JP2017178648A - Method for prospecting quality of cement or product condition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method, in which quality of cement may be prospected in a short time period and at a high accuracy.SOLUTION: There is provided a method for prospecting quality of cement or product condition, in which with neural network, groups of study data within a specific range where actual value of observation data of cement product is obtained from study data, and monitor data, after studies of neural network at a large trial number of study, which becomes σ<σ, the study of the neural network is repeated while reducing the trial number of study until reaching σ≥σ, and if a case where analysis degree determination value does not meet with a first setting value and a case where the analysis degree determination value meets with a predetermined second setting value, or a case where actual value of the observation data is within numerical value range created by specific restriction condition, actual value of the observation data is inputted to an input layer of the neural network and a prospective value of evaluation data is outputted from an output layer.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a method for predicting cement quality or manufacturing conditions using a neural network.

Many qualities are required for cement, such as fluidity and workability before curing, and various chemical and physical characteristics after curing.
As an important quality item, for example, compressive strength of mortar at a material age of 28 days in accordance with “JIS R 5201 (Cement physical test method)” can be mentioned. However, quality items such as compressive strength of the mortar described above, which take a long time to understand the quality test results, make it difficult to ship cement after confirming the quality test results. There was a problem that there was.
For this reason, quality control items in the manufacturing process such as the composition of the cement clinker (chemical composition and mineral composition) and the fineness of the cement are set at the cement manufacturing site, and the quality items based on the age of the material are controlled as prescribed. Management reference values based on experience are set for quality control items in the manufacturing process so as to satisfy the reference values.

However, the cement quality control method using quality control items in the manufacturing process as an index is indirect management based on limited alternative indices, so it is unavoidable that the management accuracy has some degree of ambiguity. Therefore, it is a method that must set management standards on the safe side excessively.
Therefore, while suppressing the occurrence of over-spec products resulting from quality control trends set on the safe side, we can stably manufacture products with a specified quality, and also detect abnormal products that are out of quality standards. In order to prevent manufacturing, various techniques for predicting cement quality with higher accuracy by utilizing various other information related to cement manufacturing in addition to information on quality control items in the manufacturing process have been proposed. .
As a cement quality prediction method including a neural network learning process, for example, Patent Document 1 has an input layer for inputting an actual measurement value of monitoring data and an output layer for outputting an estimated value of evaluation data. A method for predicting cement quality or manufacturing method using a neural network, and learning the neural network with a sufficiently large number of learning times such that σ _L <σ _M using the learning data and the monitor data. Later, learning of the neural network is repeated while reducing the number of learning until σ _L ≧ σ _M, and when the analytical determination value after learning is less than a preset value, cement manufacturing is performed in the input layer of the neural network. Enter the actual value of the monitoring data in the And outputs the estimated value of the evaluation data related to the evaluation of the quality and production conditions, the prediction method of quality or production conditions of the cement is described. According to the prediction method, cement quality or production conditions can be predicted in a short time and with high accuracy.

Japanese Patent No. 5323290

In the prediction method described in Patent Document 1, the actual value of the monitoring data in cement production that is input to the input layer of the neural network is near the upper or lower limit of the data range of the actual value of the monitoring data used for learning of the neural network. In some cases, the prediction accuracy is low.
The object of the present invention is to provide high-precision cement quality or production conditions even when the actual measurement value of the monitoring data in cement production is near the upper or lower limit of the data range of the actual monitoring data value used for learning of the neural network. It is to provide a method that can predict.

As a result of intensive studies to solve the above problems, the inventors of the present invention are methods for predicting cement quality or manufacturing conditions using a neural network, and after learning the neural network by a specific method, the cement When the actual measurement value of the monitoring data in manufacturing is a value within the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data used for learning of the neural network, The actual value of monitoring data in cement production is input to the input layer, the estimated value of evaluation data related to the evaluation of cement quality or manufacturing conditions is output from the output layer, and the actual value of monitoring data in cement manufacturing is output. If the numerical value is out of the above range, a neural network is added to the learning data group to which specific learning data is newly added. After learning the network, the actual value of the monitoring data in cement production is input to the input layer of the obtained neural network, and the evaluation data related to the evaluation of cement quality or manufacturing conditions is estimated from the output layer It has been found that the above object can be achieved by a method for outputting values, and the present invention has been completed.

（上記式（１）中、学習データの平均２乗誤差（σ_Ｌ）とは、学習データの監視データの実測値を学習後のニューラルネットワークの入力層に入力して得られた評価データの推測値と学習データの評価データの実測値との平均２乗誤差（σ_Ｌ）である。評価データの推測値の平均値とは、学習データの監視データの実測値を学習後のニューラルネットワークの入力層に入力して得られた評価データの推測値の平均値である。） That is, the present invention provides the following [1] to [6].
[1] A method for predicting cement quality or production conditions using a neural network having an input layer and an output layer, wherein the input layer is for inputting an actual value of monitoring data in cement production, The output layer is for outputting an estimated value of evaluation data related to evaluation of cement quality or production conditions, and a combination of the monitoring data and the evaluation data is
(I) The monitoring data is one or more types of data selected from data on cement clinker raw materials, data on cement clinker firing conditions, data on cement grinding conditions, and data on cement clinker, and A combination in which the evaluation data is one or more types of data selected from data on raw materials of cement clinker, data on firing conditions of cement clinker, data on grinding conditions of cement, data on cement clinker, and data on cement, or
(Ii) The monitoring data is one or more types of data selected from among data relating to cement clinker raw materials, data relating to cement clinker firing conditions, data relating to cement grinding, data relating to cement clinker, and data relating to cement. And the evaluation data is a combination of data relating to physical properties of a composition formed by kneading cement and water,
(A) initializing a learning data group composed of a plurality of learning data, which is a combination of the actual measurement value of the monitoring data and the actual measurement value of the evaluation data, used for learning of the neural network;
(B) initial setting of the number of learning times;
(C) using the set learning data group, learning the neural network, the number of learning times set in the previous step,
(D) An estimated value of the evaluation data obtained by inputting the actual measurement value of the monitoring data of the learning data to the input layer of the neural network obtained by the learning of the latest step (C) and the actual measurement of the evaluation data of the learning data Of the monitoring data in the monitor data, which is a combination of the measured value of the monitoring data and the evaluation value of the evaluation data for confirming the reliability of the mean square error (σ _L ) with the value and the learning result of the neural network Mean square error between the estimated value of the evaluation data obtained by inputting the actual measurement value to the input layer of the neural network obtained by the learning of the most recent step (C) and the actual measurement value of the evaluation data in the monitor data (Σ _M ) is calculated, and when the calculated relationship between σ _L and σ _M is σ _L ≧ σ _M , step (E) is performed, and the calculated relationship between σ _L and σ _M is σ _L < If σ _M , perform step (F) And a process of
(E) A learning number larger than any of the learning number set in the most recent step (B) and the reset number of learning times of the latest neural network is set as a new learning number in step (B). Setting and performing steps (C) to (D) again;
(F) resetting the number of learnings obtained by reducing the number of learnings performed in the learning of the latest neural network as a new number of learnings;
(G) Using the set learning data group, performing neural network learning the number of learning times set in the most recent step (F);
(H) The actual value of the monitoring data of the learning data is input to the input layer of the neural network obtained by the learning of the latest step (G), and the estimated value of the evaluation data obtained and the actual measurement of the evaluation data of the learning data Obtained by inputting the mean square error (σ _L ) with the measured value and the actual measurement value of the monitoring data in the monitoring data to the input layer of the neural network obtained by the learning of the latest step (G). When the mean square error (σ _M ) between the estimated value of the evaluation data and the actually measured value of the evaluation data in the monitor data is calculated, and the relationship between the calculated σ _L and σ _M is σ _L ≧ σ _M , Performing the step (J) and performing the step (I) when the calculated relationship between σ _L and σ _M is σ _L <σ _M ;
(I) If the number of learning times of the neural network in the most recent step (G) exceeds a preset number, steps (F) to (H) are performed again, and the neural network in the most recent step (G) is executed. A step of performing the step (M) when the number of learning is less than or equal to a predetermined number;
(J) An analytical determination value is calculated using the following equation (1), and when the analytical determination value is less than a predetermined first set value, learning of the neural network is terminated and step (K) is performed. A step of performing and a step of performing the step (M) when the analysis degree determination value is equal to or greater than a predetermined first set value;
(K) When the actual measurement value of the monitoring data in cement production is a numerical value within the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data in the set learning data group, (R) is carried out, and when it is a numerical value outside the range, the step of carrying out the step (L),
(L) The learning data in which the actual measurement value of the monitoring data is a numerical value outside the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data in the set learning data group, After one or more learning data not included in the data group is added to the learning data group and set as a new learning data group, the step (C) and the subsequent steps are performed. (M) Step (B) A determination is made as to the size of the performed number of times, and if the number of times is less than or equal to a preset number, step (N) is performed, and if the number exceeds the preset number of times, step (O) is performed. The steps to implement;
(N) initializing the learning conditions and performing steps (B) to (J) again;
(O) Of all the analytical determination values calculated in step (J), when the smallest analytical determination value is less than a predetermined second set value, the smallest analytical determination value can be obtained. If the neural network in the step (J) is a learned neural network, then the step (P) is performed, and if the smallest analytical determination value is greater than or equal to a predetermined second set value, the cement quality Or the process of determining that the manufacturing conditions cannot be predicted and terminating the prediction,
(P) The actual measurement value and evaluation data of the monitoring data used as learning data in the step (J) in which the smallest analysis degree determination value can be obtained among all the analysis degree determination values calculated in the step (J). An uncorrelated test was performed on the combination of measured values of, and when there were two or more types of monitoring data judged to be significant at the 5% significance level, it was judged to be significant at the 5% significance level. Plot actual values of monitoring data used as learning data in a coordinate space with all types of monitoring data as coordinate axes, and include all of the monitoring data formed by connecting the plotted monitoring data in the coordinate space After setting the area formed by connecting the monitoring data so that the area is maximized as the predictable monitoring data area, the process (Q) is performed, % If monitoring data is determined to be significant at a significance level of 0 or 1 kind, the step of terminating the prediction is determined that it is not possible to predict the quality or production conditions of the cement,
(Q) If the actual value of monitoring data in cement production is included in the predictable monitoring data area set in step (P), it is determined that the quality of cement or the production conditions can be predicted with high accuracy, Step (K) is carried out, and if the actual value of the monitoring data in cement production is not included in the predictable monitoring data area, it is determined that the cement quality or manufacturing conditions cannot be predicted and the prediction is terminated. Process,
(R) An estimated value of evaluation data related to evaluation of cement quality or manufacturing conditions is input from the output layer of the neural network by inputting an actual value of monitoring data in cement manufacturing to the input layer of the learned neural network. To output cement and predict cement quality or manufacturing conditions;
A method for predicting cement quality or manufacturing conditions, comprising:

(In the above equation (1), the mean square error (σ _L ) of the learning data is an estimation of the evaluation data obtained by inputting the measured value of the monitoring data of the learning data to the input layer of the neural network after learning. The mean square error (σ _L ) between the value and the actual measurement value of the evaluation data of the learning data The average value of the estimated value of the evaluation data is the input of the neural network after learning the actual measurement value of the monitoring data of the learning data (This is the average estimated value of the evaluation data obtained by entering the layer.)

[2] The method for predicting cement quality or manufacturing conditions according to [1], wherein the number of learning data added in the step (L) is 8% or more of the number of learning data in the set learning data group. .
[3] The predetermined first set value of the analytic degree determination value is 6% or less, the predetermined second set value of the analytical degree determination value is larger than the first set value and 20 %. The method for predicting cement quality or production conditions according to [1] or [2].
[4] The cement quality or the production condition according to any one of [1] to [3], wherein the neural network is a hierarchical neural network having an intermediate layer between the input layer and the output layer. Prediction method.
[5] The combination of the monitoring data and the evaluation data is a combination in which the monitoring data is data relating to cement, and the evaluation data is data relating to physical properties of a composition obtained by kneading cement and water. Yes, the data relating to the cement, which is the monitoring data, is the Blaine specific surface area, the mineral composition, and the chemical composition of the cement, and the data relating to the physical properties of the composition formed by kneading the cement and water is the evaluation data. The method for predicting cement quality or production conditions according to any one of [1] to [4] above, which is a compressive strength of mortar.
[6] The method according to any one of [1] to [5], wherein a cement production condition is optimized based on an estimated value of the evaluation data obtained by artificially changing the value of the monitoring data. A method of predicting cement quality or manufacturing conditions.

By using the method for predicting cement quality or manufacturing conditions of the present invention, even in conditions that are difficult with a prediction method including a learning process of a conventional neural network, such as a small number of learning data, it is short and high. Cement quality or manufacturing conditions can be predicted with accuracy.
In addition, even when the actual value of the monitoring data in cement production that is input to the input layer of the neural network is near the upper or lower limit of the data range of the actual monitoring data used for learning the neural network, Moreover, it is possible to predict cement quality or production conditions with high accuracy.
Moreover, it is possible to optimize manufacturing conditions in real time based on the obtained estimated value, and to improve the stabilization of cement quality.
Furthermore, high prediction accuracy can be maintained by continuing learning of the neural network.

It is a flowchart which shows an example of the prediction method of this invention. It is a figure which shows the predictable monitoring data area | region set in Example 2. FIG.

Hereinafter, the present invention will be described in detail.
The prediction method of the present invention includes a neural network having an input layer for inputting an actual measurement value of monitoring data in cement production and an output layer for outputting an estimated value of evaluation data related to evaluation of cement quality or production conditions. It is a method for predicting cement quality or production conditions using a network.
The neural network of the present invention may be a hierarchical neural network having an intermediate layer between an input layer and an output layer.

The following (i) or (ii) is mentioned as a combination of the monitoring data and the evaluation data.
(I) The monitoring data is one or more types of data selected from data on cement clinker raw materials, data on cement clinker firing conditions, data on cement grinding conditions, and data on cement clinker, and the evaluation Combination (ii) the above monitoring in which the data is one or more data selected from data on cement clinker raw material, data on cement clinker firing conditions, data on cement grinding conditions, data on cement clinker, and data on cement Data includes data on cement clinker raw materials, data on cement clinker firing conditions, data on cement grinding conditions, data on cement clinker, and data on cement It represents one or more data selected from within, and the evaluation data is data on the physical properties of the composition obtained by kneading cement and water combination

“Data on cement clinker raw material”, which is one of the monitoring data in the combination of (i), includes the chemical composition of the raw material of cement clinker, hydraulic modulus, residual amount of sieve test, brain specific surface area (fineness), Ignition loss, time point a certain time before the kiln input (for example, one time point before 5 hours, 4 time points before 3 hours, 4 hours before, 5 hours before and 6 hours before) The chemical composition of the raw material of cement clinker (mixed raw material of cement clinker from which fine particles etc. have been extracted by the air current flowing during conveyance). Hardening rate, supply amount, supply amount of cement clinker auxiliary material consisting of special raw materials such as waste, storage amount (blending amount) of blending silo of blended material, storage of blended material Reservoir amount (remaining amount), current value of the cyclone located between the raw material mill and the blending silo of the blended raw material (representing the number of revolutions of the cyclone and correlating with the speed of the raw material passing through the cyclone), cement The chemical composition, hydraulic modulus, Blaine specific surface area, sieve test residual amount, decarboxylation rate, moisture content, etc. of the raw material obtained by mixing the clinker kiln input material and the auxiliary material are listed. These data are used singly or in combination of two or more.
Here, the chemical composition of the raw material of cement clinker (prepared raw material or raw material in kiln) is SiO ₂ , Al ₂ O ₃ , Fe ₂ O ₃ , CaO, MgO, SO ₃ , Na ₂ O in the raw material of cement clinker. , K ₂ O, Na ₂ Oeq (total alkali), TiO ₂ , P ₂ O ₅ , MnO, Cl, Cr, Zn, Pb, Cu, Ni, V, As, Zr, Mo, Sr, Ba, F, etc. It is a content rate.

“Data relating to the firing conditions of cement clinker”, which is one of the monitoring data in the combination of (i), is the amount of cement clinker raw material inserted, kiln rotation speed, outlet temperature, firing zone temperature, cement clinker temperature, kiln Examples thereof include average torque, O ₂ concentration, NO _X concentration, clinker cooler temperature, preheater gas flow rate (correlation with preheater temperature), and the like. These data are used singly or in combination of two or more.

One of the monitoring data in the combination of (i) above, “data on cement grinding conditions” is the grinding temperature, the amount of water spray in the finishing mill, the separator air volume, the type of gypsum, the amount of gypsum added, the input of cement clinker Examples thereof include the amount, the number of revolutions of the finishing mill, the temperature of the powder discharged from the finishing mill, the amount of the powder discharged from the finishing mill, and the amount of the powder not discharged from the finishing mill.
“Data on cement clinker”, which is one of the monitoring data in the combination of (i), includes the mineral composition of cement clinker, crystallographic properties (such as lattice constant and crystallite diameter) of each mineral, and two or more minerals Composition ratio, chemical composition, wet f. CaO (free lime), weight, etc. are mentioned. These data are used singly or in combination of two or more.

Here, the mineral composition of the cement clinker is 3CaO · SiO ₂ (C ₃ S), 2CaO · SiO ₂ (C ₂ S), 3CaO · Al ₂ O ₃ (C ₃ A), 4CaO · Al ₂ O ₃ · Fe ₂ O ₃ (C ₄ AF), f. CaO, f. The content of MgO or the like. As "the ratio of the two or more mineral composition" includes, for example, the ratio of C 3 _{S /} C 2 _S.
The mineral composition of the cement clinker can be obtained by, for example, the XRD-Riet belt method.
The chemical composition of the cement clinker means SiO ₂ , Al ₂ O ₃ , Fe ₂ O ₃ , CaO, MgO, SO ₃ , Na ₂ O, K ₂ O, Na ₂ Oeq (total alkali), TiO _{2 in} the cement clinker. , P ₂ O ₅ , MnO, Cl, Cr, Zn, Pb, Cu, Ni, V, As, Zr, Mo, Sr, Ba, F, and the like.
In the combination (i), as monitoring data, only one kind of data selected from among data relating to cement clinker raw material, data relating to cement clinker firing conditions, data relating to cement grinding conditions, and data relating to cement clinker is provided. However, it is preferable to use two or more (plural) of these four types of data from the viewpoint of improving the accuracy of prediction of evaluation data.

The evaluation data in the combination of (i), “data on raw materials of cement clinker”, “data on firing conditions of cement clinker”, “data on grinding conditions of cement”, and “data on cement clinker” are respectively This is the same as the above-mentioned monitoring data “data related to cement clinker raw material”, “data related to cement clinker firing conditions”, “data related to cement grinding conditions”, and “data related to cement clinker”.
In addition, the above-mentioned “data on cement clinker raw material”, “data on cement clinker firing conditions”, “data on cement grinding conditions”, and “data on cement clinker” can also serve as monitoring data.
Examples of the “data on cement” which is one of the evaluation data in the combination of (i) include the brain specific surface area, the sieve test residual amount, the gypsum hemihydrate, the color tone, and the like.
The evaluation data in the combination (i) is one or more types of data selected from the data on the raw materials of the cement clinker described above.

The monitoring data in the combination of (ii) is “data on raw materials of cement clinker”, “data on firing conditions of cement clinker”, “data on grinding conditions of cement”, and “data on cement clinker”, respectively. This is the same as “data regarding cement clinker raw material”, “data regarding firing conditions of cement clinker”, “data regarding grinding conditions of cement”, and “data regarding cement clinker”, which are monitoring data in the combination of (i).
“Data on cement” which is monitoring data in the combination of (ii) includes chemical composition of cement, mineral composition of cement, crystallographic properties of each mineral (such as lattice constant and crystallite diameter), wet f. Examples include CaO, loss on ignition, specific surface area of branes, particle size distribution, sieve test residue, gypsum hemihydrate, color tone, and the like.
These data are used singly or in combination of two or more.

Here, the chemical composition of cement refers to SiO ₂ , Al ₂ O ₃ , Fe ₂ O ₃ , CaO, MgO, SO ₃ , Na ₂ O, K ₂ O, Na ₂ Oeq (total alkali) in the cement raw material, _{TiO 2, P 2 O 5,} MnO, is Cl, Cr, Zn, Pb, Cu, Ni, V, as, Zr, Mo, Sr, Ba, the content of such F.
The mineral composition of the _{cement, 3CaO · SiO 2 (C 3} S), 2CaO · SiO 2 (C 2 S), 3CaO · Al 2 O 3 (C 3 A), 4CaO · Al 2 O 3 · Fe 2 O 3 (C ₄ AF), f. CaO, f. It is the content of MgO, gypsum, calcite and the like.
In addition, as the data on the chemical composition and mineral composition of cement, “data on cement clinker” which is evaluation data in the combination (i) may be used.

  The “physical properties of the composition obtained by kneading cement and water”, which is evaluation data in the combination of (ii), is the compression strength, bending strength, fluidity (flow value), heat of hydration, setting time, Examples thereof include drying shrinkage rate, stability, swelling in water, sulfate resistance, neutralization, ASR resistance, and the like.

In the method for predicting the quality or production conditions of the cement of the present invention, the target cement is not particularly limited. , Mixed cement such as blast furnace cement and fly ash cement, and cement obtained by adding an admixture such as limestone powder and silica fume to Portland cement.
The manufacturing process of Portland cement is roughly divided into three processes: a raw material process, a firing process, and a finishing process. The raw material process is a process in which cement raw materials such as limestone, clay, silica stone, and iron oxide raw materials are prepared at an appropriate ratio and finely pulverized by a raw material mill to obtain a mixed raw material of Portland cement clinker. The firing step is a step in which the raw material for preparing Portland cement clinker is supplied to a rotary kiln via a suspension preheater and the like, sufficiently fired and then cooled to obtain a Portland cement clinker. The finishing step is a step of obtaining Portland cement by adding an appropriate amount of gypsum and the like to the obtained Portland cement clinker and finely pulverizing with a finishing mill.

In the present invention, the relationship between the monitoring data in cement production and the evaluation data related to the evaluation of cement quality or manufacturing conditions is learned in advance by a neural network, and using the learning result, based only on the monitoring data, Predict the evaluation data.
Hereinafter, the prediction method of the present invention will be described in detail with reference to FIG.
[Step (A)]
In step (A), an initial setting of a learning data group composed of a plurality of learning data, which is a combination of the actual measurement value of the monitoring data and the actual measurement value of the evaluation data, used for learning the neural network is performed.
The learning data can be obtained by preparing a sample for learning and measuring the actual measurement value of the monitoring data of the sample and the actual measurement value of the evaluation data. From the viewpoint of workability, it is preferable to prepare a sufficient number of learning data in advance. However, in the step (L) described later, new learning data is added to the learning data group as necessary. A new learning sample may be prepared, and new learning data may be obtained from the sample.

The learning data group may consist of all learning data obtained from a plurality of learning samples, and a plurality selected from all learning data (however, one or more data are deleted from all the data) It may consist of learning data.
When a plurality of learning data are selected from all the learning data and set as a learning data group, the learning data is selected from all the learning data by monitoring the monitoring data (the monitoring data in the combination of (i) above). It is preferable to select an actual measurement value of (ii) or an actual measurement value of the monitoring data in the combination of (ii)) that is not significantly different from the actual measurement value of the monitoring data in cement production that is the target of the prediction method of the present invention.
Specifically, among all the learning data, the learning clinker raw material composition is not significantly different from the cement clinker raw material composition in the cement production subject to the prediction method of the present invention, the production equipment change, Learning data that has not been updated or changed in manufacturing process is selected. When learning data having a significantly different composition of the cement clinker raw material or learning data in which a change in manufacturing equipment or the like is selected, the learning data may become noise and it may be difficult to improve prediction accuracy.

The number of learning data constituting the learning data group is not particularly limited, but is preferably 5 or more, more preferably 50 or more, still more preferably 100 or more, and particularly preferably 150 from the viewpoint of increasing prediction accuracy. That's it. Although the upper limit of the said number is not specifically limited, For example, it is 1,000.
A process (B) is implemented after completion | finish of a process (A).

[Step (B)]
In step (B), an initial setting of the number of learning is performed. The number of learning times to be set is not particularly limited, but is preferably a sufficiently large number such that overlearning (overlearning) of the neural network occurs. Specifically, it is usually 5,000 to 1,000,000 times, preferably 10,000 to 100,000 times.
In the step (B), it is preferable to set the number of learnings that causes over-learning of the neural network, specifically, the number of learnings such that σ _L <σ _M (details will be described later). Since the number of learning times is increased / decreased, there is no problem even if the number of learning times initially set in the step (B) is the number of learning times normally performed for learning of the neural network.
A process (C) is implemented after completion | finish of a process (B).

[Step (C)]
In the step (C), learning of the neural network is performed by the number of learnings set in the previous step, using the set learning data group.
Here, the “set learning data group” refers to the learning data group set in the step (A) or the step (L) and set most recently.
The “number of learnings set in the previous step” is the number of learnings set in the step (B) or the new number of learnings reset in the step (E), and the most recent step (step (B) or the number of learning times set in step (E)).
Specifically, from the set learning data group, the actual value of the monitoring data of the learning data is input to the input layer of the neural network, the estimated value of the evaluation data output from the output layer, and the estimation of the evaluation data The neural network is learned by comparing and evaluating the actual measurement value of the evaluation data of the learning data corresponding to the value and correcting the neural network by performing the set number of learning times.
When the learning number is changed and the neural network is re-learned, the neural network obtained as a result of the previous learning is initialized and learning is performed again.
A process (D) is implemented after completion | finish of a process (C).

[Step (D)]
In step (D), σ _L and σ _M are calculated. From the magnitude relationship between σ _L and σ _M , it can be determined whether learning has been performed a sufficiently large number of times to cause over-learning of the neural network.
Specifically, the estimated value of the evaluation data and the evaluation data of the learning data obtained by inputting the actual measurement value of the monitoring data of the learning data to the input layer of the neural network learned in the most recent step (C) The mean square error (σ _L ) with the actual measurement value is calculated. Next, the estimated value of the monitoring data and the estimated value of the evaluation data obtained by inputting the actual value of the monitoring data of the monitoring data to the input layer of the neural network learned in the most recent step (C) The mean square error (σ _M ) is calculated. Thereafter, by comparing the calculated values of σ _L and σ _M , it can be determined whether the learning of the neural network has been performed a sufficiently large number of times.
Here, the monitor data is a combination of the actual measurement value of the monitoring data and the actual measurement value of the evaluation data obtained from a sample different from the sample used for obtaining the learning data. This is data for confirmation.
From the viewpoint of workability, the number of samples of monitor data (a combination of actual values of monitoring data and actual values of evaluation data) is preferably 5 to 50%, more preferably 10 to 30% of the number of samples of learning data. is there.

When the relationship between σ _L and σ _M calculated in the step (D) is σ _L ≧ σ _M (“No” in the overlearning determination in FIG. 1), the number of learnings in the most recently performed step (C) is It can be determined that the number of times is not sufficiently large. In this case, step (E) is performed. When the relationship between σ _L and σ _M calculated in the step (D) is σ _L <σ _M (“Yes” in the overlearning determination in FIG. 1), the number of learnings in the most recent step (C) is It can be determined that the number of times was sufficiently large. In this case, step (F) is performed.

[Step (E)]
In the step (E), a learning number larger than any of the learning number set in the most recent step (B) and the reset number of the latest neural network is reset as a new learning number ( For example, a number obtained by multiplying the number of learning performed in the most recent step (C) by 2.0 is set as a new number of learning. After resetting the new number of learning times, steps (C) to (D) are performed again.

[Step (F)]
In the step (F), the learning number obtained by reducing the number of learnings performed in the latest neural network learning is reset as a new learning number (for example, 0 is set as the number of learnings performed in the latest neural network learning). ..95 is set as the new number of learning times.)
Note that learning of the latest neural network refers to learning performed in the near past. Specifically, it refers to learning performed in the nearer past in step (C) or step (G) described later. A process (G) is implemented after completion | finish of a process (F).

[Step (G)]
In step (G), learning of the neural network is performed using the set learning data group (the learning data group set in step (A) or step (L) and set most recently). The number of learning times set in step (F) is performed. The contents to be implemented in the step (G) are the same as those in the step (C) except that learning of the neural network is performed the number of times newly set in the step (F).
A process (H) is implemented after completion | finish of a process (G).

[Step (H)]
In step (H), end determination is performed using the neural network obtained in the learning of the latest step (G). Specifically, the actual value of the monitoring data of the learning data is input to the input layer of the neural network obtained by the learning in the latest step (G), and the estimated value of the evaluation data obtained by learning and the evaluation data of the learning data Obtained by inputting the mean square error (σ _L ) from the actual measurement value and the actual measurement value of the monitoring data in the monitor data to the input layer of the neural network obtained by the learning of the most recent process (G). The mean square error (σ _M ) between the estimated value of the evaluation data and the actually measured value of the evaluation data in the monitor data is calculated, and the relationship between the calculated σ _L and σ _M is σ _L ≧ σ _M In the case (“Yes” in the end determination in FIG. 1), it is possible to determine that the number of times of learning in the most recently performed step (G) is no longer sufficiently large. In this case, step (J) described later is performed. When the relationship between the calculated σ _L and σ _M is σ _L <σ _M (“No” in the end determination in FIG. 1), the number of times of learning in the most recently performed step (G) is still a sufficiently large number. It can be judged that there was. In this case, step (I) described later is performed.

[Step (I)]
In step (I), it is determined whether the number of learning times of the neural network in the latest step (G) has exceeded a predetermined numerical value. Step (I) is performed to avoid repeating steps (F) to (H) indefinitely. If the number of times the neural network has been learned in step (G) most recently performed in step (I) exceeds a predetermined value ("Yes" in FIG. 1), steps (F) to (H) are performed again. To do. When the number of times of learning in step (G) performed most recently in step (I) is equal to or smaller than a predetermined numerical value ("No" in FIG. 1), step (M) described later is performed.
The predetermined numerical value is not particularly limited, and may be, for example, a numerical value equal to or less than 1/100 of the number of learning times set in the step (F), 1 or less, or 0 or less.

上記式（１）中、学習データの平均２乗誤差（σ_Ｌ）とは、直近の工程（Ｈ）で算出された平均２乗誤差（σ_Ｌ）と同じである。評価データの推測値の平均値とは、学習データの監視データの実測値を、直近の工程（Ｇ）にて得られたニューラルネットワークの入力層に入力して得られた評価データの推測値の平均値である。
解析度の判定を行うことで学習を行ったニューラルネットワークを用いて、セメントの品質等の予測を高い精度で行うことができるか否かを判断することができる。解析度判定値が予め定めた第一の設定値未満（図１の第一の解析度判定における「Ｙｅｓ」）であれば、解析は十分であると判断され、ニューラルネットワークの学習は終了し、工程（Ｋ）を実施する。
解析度判定値が予め定めた第一の設定値以上（図１の第一の解析度判定における「Ｎｏ」）であれば、学習データを用いて学習を行ったニューラルネットワークをそのまま用いて、セメントの品質等の予測を高い精度で行うことはできないと判断され、工程（Ｍ）を実施する。 [Process (J)]
In the step (J), the analysis degree can be determined depending on whether or not the analysis degree determination value is less than a predetermined first set value. The analysis degree determination value is calculated using the following formula (1).

In the above formula (1), the average of the training data square error between (sigma _L) is the same as the mean square error calculated in the last step (H) (sigma _L). The average value of the estimated value of the evaluation data is the estimated value of the evaluation data obtained by inputting the actually measured value of the monitoring data of the learning data to the input layer of the neural network obtained in the most recent step (G). Average value.
It is possible to determine whether or not the cement quality and the like can be predicted with high accuracy by using the neural network that has been learned by determining the analysis degree. If the analysis level determination value is less than a predetermined first set value (“Yes” in the first analysis level determination in FIG. 1), it is determined that the analysis is sufficient, and learning of the neural network is terminated. Step (K) is performed.
If the analysis degree determination value is equal to or greater than a predetermined first set value (“No” in the first analysis degree determination in FIG. 1), the neural network learned using the learning data is used as it is, and the cement is used. It is determined that the quality and the like cannot be predicted with high accuracy, and the step (M) is performed.

The predetermined first set value is not particularly limited, but is preferably 10% or less, more preferably 8% or less, and particularly preferably 7% or less from the viewpoint of performing prediction with higher accuracy.
In the steps (B) to (J), the analysis determination value in the step (J) is less than the first set value set in advance, or the number of times set in the step (M) is set in advance. Repeat until it exceeds. The analysis degree determination value and the learned neural network obtained each time the step (J) is performed are stored as data for use in the step (O).
In addition, from the viewpoint of improving the prediction accuracy, when the step (L) is performed, the analytical value determination value and the learned value obtained each time the step (J) is performed before the step (L) is performed. The data of the neural network is discarded without being used in the step (O).

[Step (K)]
In step (K), the actual value of the monitoring data in cement production (what is input to the input layer of the neural network in which learning has been completed in step (R)) is a set learning data group (step (A) or step). In the learning data group set in (L), which is set most recently), it is a numerical value within the range of the average value ± average square error (σ _G ) of the actually measured values of the monitoring data (FIG. 1 (“Yes” in the determination of the learning data), the process (R) is performed. If the numerical value is outside the range (“No” in the determination of the learning data in FIG. 1), the process (L) is performed. carry out.
When there are a plurality of types of actually measured values of monitoring data in cement manufacturing, and at least one of the values is out of the range, the step (L) is performed.
In this step, by determining whether the actual measurement value of the monitoring data in cement production is a numerical value within the above range, the number of learning data constituting the learning data group is increased, and the neural network learning is performed. It can be determined whether it is necessary to do it again. Thereby, the prediction accuracy of the estimated value of the evaluation data related to the evaluation of the quality of the cement or the manufacturing conditions can be further improved.

Note that the numerical value within the range of the average value ± average square error (σ _G ) of the actual measurement values of the monitoring data is equal to or greater than the average value of the actual measurement values of the monitoring data−the mean square error (σ _G ). It means that it is less than or equal to the average value of the actually measured values of the monitoring data + the mean square error (σ _G ). Moreover, the numerical value outside the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data is greater than the average value of the actual measurement value of the monitoring data + the average square error (σ _G ). Or it means that it is less than the average value of the measured values of the monitoring data minus the mean square error (σ _G ).

[Step (L)]
In the step (L), the actual value of the monitoring data is monitored in the set learning data group (the learning data group set in the step (A) or the step (L) and set most recently). One or more pieces of learning data that are values outside the range of the average value ± average square error (σ _G ) of the actual measured values of the data and not included in the learning data group are stored in the learning data group. After adding and setting to a new learning data group, step (C) and subsequent steps are performed (steps appropriately selected from steps (C) to (T) are performed).
When there are a plurality of types of measured values of the monitoring data that are values outside the above range, for each of the measured values of the plurality of types of monitoring data, the average value ± average square error of the measured values of the monitoring data of the learning data One or more pieces of learning data that are numerical values outside the range of (σ _G ) and are not included in the learning data group are added to the learning data group and set as a new learning data group.
Note that, among the types of actually measured values of the monitoring data of the learning data to be added, the monitoring data of the learning data group, which is a numerical value within the range of the average value ± average square error (σ _G ) of the actually measured values of the monitoring data. There is no need to consider the actual measured value of the type.
By adding the above learning data, the actual value of the monitoring data in cement production that is input to the input layer of the neural network is the upper or lower limit of the data range of the actual value of the monitoring data of the learning data used for learning of the neural network. Even in the vicinity, the prediction accuracy can be improved.

The number of learning data to be added is preferably 8% or more, more preferably 10% or more, and particularly preferably 15% or more of the number of learning data in the set learning data group. If the number is 8% or more, the prediction accuracy can be further improved.
In the process (L), in the process (K), the actual measurement value of the monitoring data in cement production is an average value ± average square error (σ _G ) of the actual measurement values of the monitoring data in the set learning data group. Repeatedly until a numerical value within the range is reached.

[Process (M)]
In the step (M), it is determined whether or not the number of times that the step (B) for setting the number of learning is performed is equal to or less than a preset numerical value. By performing the determination, it is possible to avoid the steps (B) to (J) from being repeated infinitely.
In the step (M), when the number of times the step (B) is performed is equal to or less than the preset number (“Yes” in the number determination of FIG. 1), the learning condition is initialized, and the steps (B) to (B) are performed again. (J) is performed, and when the number of times exceeds a preset number of times (“No” in the number of times determination of FIG. 1), the step (O) is performed.
The number of times set in advance is not particularly limited, but is usually 5 times or more. The upper limit of the preset number of times is preferably 100 times or less from the viewpoint of preventing the steps (B) to (J) from being repeated greatly.

[Step (N)]
In step (N), learning conditions are initialized, and steps (B) to (J) are performed again.
As a method for initializing learning conditions, for example, a method of re-inputting learning data after randomly changing a threshold value of units constituting a neural network or a weight combining units, and obtaining learning data Examples include a method of inputting new learning data after increasing the number of samples, changing the type of monitoring data to be used, or excluding inappropriate learning data.

[Step (O)]
In the step (O), all analytical degree determination values calculated in the step (J) (however, each time the step (J) is performed before the step (L) is performed in the step (J), When the analysis degree determination value and the learned neural network data are discarded, the smallest analysis degree determination value among all the analysis degree determination values calculated in step (J) after performing step (L) is Whether or not the next prediction can be performed can be determined depending on whether or not it is less than a predetermined second set value.
By adding the determination of the step (O), even in the learned neural network that has been determined that the prediction of the cement quality or the like cannot be performed with high accuracy in the step (J), the next step (P ) To (Q), it is possible to determine whether or not the cement quality and the like can be predicted with high accuracy. When the smallest analytical determination value is less than a predetermined second set value (“Yes” in the second analytical determination of FIG. 1), the process (J ) Is obtained as a learned neural network, and then the step (P) is performed.

If the smallest analysis degree determination value is equal to or greater than a predetermined second set value (“No” in the second analysis degree determination in FIG. 1), using a neural network that has learned using learning data, Judging that it is impossible to predict the cement quality and the like with high accuracy, the prediction is terminated.
The predetermined second set value is larger than the first set value. Further, the upper limit is preferably 30% or less, more preferably 25%, from the viewpoint of performing prediction with higher accuracy.

[Step (P)]
In the process (P), a predictable monitoring data area used in the next process (Q) is set.
First, all the analytical determination values calculated in the step (J) (however, the analytical determination that is obtained every time the step (J) is performed before the step (L) is performed in the step (J)). When the value and the learned neural network data are discarded, the smallest analysis degree determination value among all the analysis degree determination values calculated in the step (J) can be obtained after the step (L) is performed. In the completed step (J), an uncorrelated test is performed on the combination of the actual measurement value of the monitoring data and the actual measurement value of the evaluation data used as learning data. When there are two or more types of monitoring data judged to be significant at the significance level of 5% in the uncorrelated test (“Yes” in the decorrelation test in FIG. 1), the significance is significant at the significance level of 5%. A coordinate space having all the types of monitoring data determined to be coordinate axes is created.
For example, when there are two types of monitoring data judged to be significant at the significance level of 5%, the total alkali amount of cement and the amount of C ₃ S of cement, the total alkali amount of cement is x-axis, A coordinate space with the amount of C ₃ S as the y-axis is created.

Next, among the actual measurement values of the monitoring data used as learning data, all the actual measurement values of the types of monitoring data determined to be significant at the significance level of 5% were plotted in the coordinate space and plotted in the coordinate space. A predictable monitoring data area is set by connecting monitoring data. The predictable monitoring data area is an area including all of the plotted monitoring data, and is an area formed by connecting the monitoring data so that the area becomes maximum. After setting the predictable monitoring data area, step (Q) is performed.
When the monitoring data judged to be significant at the significance level of 5% is 0 or 1 type (“No” in the uncorrelated test in FIG. 1), it is judged that the cement quality or the production condition cannot be predicted. To finish the prediction.

[Step (Q)]
In step (Q), using the actual value of monitoring data in cement production and the coordinate space set in step (P), the actual value of monitoring data in cement production and the learned neural network in step (O) It can be determined whether or not the cement quality and the like can be predicted with high accuracy.
When the actual value of the monitoring data in cement production used for prediction of cement quality is included in the predictable monitoring data area set in step (Q) (“Yes” in the coordinate determination of FIG. 1), the cement It is determined that the quality or manufacturing conditions can be predicted with high accuracy, and the step (K) is performed. When the actual value of monitoring data in cement production is not included in the predictable monitoring data area (“No” in the coordinate determination of FIG. 1), it is determined that the cement quality or manufacturing conditions cannot be predicted. Exit.
Note that the monitoring data in cement production is determined to be significant at the significance level of 5% in the uncorrelated test in actual measurement values of types of monitoring data that are not used as coordinate axes in the coordinate space set in step (P). If there is a type of monitoring data that does not exist), there is no limitation on the actual measurement value of the monitoring data from the types of monitoring data that are not used as the coordinate axes.

[Step (R)]
In step (R), actual values of monitoring data in cement manufacturing are input to the input layer of the learned neural network, and the quality of the cement or manufacturing conditions are evaluated from the output layer of the learned neural network. By outputting the estimated value of the evaluation data, it is possible to predict cement quality or production conditions.

  According to the prediction method of the present invention, the actual measurement value of the monitoring data in cement production input to the input layer of the neural network is near the upper limit or lower limit of the data range of the actual monitoring data used for learning the neural network. Even in the condition where it is difficult to predict with high accuracy by the conventional neural network, it is possible to predict cement quality or manufacturing conditions with high accuracy.

  In order to maintain the prediction accuracy at a high level, the neural network periodically checks the magnitude of the difference between the estimated value of the evaluation data and the actually measured value corresponding to the estimated value, and based on the inspection result, It is preferable to update the neural network. In the combination (i) (the neural network relating to prediction of the mineral composition of the cement clinker), the update cycle is preferably once per hour, more preferably once every 30 minutes. In the combination (ii) (a neural network for predicting the physical properties of a composition formed by mixing cement and water), preferably once a month, more preferably once a week, particularly preferably one day. Once.

According to the method for predicting cement quality or production conditions of the present invention, by using a neural network, a cement clinker mineral composition or a composition obtained by kneading cement and water (for example, by simply inputting monitoring data) The estimated value of the evaluation data such as the compressive strength of the mortar can be obtained within one hour.
In addition, based on the estimated value of the obtained evaluation data, it is possible to detect abnormalities in the quality of cement at an early stage during cement production, and optimize various conditions in the raw material process, firing process, and finishing process to achieve an appropriate quality. Of cement can be manufactured.
Specifically, when an abnormality is observed in the estimated value of the mineral composition of the cement clinker, the mineral composition of the cement clinker can be achieved by adjusting the raw material preparation, the firing conditions, and the like.
It is also possible to correct the management target value of the cement manufacturing process based on the estimated value of the evaluation data.
For example, when it is predicted that the mortar compressive strength will not reach the target value, the relationship between the monitoring data (factors) used for learning and the mortar compressive strength is analyzed, and the management target of the optimal cement manufacturing process is analyzed. By confirming the value, the quality of the cement can be achieved.

Furthermore, the monitoring data is artificially varied based on the evaluation data by connecting a computer for controlling the cement manufacturing process and a computer used for implementing the cement quality or manufacturing condition prediction method of the present invention. The control system for this can also be automated.
In the present invention, examples of software for performing an operation using a neural network include “Neural Network Library” (trade name) manufactured by OLSOFT.

Hereinafter, the present invention will be described by way of examples.
[When the first analysis degree determination value is less than the first set value]
[Example 1]
As a learning sample, 132 ordinary Portland cements with different sampling times were kneaded according to “JIS R 5201 (Cement physical test method)”, and the compressive strength of the mortar at the age of 28 days was measured. Learning data (actually measured evaluation data) was used.
In addition, the above-mentioned 132 ordinary Portland cements were measured for the Blaine specific surface area, the amount of each mineral, the total alkali amount (Na ₂ O + 0.658 × K ₂ O), and the amount of P ₂ O ₅ to obtain learning data (monitoring data). Measured value).
The amount of each mineral was measured with a powder X-ray diffractometer in the measurement range: 2θ = 10 to 65 °, and C ₃ S, C ₂ S, C calculated by Rietveld analysis software. ₄ AF, the amount of C ₃ A.
In addition, 10 normal Portland cements with different sampling times from the 132 samples were used as monitoring samples, and the mortar compressive strength at the age of 28 days was measured in the same manner as the learning data. Data (actually measured evaluation data) was used.
Further, the Blaine specific surface area, the amount of each mineral, the total alkali amount, and the P ₂ O ₅ amount of the above 10 ordinary Portland cements were measured in the same manner as the learning data to obtain monitor data (actual value of monitoring data). .

The neural network was learned using the learning data. As the neural network, a hierarchical neural network having an input layer, an intermediate layer, and an output layer was used. The neural network was first learned 10,000 times using the learning data and the monitor data. When σ _L and σ _M were calculated using the obtained neural network, the relationship between σ _L and σ _M was σ _L <σ _M. After that, the neural network is initialized, and the learning data and the monitor data are used to learn the neural network by performing the number of learning times (rounded down) by multiplying the number of learning times by 0.95. The relationship between σ _L and σ _M calculated using the network was repeated until σ _L ≧ σ _M.
After the relationship between σ _L and σ _M became σ _L ≧ σ _M , the analysis degree determination value was calculated to be 5.2%, which was less than the predetermined first set value of 6%. So I finished learning neural networks.

The average value and the mean square error were calculated for the actual values (the brain specific surface area, the amount of each mineral, the total alkali amount, and the P ₂ O ₅ amount) of the monitoring data of the learning data used for learning of the neural network.
In addition, as a sample for predicting cement quality using a learned neural network, normal Portland cement (hereinafter, also referred to as “prediction sample”), which is different from the above-described learning sample and monitoring sample, is used. Monitoring data (brane specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount) were measured.
The results are shown in Table 1.

Of the measured values of the monitoring data of the prediction sample (brain specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount), the measured value of the brane specific surface area is the measured value of the brane specific surface area of the learning data. Is a numerical value outside the range of the average value ± average square error (σ). Specifically, the actual value (3,090 cm ² / g) of the brain specific surface area of the prediction sample is the mean square from the average value (3,145 cm ² / g) of the actual brain surface area of the learning data. It is smaller than the numerical value (3,096 cm ² / g) obtained by subtracting the error (σ).
Therefore, when the estimated value of the pre-evaluation data (for example, the compressive strength of mortar at the age of 28 days) is predicted using the learned neural network and the actual measurement value of the monitoring data in cement production, the prediction accuracy May be low.

The actual measurement value of the monitoring data (Brain specific surface area) is learning data that is a numerical value outside the range of the average value ± average square error (σ) of the actual measurement value of the monitoring data (Brain specific surface area) of the learning data, And, 13 new learning data different from the above learning data and monitor data (a number corresponding to about 10% of the learning data group used for learning of the previous neural network) are used as new learning data. In addition to the 132 learning data, a total of 145 new learning data groups were obtained.
Note that the actual values of monitoring data (amount of minerals, total alkali amount, and P ₂ O ₅ amount) other than the brain specific surface area of the new learning data are the learning data (learning before adding new learning data). Data) of the monitoring data (amount of each mineral, total alkali amount, and P ₂ O ₅ amount) other than the Blaine specific surface area, the average value ± the mean square error (σ) outside the range and within the range The numbers were mixed.

Neural network learning was performed 10,000 times using new learning data (145 pieces) and monitor data. When σ _L and σ _M were calculated using the obtained neural network, the relationship between σ _L and σ _M was σ _L <σ _M. After that, the neural network is initialized, and the learning data and the monitor data are used to learn the neural network by performing the number of learning times (rounded down) by multiplying the number of learning times by 0.95. The relationship between σ _L and σ _M calculated using the network was repeated until σ _L ≧ σ _M.
After the relationship between σ _L and σ _M became σ _L ≧ σ _M , the analysis degree determination value was calculated to be 5.0%, which was less than the predetermined first set value of 6%. So I finished learning neural networks.
Calculate the average value and mean square error for the actual values (brane specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount) of the monitoring data of the learning data used for the second neural network learning. did.
The results are shown in Table 2. Incidentally, it is shown in Table 2, the monitoring data of the prediction for the sample by reference to the measured value of (Blaine specific surface area, the amount of each mineral, total alkali content, and P ₂ O ₅ amount).

Measured values of the monitoring data of the prediction data (the brain specific surface area, the amount of each mineral, the total alkali amount, and the amount of P ₂ O ₅ ) are the actual values of the monitoring data of the learning data used for learning the neural network again. It was a numerical value within the range of the average value ± average square error (σ).
The measured values of the monitoring data of the prediction sample (brane specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount) are input to the input layer of the learned neural network obtained by learning the neural network again. The quality of the cement was predicted by obtaining the compressive strength of the mortar at the age of 28 days as an estimated value of the evaluation data from the output layer.
In addition, when 30 samples (a number corresponding to about 25% of the learning samples used in the previous learning) are added to the 132 ordinary Portland cements as new learning samples, 65 (the previous learning) New learning data other than adding 107 (the number corresponding to about 80% of the learning samples used in the previous learning) or the number of learning samples used in In the same manner as when 13 pieces were added, the compressive strength of the mortar at the age of 28 days of the prediction sample was predicted.
Each result is shown in Table 3.
In addition, the measured value of the compressive strength of the mortar of the age 28 of the sample for prediction was 59.4 N / mm ² .

[Comparative Example 1]
In Example 1, the learned neural network obtained first (in the case where learning data is not added) and the monitoring data of the prediction sample (brain specific surface area, amount of each mineral, total alkali amount and P ₂ O) ₅ ), the mortar compressive strength at 28 days of age was predicted. The estimated value of mortar compressive strength at 28 days of age was 62.1 ± 2.7 (deviation was 3σ). N / mm ³

[When the first analysis degree determination value is greater than or equal to the first set value]
[Example 2]
59 normal Portland cements with different sampling times as samples for learning were kneaded according to “JIS R 5201 (cement physical test method)”, and the compressive strength of the mortar at the age of 28 days was measured. Learning data (actually measured evaluation data) was used.
In addition, the above 59 ordinary Portland cements were measured for the Blaine specific surface area, the amount of each mineral, the total alkali amount (Na ₂ O + 0.658 × K ₂ O), and the amount of P ₂ O ₅ to obtain learning data (monitoring data). Obtained).
The amount of each mineral was measured with a powder X-ray diffractometer in the measurement range: 2θ = 10 to 65 °, and C ₃ S, C ₂ S, C calculated by Rietveld analysis software. It is the amount of each mineral of ₃ A and C ₄ AF.
In addition, 10 normal Portland cements with different sampling times from the 59 samples were used as monitoring samples, and the mortar compressive strength at the age of 28 days was measured in the same manner as the learning data. Data (actually measured evaluation data) was used.
Further, the Blaine specific surface area, the amount of each mineral, the total alkali amount, and the P ₂ O ₅ amount of the above 10 ordinary Portland cements were measured in the same manner as the learning data to obtain monitor data (actual value of monitoring data). .

A neural network was learned using the learning data. As the neural network, a hierarchical neural network having an input layer, an intermediate layer, and an output layer was used. The initial learning number of the neural network was set to 10,000 times.
When σ _L and σ _M were calculated using the obtained neural network, the relationship between σ _L and σ _M was σ _L <σ _M.
Thereafter, the neural network is initialized, and the learning network and the monitor data are used to perform learning of the neural network by performing the number of learning times (rounded down) by multiplying the number of learning times by 0.95. It repeated until the relationship of (sigma) _L calculated using (sigma) _L and (sigma) _M became (sigma) _L > = (sigma) _M.
After the relationship between σ _L and σ _M becomes σ _L ≧ σ _M , the analytical determination value is calculated, but it is not less than 6%, which is the predetermined first setting value, and is the smallest analytical determination. The value was 7.4%. The neural network at the time of calculating the smallest analytical determination value is a learned neural network.
The set value of the second analysis degree determination value was 20%.

Compressive strength of mortar at the age of 28 days, which is an actual measurement value of learning data, and a brane specific surface area, an amount of each mineral, a total alkali amount, and a P ₂ O ₅ amount, which are actual values of monitoring data of learning data An uncorrelated test was performed for each of the above.
Table 4 shows the correlation coefficient between the compressive strength of the mortar at the age of 28 days and the specific surface area of the brane, the amount of each mineral, the total alkali amount, and the P ₂ O ₅ amount. Of the measured values of the monitoring data of the learning data, the learning data that is significant at the significance level of 5% (the correlation coefficient is greater than 0.197 or less than −0.197) is C ₃ A Amount and total alkali amount.

For the C ₃ A amount and the total alkali amount determined to be significant from the above uncorrelated test, the learning data (actual value of the monitoring data is displayed in the coordinate space where the total alkali amount is the x axis and the C ₃ A amount is the y axis. ) Data of total alkali amount and C ₃ A amount were plotted.
A region that is formed by connecting the plotted learning data and that includes all of the learning data and has the maximum area is set as a predictable monitoring data region (see FIG. 2a).

For the actual values (brane specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount) of the monitoring data of the learning data used for learning of the obtained learned neural network, the average value and the average The square error was calculated.
The results are shown in Table 5. In Table 5 shows the measured values of the monitoring data of the prediction for the sample (Blaine specific surface area, the amount of each mineral, total alkali content, and P ₂ O ₅ amount).

Of the measured values of the monitoring data of the prediction sample (brain specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount), the measured value of the brane specific surface area is the measured value of the brane specific surface area of the learning data. It was a numerical value outside the range of the mean value ± mean square error (σ). Specifically, the actual value (3,290 cm ² / g) of the brain specific surface area of the prediction sample is the mean square of the average value (3,180 cm ² / g) of the actual brain surface area of the learning data. It is larger than the numerical value (3,228 cm ² / g) obtained by adding the error (σ).
Therefore, when the estimated value of the pre-evaluation data (for example, the compressive strength of mortar at the age of 28 days) is predicted using the learned neural network and the actual measurement value of the monitoring data in cement production, the prediction accuracy May be low.

Therefore, the actual measurement value of the monitoring data (Blaine specific surface area) is learning data that is a value outside the range of the average value ± average square error (σ) of the actual measurement value of the monitoring data (Brain specific surface area) of the learning data. In addition, 15 new learning data different from the above learning data and monitor data (the number corresponding to about 25% of the learning data used in the previous neural network learning) are used as new learning data. In addition to the 59 learning data, a total of 74 learning data was used as a new learning data group.
Note that the actual values of monitoring data (amount of minerals, total alkali amount, and P ₂ O ₅ amount) other than the brain specific surface area of the new learning data are the learning data (learning before adding new learning data). Data) of the monitoring data (amount of each mineral, total alkali amount, and P ₂ O ₅ amount) other than the Blaine specific surface area, the average value ± the mean square error (σ) outside the range and within the range The numbers were mixed.

Neural network learning was performed 10,000 times using new learning data (74 pieces) and monitor data. When σ _L and σ _M were calculated using the obtained neural network, the relationship between σ _L and σ _M was σ _L <σ _M. After that, the neural network is initialized, and the learning data and the monitor data are used to learn the neural network by performing the number of learning times (rounded down) by multiplying the number of learning times by 0.95. The relationship between σ _L and σ _M calculated using the network was repeated until σ _L ≧ σ _M.
After the relationship between σ _L and σ _M became σ _L ≧ σ _M , the analytical determination value was calculated to be 5.8%, which was less than the predetermined first set value of 6%. So I finished learning neural networks.
The average value of the measured values of the brain specific surface area (monitoring data) of the learning data used for the second neural network learning is 3,190 cm ² / g, the mean square error (σ) is 65, and the prediction sample The actual measured value of Blain specific surface area was still a value outside the numerical range of the average value ± average square error (σ) of the actual measured value of Blain specific surface area of the learning data.

Therefore, the actual measurement value of the brain specific surface area (monitoring data) is the learning data that is a value outside the range of the average value ± average square error (σ) of the actual measurement value of the brain specific surface area (monitoring data) of the learning data. 44 new learning data different from the above learning data and monitor data (the total number of added learning data (59) is 100% of the number of learning data used for learning of the first neural network. Is added to the 74 learning data as new learning data, and a total of 118 learning data is used as a new learning data group, similar to the above-described neural network learning. The neural network was learned again.
Calculate the average value and mean square error for the actual values (brane specific surface area, amount of each mineral, total alkali amount, and P ₂ O ₅ amount) of the monitoring data of the learning data used for the second neural network learning. did.
The results are shown in Table 6. Incidentally, it is shown in Table 6, the monitoring data of the prediction for the sample by reference to the measured value of (Blaine specific surface area, the amount of each mineral, total alkali content, and P ₂ O ₅ amount).
The actual value of the brain specific surface area (monitoring data) of the learning data used for the second neural network learning is 3,210 cm ² / g, and the mean square error (σ) is 94, which is used for prediction. The actual measured value of the brain specific surface area of the sample was a numerical value within the range of the average value ± average square error (σ) of the actual measured value of the brain specific surface area of the learning data.

Compressive strength of mortar at the age of 28 days, which is an actual measurement value of the learning data (118) used for learning of the neural network, a brain specific surface area which is an actual measurement value of the monitoring data of the learning data, and the amount of each mineral For each of the total alkali amount and the P ₂ O ₅ amount, an uncorrelated test was performed.
Table 7 shows the correlation coefficient between the compressive strength of the mortar at the age of 28 days, the specific surface area of the brane, the amount of each mineral, the total alkali amount, and the P ₂ O ₅ amount. Of the measured values of the monitoring data of the learning data, the learning data that was significant at a significance level of 5% (those with a correlation coefficient exceeding 0.160 or less than -0.160) The amount of C ₃ A and the total amount of alkali were the same as in the correlation test.

Regarding the C ₃ A amount and the total alkali amount determined to be significant from the above-described uncorrelated test, the learning data (actual measurement of monitoring data) is performed in a coordinate space in which the total alkali amount is the x axis and the C ₃ A amount is the y axis. Data) of the total alkali amount and C ₃ A amount in (Value).
A region that is formed by connecting the plotted learning data and that includes all of the learning data and that has the maximum area is set as a predictable monitoring data region ( See FIG. 2b).
When the measured values of the monitoring data (total alkali amount, C ₃ A amount) of the prediction sample were plotted in the coordinate space, they were within the predictable monitoring data area range.

The actual values (brane specific surface area, amount of each mineral, total alkali amount and P ₂ O ₅ amount) of the monitoring data of the prediction sample are input to the input layer of the obtained learned neural network, and the material age is 28 days. An estimate of the compressive strength of the mortar was obtained. The estimated value was 64.7 ± 4.8 (deviation shows 3σ) N / mm ² .
Moreover, the measured value of the compressive strength of the mortar at the age of 28 days of the sample for prediction was 63.5 N / mm < ² >.

[Comparative Example 2]
After learning the neural network using 59 pieces of learning data, the obtained learned neural network was obtained in the same manner as in Example 2 except that the learning data was not added and the neural network was not re-learned. In the input layer, the measured values of the monitoring data (brane specific surface area, amount of each mineral, total alkali amount and P ₂ O ₅ amount) of the sample for prediction are input, and the estimated value of the compressive strength of the mortar at the age of 28 days Got. The estimated value was 65.5 ± 5.6 N (deviation shows 3σ) / mm ² .

Estimated value of evaluation data obtained in Example 2 (64.7 ± 4.8 N / mm ² ) and estimated value of evaluation data obtained in Comparative Example 2 (65.5 ± 5.6 N / mm ² ) The estimated value obtained in Example 2 is closer to the actually measured value (63.5 N / mm ² ) than the estimated value obtained in Comparative Example 2, and it can be seen that the reliability is higher.

Claims (6)

A method for predicting cement quality or manufacturing conditions using a neural network having an input layer and an output layer,
The input layer is for inputting actual values of monitoring data in cement manufacturing, and the output layer is for outputting estimated values of evaluation data related to evaluation of cement quality or manufacturing conditions. Yes,
The combination of the monitoring data and the evaluation data is
(I) The monitoring data is one or more types of data selected from data on cement clinker raw materials, data on cement clinker firing conditions, data on cement grinding conditions, and data on cement clinker, and A combination in which the evaluation data is one or more types of data selected from data on raw materials of cement clinker, data on firing conditions of cement clinker, data on grinding conditions of cement, data on cement clinker, and data on cement, or
(Ii) The monitoring data is one or more types of data selected from among data relating to cement clinker raw materials, data relating to cement clinker firing conditions, data relating to cement grinding, data relating to cement clinker, and data relating to cement. And the evaluation data is a combination of data relating to physical properties of a composition formed by kneading cement and water,
(A) initializing a learning data group composed of a plurality of learning data, which is a combination of the actual measurement value of the monitoring data and the actual measurement value of the evaluation data, used for learning of the neural network;
(B) initial setting of the number of learning times;
(C) using the set learning data group, learning the neural network, the number of learning times set in the previous step,
(D) An estimated value of the evaluation data obtained by inputting the actual measurement value of the monitoring data of the learning data to the input layer of the neural network obtained by the learning of the latest step (C) and the actual measurement of the evaluation data of the learning data Of the monitoring data in the monitor data, which is a combination of the measured value of the monitoring data and the evaluation value of the evaluation data for confirming the reliability of the mean square error (σ _L ) with the value and the learning result of the neural network Mean square error between the estimated value of the evaluation data obtained by inputting the actual measurement value to the input layer of the neural network obtained by the learning of the most recent step (C) and the actual measurement value of the evaluation data in the monitor data (Σ _M ) is calculated, and when the calculated relationship between σ _L and σ _M is σ _L ≧ σ _M , step (E) is performed, and the calculated relationship between σ _L and σ _M is σ _L < If σ _M , perform step (F) And a process of
(E) A learning number larger than any of the learning number set in the most recent step (B) and the reset number of learning times of the latest neural network is set as a new learning number in step (B). Setting and performing steps (C) to (D) again;
(F) resetting the number of learnings obtained by reducing the number of learnings performed in the learning of the latest neural network as a new number of learnings;
(G) Using the set learning data group, performing neural network learning the number of learning times set in the most recent step (F);
(H) The actual value of the monitoring data of the learning data is input to the input layer of the neural network obtained by the learning of the latest step (G), and the estimated value of the evaluation data obtained and the actual measurement of the evaluation data of the learning data Obtained by inputting the mean square error (σ _L ) with the measured value and the actual measurement value of the monitoring data in the monitoring data to the input layer of the neural network obtained by the learning of the latest step (G). When the mean square error (σ _M ) between the estimated value of the evaluation data and the actually measured value of the evaluation data in the monitor data is calculated, and the relationship between the calculated σ _L and σ _M is σ _L ≧ σ _M , Performing the step (J) and performing the step (I) when the calculated relationship between σ _L and σ _M is σ _L <σ _M ;
(I) If the number of learning times of the neural network in the most recent step (G) exceeds a preset number, steps (F) to (H) are performed again, and the neural network in the most recent step (G) is executed. A step of performing the step (M) when the number of learning is less than or equal to a predetermined number;
(J) An analytical determination value is calculated using the following equation (1), and when the analytical determination value is less than a predetermined first set value, learning of the neural network is terminated and step (K) is performed. A step of performing and a step of performing the step (M) when the analysis degree determination value is equal to or greater than a predetermined first set value;
(K) When the actual measurement value of the monitoring data in cement production is a numerical value within the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data in the set learning data group, (R) is carried out, and when it is a numerical value outside the range, the step of carrying out the step (L),
(L) The learning data in which the actual measurement value of the monitoring data is a numerical value outside the range of the average value ± average square error (σ _G ) of the actual measurement value of the monitoring data in the set learning data group, After one or more learning data not included in the data group is added to the learning data group and set as a new learning data group, the step (C) and the subsequent steps are performed. (M) Step (B) A determination is made as to the size of the performed number of times, and if the number of times is less than or equal to a preset number, step (N) is performed, and if the number exceeds the preset number of times, step (O) is performed. The steps to implement;
(N) initializing the learning conditions and performing steps (B) to (J) again;
(O) Of all the analytical determination values calculated in step (J), when the smallest analytical determination value is less than a predetermined second set value, the smallest analytical determination value can be obtained. If the neural network in the step (J) is a learned neural network, then the step (P) is performed, and if the smallest analytical determination value is greater than or equal to a predetermined second set value, the cement quality Or the process of determining that the manufacturing conditions cannot be predicted and terminating the prediction,
(P) The actual measurement value and evaluation data of the monitoring data used as learning data in the step (J) in which the smallest analysis degree determination value can be obtained among all the analysis degree determination values calculated in the step (J). An uncorrelated test was performed on the combination of measured values of, and when there were two or more types of monitoring data judged to be significant at the 5% significance level, it was judged to be significant at the 5% significance level. Plot actual values of monitoring data used as learning data in a coordinate space with all types of monitoring data as coordinate axes, and include all of the monitoring data formed by connecting the plotted monitoring data in the coordinate space After setting the area formed by connecting the monitoring data so that the area is maximized as the predictable monitoring data area, the process (Q) is performed, % When monitoring data is determined to be significant at a significance level of 0 or 1 kind, the step of terminating the prediction is determined that it is not possible to predict the quality or production conditions of the cement,
(Q) If the actual value of monitoring data in cement production is included in the predictable monitoring data area set in step (P), it is determined that the quality of cement or the production conditions can be predicted with high accuracy, Step (K) is carried out, and if the actual value of the monitoring data in cement production is not included in the predictable monitoring data area, it is determined that the cement quality or manufacturing conditions cannot be predicted and the prediction is terminated. Process,
(R) An estimated value of evaluation data related to evaluation of cement quality or manufacturing conditions is input from the output layer of the neural network by inputting an actual value of monitoring data in cement manufacturing to the input layer of the learned neural network. To output cement and predict cement quality or manufacturing conditions;
A method for predicting cement quality or manufacturing conditions, comprising:

(In the above equation (1), the mean square error (σ _L ) of the learning data is an estimation of the evaluation data obtained by inputting the measured value of the monitoring data of the learning data to the input layer of the neural network after learning. The mean square error (σ _L ) between the value and the actual measurement value of the evaluation data of the learning data The average value of the estimated value of the evaluation data is the input of the neural network after learning the actual measurement value of the monitoring data of the learning data (This is the average estimated value of the evaluation data obtained by entering the layer.)   The method for predicting cement quality or manufacturing conditions according to claim 1, wherein the number of learning data added in the step (L) is 8% or more of the number of learning data in the set learning data group.   The predetermined first setting value of the analysis degree determination value is 10% or less, and the predetermined second setting value of the analysis degree determination value is larger than the first setting value and 30% or less. 3. A method for predicting cement quality or production conditions according to claim 1 or 2.   The method for predicting cement quality or manufacturing conditions according to any one of claims 1 to 3, wherein the neural network is a hierarchical neural network having an intermediate layer between the input layer and the output layer.   The combination of the monitoring data and the evaluation data is a combination in which the monitoring data is data relating to cement, and the evaluation data is data relating to physical properties of a composition obtained by kneading cement and water, The data relating to cement, which is monitoring data, is the brane specific surface area, mineral composition, and chemical composition of the cement, and the data relating to the physical properties of the composition obtained by kneading the cement and water, which is the evaluation data, is that of mortar. It is compressive strength, The prediction method of the quality of cement or manufacturing conditions of any one of Claims 1-4. The cement quality or production conditions according to any one of claims 1 to 5, wherein the cement production conditions are optimized based on an estimated value of the evaluation data obtained by artificially changing the value of the monitoring data. Prediction method.

Images