for predicting cement quality, comprising inputting actual measured values ​​of prediction input data into the second prediction model, outputting predicted values ​​of prediction output data relating to cement quality from the second prediction model, and predicting cement quality using the predicted values ​​of the prediction output data. (A) a first prediction model creation step of changing the value of the tuning parameter α within an arbitrarily determined range and interval, and creating the first prediction model by Lasso regression using a plurality of first learning data that are combinations of actual measurement values ​​of first learning input data and actual measurement values ​​of first learning output data related to cement quality for each changed value of the tuning parameter α, thereby obtaining the same number of first prediction models as the number of changes in the value of the tuning parameter α; (B) For each of the plurality of first prediction models obtained in the step (A), a plurality of test data are used, which are combinations of actual measurement values ​​of test input data and actual measurement values ​​of test output data related to cement quality, and (b1) a predicted value of test output data related to cement quality obtained by inputting the actual measurement values ​​of the test input data into the first prediction model, and (b2) a coefficient of determination (R ２ ) is calculated, the coefficient of determination calculation step (C) The plurality of coefficients of determination (R ２ ), the coefficient of determination (R ２ a tuning parameter determination step of determining the value of the tuning parameter α used in creating the first prediction model having the largest value of (a) above as the value of the tuning parameter α used in creating the second prediction model;

2. the first learning input data is made up of a plurality of types of data, The coefficient of determination (R ２ 2. The cement quality prediction method of claim 1, further comprising: calculating a standard partial regression coefficient for each type of data constituting the first learning input data used to create the first prediction model for which the value of (i) is the largest; and determining the type of data constituting the second learning input data based on the values ​​of the standard partial regression coefficients.

3. A cement quality prediction method as described in claim 2, wherein the data constituting the first learning input data, excluding data whose standard partial regression coefficient has a value of 0, is defined as the data constituting the second learning input data.

4. In the step (A), the arbitrarily determined numerical range is a numerical range that is arbitrarily determined within a numerical range of more than 0 and not more than 10, 4. The method for predicting cement quality according to claim 1, wherein the arbitrarily determined numerical interval is a numerical interval determined arbitrarily within a numerical range of 0.0001 to 0.

01.

5. the first learning input data, the second learning input data, the test input data, and the prediction input data are one or more types of data selected from the group consisting of data on cement, data on materials other than cement that constitute a cement composition obtained by kneading cement and water, data on blending conditions for the cement composition, data on means and methods for mixing the cement composition, data on the environment during mixing of the cement composition, and data on transportation of the cement composition; The cement quality prediction method according to any one of claims 1 to 4, wherein the first learning output data, the second learning output data, the test output data, and the prediction output data are data relating to the quality of the cement composition.

Images