Method for predicting quality of fresh concrete, quality prediction program, method for manufacturing fresh concrete, apparatus for predicting quality of fresh concrete, and manufacturing system for fresh concrete

Abstract

Improve the prediction accuracy when predicting the quality of fresh concrete using a prediction model. 【Solution means】A construction step of constructing a prediction model by machine learning based on training data including a plurality of data sets, an acquisition step of acquiring a plurality of feature amounts obtained when producing fresh concrete using a mixer, and a prediction step of using the prediction model to acquire a predicted value of quality corresponding to the plurality of feature amounts acquired in the acquisition step. In each of the plurality of data sets, a plurality of feature amounts and a correct value of quality are associated with each other. The plurality of feature amounts include one or more load feature amounts related to the power load value of the mixer when mixing the concrete material or stirring the fresh concrete. In the construction step, the prediction model is constructed such that one or more feature amounts selected from the plurality of feature amounts and the predicted value of quality are associated with each other by a decision tree. A method for predicting the quality of fresh concrete.

Description

Technical Field

[0001] The present disclosure relates to a method for predicting the quality of fresh concrete, a quality prediction program, a method for manufacturing fresh concrete, an apparatus for predicting the quality of fresh concrete, and a manufacturing system for fresh concrete.

Background Art

[0002] Patent Documents 1 and 2 disclose methods for predicting the quality of fresh concrete using a prediction model.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] The present disclosure provides a method for predicting the quality of fresh concrete, a quality prediction program, a method for manufacturing fresh concrete, an apparatus for predicting the quality of fresh concrete, and a manufacturing system for fresh concrete, which are useful for improving the prediction accuracy when predicting the quality of fresh concrete using a prediction model.

Means for Solving the Problems

[0005] [1]A quality prediction method for predicting the quality of fresh concrete, comprising: a construction step of constructing a prediction model by machine learning based on training data including a plurality of data sets; an acquisition step of acquiring a plurality of feature quantities obtained when producing fresh concrete using a mixer; and a prediction step of using the prediction model constructed in the construction step to acquire a predicted value of the quality corresponding to the plurality of feature quantities acquired in the acquisition step, wherein in each of the plurality of data sets, the plurality of feature quantities and the correct value of the quality are associated with each other, the plurality of feature quantities include one or more load feature quantities related to the power load value of the mixer when mixing concrete materials or stirring fresh concrete, and in the construction step, the prediction model is constructed such that one or more feature quantities selected from the plurality of feature quantities and the predicted value of the quality are associated with each other by a decision tree. A quality prediction method for fresh concrete.

[0006] [2]The quality prediction method for fresh concrete according to [1] above, wherein in the construction step, the prediction model is constructed by machine learning including ensemble learning.

[0007] [3]The quality prediction method for fresh concrete according to [2] above, wherein in the construction step, the prediction model is constructed by machine learning using boosting.

[0008] [4]The quality prediction method for fresh concrete according to [2] or [3] above, wherein in the construction step, the prediction model is constructed by machine learning using gradient boosting.

[0009] [5]The one or more load feature quantities include at least one of a value at a predetermined time point in the time series data of the power load value of the mixer when mixing concrete materials or stirring fresh concrete, and a statistic obtained from the time series data, and the statistic is one or more values selected from the group consisting of the difference between the maximum value and the minimum value, the difference between the maximum value and the final value, the total value, the average value, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness. The quality prediction method for fresh concrete according to any one of [1] to [4] above.

[0010] [6] The plurality of feature amounts further include one or more time feature amounts related to the timing when the mixer performs the kneading of the concrete material. The one or more time feature amounts include the date and time when the mixer performs the kneading of the concrete material, the cumulative number of batches in one day of the kneading by the mixer, the time from an arbitrarily set reference time point to the time when the mixer performs the kneading, and the cumulative number of batches of the kneading by the mixer from an arbitrarily set reference time point, and include at least one of them. The method for predicting the quality of fresh concrete according to any one of [1] to [5] above.

[0011] [7] A quality prediction program for causing a computer to execute the quality prediction method according to any one of [1] to [6] above.

[0012] [8] A method for manufacturing fresh concrete, including a manufacturing process for manufacturing fresh concrete, and a quality prediction process for predicting the quality by the quality prediction method according to any one of [1] to [6] above after or during the execution of the manufacturing process.

[0013] [9] A quality prediction device for predicting the quality of fresh concrete, including a model construction unit for constructing a prediction model by machine learning based on training data including a plurality of data sets, an acquisition unit for acquiring a plurality of feature amounts obtained when manufacturing fresh concrete using a mixer, and a prediction unit for obtaining a predicted value of the quality corresponding to the plurality of feature amounts acquired by the acquisition unit by using the prediction model constructed by the model construction unit. In each of the plurality of data sets, the plurality of feature amounts and the correct value of the quality are associated with each other. The plurality of feature amounts include one or more load feature amounts related to the power load value of the mixer when kneading the concrete material or stirring the fresh concrete. The model construction unit constructs the prediction model so that one or more feature amounts selected from the plurality of feature amounts and the predicted value of the quality are associated with each other by a decision tree. A quality prediction device for fresh concrete.

[0014]

[10] A manufacturing system for fresh concrete, comprising a manufacturing apparatus for manufacturing fresh concrete and the quality prediction apparatus described in [9] above, wherein the quality prediction apparatus predicts the quality of the fresh concrete manufactured by the manufacturing apparatus.

Advantages of the Invention

[0015] According to the present disclosure, there are provided a method for predicting the quality of fresh concrete, a quality prediction program, a method for manufacturing fresh concrete, a quality prediction apparatus for fresh concrete, and a manufacturing system for fresh concrete, which are useful for improving the prediction accuracy when predicting the quality of fresh concrete using a prediction model. [[ID=�]]

Brief Description of the Drawings

[0016]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, an embodiment will be described with reference to the drawings. In the description, the same reference numerals are given to the same elements or elements having the same function, and duplicate descriptions are omitted. FIG. 1 schematically shows a fresh concrete manufacturing system including a quality prediction device according to an embodiment.

[0018] [Fresh Concrete Manufacturing System] First, the outline of the fresh concrete manufacturing system will be described. The manufacturing system 1 (fresh concrete manufacturing system) shown in FIG. 1 is a system for manufacturing fresh concrete. The manufacturing system 1 manufactures fresh concrete through at least a process of kneading concrete materials. The concrete materials used in the manufacturing system 1 include cement, admixtures, coarse aggregates, fine aggregates, water, and admixtures.

[0019] Examples of the coarse aggregates include natural coarse aggregates or artificial coarse aggregates. Examples of the coarse aggregates also include gravel, crushed stone, slag coarse aggregates, lightweight coarse aggregates, recycled coarse aggregates, recovered coarse aggregates, or coarse aggregates obtained by mixing these. The gravel is mountain gravel, land gravel, river gravel, or sea gravel. The slag coarse aggregates are blast furnace slag aggregates, ferronickel slag aggregates, electric furnace oxidized slag aggregates, or coal gasification slag aggregates. The lightweight coarse aggregates are natural lightweight aggregates, by-product lightweight aggregates, or artificial lightweight aggregates. The coarse aggregates may include crushed rock or lime crushed stone.

[0020] Examples of fine aggregates include natural aggregates or artificial aggregates. Examples of fine aggregates also include sand, crushed sand, slag fine aggregate, lightweight fine aggregate, recycled fine aggregate, recovered fine aggregate, or a mixture of these fine aggregates. Sand includes mountain sand, land sand, river sand, or sea sand, etc. Slag fine aggregate includes blast furnace slag aggregate, ferronickel slag aggregate, copper slag aggregate, electric furnace oxidized slag aggregate, or coal gasification slag aggregate, etc. Lightweight fine aggregate includes natural lightweight aggregate, by-product lightweight aggregate, or artificial lightweight aggregate, etc.

[0021] Examples of the rock types of crushed stone and crushed sand include igneous rocks, sedimentary rocks, metamorphic rocks, silica, limestone, dolomite, or calcareous shale, etc. Igneous rocks include granite, diorite, porphyry, rhyolite, diabase, rhyolite, andesite, basalt, or serpentine, etc. Sedimentary rocks include conglomerate, sandstone, shale, slate, or tuff, etc. Metamorphic rocks include gneiss or schist, etc. As the coarse aggregate and fine aggregate, a mixture of two or more of the above-exemplified ones may be used.

[0022] At least a part of the manufacturing system 1 is installed at a location (e.g., a factory, etc.) different from the location (e.g., the site) where fresh concrete is used. After fresh concrete is loaded onto the transport vehicle 200, the transport vehicle 200 transports the fresh concrete to the site (e.g., the construction site) where the fresh concrete is used. Examples of the transport vehicle 200 include an agitator truck (mixer truck) or a dump truck. The manufacturing system 1 may manufacture fresh concrete from concrete materials so as to meet the target quality (required quality) set for each site. For example, an operator of the manufacturing system 1 determines the mix of the concrete materials so as to meet the target quality and inputs an operation instruction to the manufacturing system 1.

[0023] In one example, ready-mixed concrete is manufactured in the manufacturing system 1 and the quality of the ready-mixed concrete before shipment (such as inspection) is managed so as to meet the target quality when the ready-mixed concrete is used, which is set for each construction site. The time of use of the ready-mixed concrete corresponds to the time when the ready-mixed concrete is received at the construction site. For the management of the quality of the ready-mixed concrete before shipment, the target quality at the time of shipment of the ready-mixed concrete may be determined based on the target quality at the time of use of the ready-mixed concrete. The quality of the ready-mixed concrete includes fresh properties. Specific examples of the fresh properties include slump, slump flow, and air content.

[0024] The manufacturing system 1 includes, for example, a manufacturing device 100 and a quality prediction device 10. The manufacturing device 100 is a device that kneads and mixes concrete materials to manufacture ready-mixed concrete. The manufacturing device 100 may be installed in a factory or the like different from the construction site where the ready-mixed concrete is used. The manufacturing device 100 includes, for example, a material storage area 101, a transport device 104, a storage tank 111, a weighing tank 112, a collecting hopper 113, a mixer 114, and a loading hopper 115.

[0025] The material storage area 101 is a place for storing concrete materials. The material storage area 101 includes a plurality of silos 102. The plurality of silos 102 are containers for storing at least a part of the concrete materials according to the type of the materials. The plurality of silos 102 include, for example, a silo 102 for storing coarse aggregate, a silo 102 for storing fine aggregate, and a silo 102 for storing cement.

[0026] The transport device 104 is a device that transports the concrete materials stored in the plurality of silos 102 to the storage tank 111. The transport device 104 includes, for example, a belt conveyor for transporting the concrete materials. The transport device 104 may transport the concrete materials according to the type of the materials at different timings. In one example, based on an operation instruction by a control device provided in the manufacturing system 1, a specific material among various concrete materials is transferred to the transport device 104 and transported to the storage tank 111.

[0027] The storage bottle 111 temporarily stores various concrete materials. Various concrete materials are transported (conveyed) to the storage bottle 111 from the material storage area 101 by the transport device 104. The storage bottle 111 is configured to store various concrete materials individually. Hereinafter, "concrete materials" may sometimes be simply referred to as "materials". The various materials stored in the storage bottle 111 are supplied to the measuring bottle 112 as needed.

[0028] The measuring bottle 112 is arranged below the storage bottle 111. The measuring bottle 112 operates based on an operation instruction from the control device provided in the manufacturing system 1 and measures various materials individually. When the measuring bottle 112 detects the target amount of material instructed by the control device, it supplies the material to the aggregate hopper 113. When water is supplied to the measuring bottle 112, an admixture may be mixed into the water. The aggregate hopper 113 is arranged below the measuring bottle 112. The aggregate hopper 113 aggregates various materials discharged from the measuring bottle 112 and supplies the aggregated various materials to the mixer 114. Note that the manufacturing apparatus 100 may not be provided with the aggregate hopper 113, and various materials may be supplied from the measuring bottle 112 to the mixer 114.

[0029] The mixer 114 is arranged below the aggregate hopper 113. The mixer 114 is a device for kneading concrete materials. The mixer 114 manufactures fresh concrete by kneading (mixing) aggregates, cement, water, admixtures, etc. Fresh concrete is discharged from the bottom of the mixer 114 into the loading hopper 115. The mixer 114 may be a tilting mixer, a horizontal single-shaft mixer, a horizontal double-shaft mixer, or a pan mixer. The mixer 114 includes, for example, two stirring members 114a and a mixer drive unit 114b.

[0030] The stirring member 114a is a member that stirs various materials supplied to the mixer 114. The two stirring members 114a are arranged side by side inside the main body portion (container portion) of the mixer 114 and are rotatably provided. Each of the two stirring members 114a includes a rotating shaft extending in a horizontal one direction. The mixer drive unit 114b rotates the rotating shafts of the two stirring members 114a respectively based on an operation instruction from a control device provided in the manufacturing system 1. The mixer drive unit 114b includes a drive source such as a motor that applies a driving force to the stirring member 114a, for example. An opening / closing port for discharging the produced fresh concrete into the loading hopper 115 is provided at the bottom of the main body portion of the mixer 114.

[0031] The loading hopper 115 is disposed below the mixer 114 and temporarily stores the fresh concrete. The loading hopper 115 supplies the temporarily stored fresh concrete to the transport vehicle 200.

[0032] The manufacturing apparatus 100 described above is an example of a fresh concrete manufacturing apparatus, and the fresh concrete manufacturing apparatus may be configured in any manner as long as it can manufacture fresh concrete through a process of kneading concrete materials with a mixer. In the present disclosure, in the manufacture of fresh concrete, not only obtaining fresh concrete by kneading concrete materials, but also transporting the obtained fresh concrete and stirring the fresh concrete during transportation may be included. In one example, the transport vehicle 200 is also included in the manufacturing system 1.

[0033] In the present disclosure, a unit of fresh concrete that is produced by one mixing in the mixer 114 and loaded onto the transport vehicle 200 is defined as "one batch". The process executed by the manufacturing system 1 to produce one batch of fresh concrete is defined as "batch process". The number of times the batch process is executed (cumulative number of executions) from a certain reference time point is referred to as the "cumulative batch number". At the reference time point, the cumulative batch number is reset to zero. For example, every day, when the manufacturing apparatus 100 starts operating, the cumulative batch number is reset to zero. In this case, the cumulative batch number in the batch process executed at the k-th (k is an integer greater than or equal to 1) time within one day is k.

[0034] Fresh concrete for 1 to 3 batches may be loaded onto one transport vehicle 200. For example, when fresh concrete for 2 batches is loaded onto one transport vehicle 200, two batch processes according to the same manufacturing conditions are performed at different timings (in sequence). During a certain period within one day, a plurality of batch processes according to the same manufacturing conditions may be continuously performed in sequence.

[0035] <Quality Prediction Device> The quality prediction device 10 is a device that predicts the quality of fresh concrete produced by the manufacturing apparatus 100. The quality prediction device 10 is composed of one or more computers. When the quality prediction device 10 is composed of a plurality of computers, these computers are communicably connected to each other. The quality prediction device 10 may have a function of controlling the manufacturing apparatus 100 in addition to the function of predicting the quality. For example, the quality prediction device 10 is configured as a part of a control device that controls the manufacturing apparatus 100. The quality prediction device 10 may control the manufacturing apparatus 100 according to the set operating conditions. At least a part of the operating conditions may be determined by an instruction from an operator such as a worker.

[0036] The quality prediction device 10 may be connected to an input device 12 and a monitor 14. The input device 12 is a device that inputs information indicating an instruction from an operator or the like to the quality prediction device 10. The input device 12 may be any device as long as it can input desired information, and may be a keyboard (keypad), an operation panel, or a mouse. The monitor 14 is a device for displaying information from the quality prediction device 10 to an operator or the like. The monitor 14 may be any device as long as it can perform graphic display, and may be a liquid crystal display. The input device 12 and the monitor 14 may be integrated like a touch panel. The quality prediction device 10, the input device 12, and the monitor 14 may be integrated like a tablet computer (tablet terminal).

[0037] The quality of the fresh concrete to be predicted by the quality prediction device 10 includes not only the quality of the fresh concrete itself but also the quality of the hardened fresh concrete (hardened concrete). Predicting the quality of fresh concrete means calculating a predicted value of an index value representing the quality.

[0038] The quality prediction device 10 predicts, for example, the fresh properties of fresh concrete as the quality of fresh concrete. The quality prediction device 10 may predict the fresh properties of fresh concrete obtained by mixing in the manufacturing device 100 and before being shipped from the factory or the like where the manufacturing device 100 is installed to the site as the quality of fresh concrete. Alternatively, the quality prediction device 10 may predict the fresh properties of fresh concrete after being shipped from the factory or the like and before being placed at the site as the quality of fresh concrete. That is, the quality prediction device 10 may predict the fresh properties of fresh concrete during transportation or may predict the fresh properties of fresh concrete at the time of unloading.

[0039] The quality prediction device 10 may predict one or more index values among slump, slump flow, and air content as the fresh properties of fresh concrete. The quality prediction device 10 may predict two or more index values among slump, slump flow, and air content as the fresh properties of fresh concrete. The quality prediction device 10 may predict index values other than slump, slump flow, and air content as the fresh properties of fresh concrete. Examples of the indexes of fresh properties other than slump, slump flow, and air content include, for example, 500 mm flow arrival time, flow stop time of flow, presence or absence of material separation, temperature, bleeding amount, bleeding rate, setting time, unit water amount, plastic viscosity, yield value, funnel flow-down time, compactability, deformability, fillability, and gap passing property.

[0040] The quality prediction device 10 may predict at least one of an index representing strength and an index representing durability as the quality of hardened fresh concrete. Examples of the index representing strength include compressive strength, flexural strength, splitting tensile strength, adhesive strength, static elastic modulus, resilience, and viscosity coefficient. Examples of the index representing durability include length change rate, expansion rate, mass reduction rate, carbonation depth, chloride ion penetration depth, chloride ion diffusion coefficient, air permeability coefficient, water permeability coefficient, freeze-thaw resistance, electrical resistivity, dynamic elastic modulus, Poisson's ratio, and creep coefficient. Hereinafter, "the quality of fresh concrete" may be simply referred to as "quality".

[0041] To explain the outline of the function of the quality prediction device 10, the quality prediction device 10 is configured to at least construct a prediction model by machine learning based on training data including a plurality of data sets, and acquire a plurality of feature amounts obtained when manufacturing fresh concrete using the mixer 114. The quality prediction device 10 is further configured to execute acquiring a predicted value of quality corresponding to the acquired plurality of feature amounts using the above prediction model. The quality prediction device 10 constructs the prediction model so that one or more feature amounts selected from the above plurality of feature amounts and the predicted value of quality are associated with each other by a decision tree.

[0042] FIG. 2 shows an example of the functional components (hereinafter referred to as "functional blocks") provided in the quality prediction device 10. The quality prediction device 10 has, for example, as functional blocks, an operation control unit 22, a feature amount acquisition unit 24, a model construction unit 26, a model holding unit 28, a prediction unit 30, and an output unit 32. The processes executed by these functional blocks correspond to the processes executed by the quality prediction device 10. Hereinafter, taking as an example the case where the quality to be predicted is the fresh property of fresh concrete (more specifically, the fresh property of fresh concrete obtained after being kneaded by the mixer 114 and before being shipped), the description of each functional block will be given.

[0043] The operation control unit 22 controls the manufacturing device 100 to manufacture fresh concrete according to predetermined operation conditions. At least a part of the above operation conditions may be determined by an operator such as a worker each time the manufacturing of fresh concrete is executed. The operation control unit 22 may control the mixer drive unit 114b so that the rotation speed of the mixer drive unit 114b of the mixer 114 follows the target rotation speed defined in the operation conditions. When controlling the mixer drive unit 114b, the operation control unit 22 may adjust the power (for example, current value) supplied to the mixer drive unit 114b. When the manufactured fresh concrete is hard, the power load value tends to increase, and when the manufactured fresh concrete is soft, the power load value tends to decrease.

[0044] The feature amount acquisition unit 24 (acquisition unit) acquires a plurality of feature amounts obtained when manufacturing fresh concrete using the mixer 114. Hereinafter, each of the plurality of feature amounts acquired by the feature amount acquisition unit 24 may be denoted as feature amount F1, feature amount F2, ···, and feature amount FN (N is an integer of 2 or more). The feature amounts F1 to FN are input data to a prediction model for predicting the fresh property. At least a part of the feature amounts F1 to FN (a plurality of feature amounts) may be a feature amount representing the result of the operation of the mixer 114 or a feature amount representing the state during the operation of the mixer 114. The feature amounts F1 to FN may include a feature amount representing the manufacturing conditions (for example, the above operation conditions) using the mixer 114.

[0045] Each of the feature quantities F1 to FN is data represented by numerical values. The feature quantities F1 to FN may include data representing classification (type), and in that case, different numerical values (for example, numerical values of 0 or 1) are assigned for each classification. In one example, it can be set that the case where it does not correspond to the classification is 0 and the case where it corresponds to the classification is 1. The feature quantities F1 to FN do not include image data constituted by a combination of coordinate information and pixel values. However, the feature quantities F1 to FN may include feature quantities (one-dimensional numerical data) obtained from an image related to such image data.

[0046] The feature quantities F1 to FN include one or more feature quantities related to the power load value of the mixer 114. Hereinafter, one or more feature quantities related to the power load value among the feature quantities F1 to FN are denoted as "load feature quantity FL" in order to distinguish them from other types of feature quantities. The load feature quantity FL related to the power load value of the mixer 114 is, for example, a feature quantity related to the power load value of the mixer 114 when kneading a concrete material. The power load value of the mixer 114 may be a value indicating the power (W) supplied to the mixer 114 itself, or may be a value indicating the current value (A) supplied to the mixer 114. Alternatively, the power load value of the mixer 114 may be a value indicating the load hydraulic pressure (Pa) or a value indicating the torque (N·m) instead of the power or the current value.

[0047] The one or more load feature quantities FL include, for example, at least one of the value at a predetermined time point in the time-series data of the power load value of the mixer 114 when kneading a concrete material and the statistic obtained from the time-series data. The time-series data of the power load value may be a data group (data group indicating the time change of the power load value) constituted by continuous power load values in at least a part of the period during which the mixer 114 operates during the processing of one batch. Specific examples of the feature quantities F1 to FN and specific examples of the load feature quantity FL will be described later.

[0048] The model construction unit 26 constructs a prediction model by machine learning based on training data including a plurality of data sets. Hereinafter, the training data including a plurality of data sets is denoted as "training data TD", and the prediction model constructed by the model construction unit 26 is denoted as "prediction model M". The prediction model M is a model for predicting the fresh properties of green concrete when the feature quantities F1 to FN are obtained. The fresh properties predicted by the prediction model M are, for example, slump, slump flow, or air content.

[0049] The model construction unit 26 constructs a prediction model M showing the relationship between the feature quantities F1 to FN and the fresh properties (predicted values) by machine learning. The prediction model M is configured to output a predicted value of the fresh properties in response to the input of the feature quantities F1 to FN. The feature quantities F1 to FN can be referred to as explanatory variables, and the fresh properties can be referred to as objective variables.

[0050] Machine learning refers to a method in which a machine (computer) repeatedly learns based on given information to autonomously find laws or rules. The prediction model M can be constructed using an algorithm and a data structure. The prediction model M is realized using a decision tree, which is a method of analyzing data by a tree structure (tree diagram). The prediction model M outputs a predicted value by sequentially setting conditions for the given input data and branching to the assumed results.

[0051] The model construction unit 26 constructs a prediction model M such that one or more feature quantities selected from among the feature quantities F1 to FN and the predicted value of the fresh properties are associated with each other using a decision tree (decision tree algorithm). One or more feature quantities used for the association with the fresh properties in the prediction model M are selected by machine learning from among the feature quantities F1 to FN. As the decision tree algorithm, LightGBM (Light Gradient Boosting Machine) may be used. Note that, as the decision tree algorithm, instead of LightGBM, XGBoost (eXtreme Gradient Boosting) or RandomForest may be used.

[0052] The model construction unit 26 may autonomously construct the prediction model M by performing machine learning using the data given as the input of the machine learning and the correct answer data (correct value of the fresh properties) of the output of the machine learning. The input of the machine learning is the feature quantities F1 to FN. The output of the machine learning is data (numerical value) indicating the fresh properties of the fresh concrete. The model construction unit 26 repeatedly learns a model that outputs a predicted value of the fresh properties using a plurality of combinations (the above-described training data TD) of the feature quantities F1 to FN and the correct values of the fresh properties. In each of the plurality of data sets constituting the training data TD, the feature quantities F1 to FN and the correct value of the fresh properties are associated with each other.

[0053] The stage of autonomously constructing the prediction model M corresponds to the training phase (learning phase). The training phase may be performed before the production phase of producing the fresh concrete, or may be performed at the initial stage of the production phase. The stage of predicting the fresh properties using the prediction model M from the feature quantities F1 to FN (combinations of the values of the feature quantities F1 to FN) whose fresh properties are unknown corresponds to the evaluation phase. In the following description, in order to distinguish whether it is in the training phase or the evaluation phase, or what data is used in which phase, the term may be appended with "for training" or "for evaluation". The same type of feature quantities F1 to FN are used between the training phase and the evaluation phase.

[0054] Here, with reference to FIG. 3, an example of the analysis process in the prediction model M will be described. FIG. 3 shows a tree diagram visually representing an example of the analysis process in the prediction model M. For ease of understanding the content of the present disclosure, a simplified example will be used for the description. In the algorithm using the decision tree exemplified below, one data group is divided step by step into two data groups according to a certain condition, and each step is referred to as the "first step", the "second step", and the "third step" in order from the top. Also, the location where division or branching is performed according to the condition is called a "node (condition node)". In FIG. 3, the combination of "G" and a numerical value such as "G1" represents the name of the data group. Also, the combination of "Th" and a numerical value such as "Th1" represents a threshold value.

[0055] The model construction unit 26 prepares training data TD composed of a plurality of data sets in which the combinations of the values of the feature amounts F1 to FN are different from each other. The data group G1 is, for example, all the data sets included in the training data TD. In the first step, one feature amount (feature amount F2 in the example shown in FIG. 3) selected by machine learning from among the feature amounts F1 to FN divides the data group G1 into two data groups, namely the data group G21 and the data group G22. In the first step, the data group G1 is divided into two data groups according to the condition of whether the feature amount F2 is smaller than the threshold value Th1 or is equal to or greater than the threshold value Th1.

[0056] The data group G21 is composed of a plurality of data sets in which the feature amount F2 is smaller than the threshold value Th1 among the data group G1, and the data group G22 is composed of the remaining plurality of data sets in which the feature amount F2 is equal to or greater than the threshold value Th1 among the data group G1. The number of data sets in the data group G21 and the number of data sets in the data group G22 may be the same as each other or may be different from each other. Although the feature amount F2 is used in the first step to divide into two data groups, the values of the feature amounts other than the feature amount F2 are also retained in the data group G21 and the data group G22. The threshold value Th1 is also set autonomously by machine learning.

[0057] In the second stage, each of the data groups G21 and G22 is divided into two data groups by a feature amount and a threshold value selected by machine learning, in the same manner as in the first stage. The feature amount selected in the second stage may be different from or the same as the feature amount selected in the first stage. The feature amounts selected for dividing each of the data groups G21 and G22 into two may be different from or the same as each other. In the example shown in FIG. 3, the feature amount F1 is selected to divide the data group G21 into two, and the feature amount F2 is selected to divide the data group G22 into two. In the second stage, the data group G21 is divided into a data group G31 and a data group G32, and the data group G22 is divided into a data group G33 and a data group G34.

[0058] In the example shown in FIG. 3, in the third stage, the data groups G31, G32, and G34 are not divided into two data groups. That is, each of the data groups G31, G32, and G34 corresponds to a "leaf" (or "terminal node") in the tree diagram. In the third stage, with respect to the data group G33, which is not a data group corresponding to a leaf, it is divided into two data groups by a feature amount and a threshold value selected by machine learning, in the same manner as in the previous stage. The feature amount F10 is selected to divide the data group G33 into two, and the data group G33 is divided into a data group G41 and a data group G42. Each of the data groups G41 and G42 corresponds to a leaf. As illustrated in FIG. 3, the depths (levels) of the leaves may be different among at least some of the plurality of leaves in the tree diagram. Alternatively, different from the example shown in FIG. 3, the depth of each leaf may be the same.

[0059] For each data group corresponding to a leaf, the predicted value of the fresh property can be determined by the correct value of the fresh property included in the data group. For example, the arithmetic mean of the correct values of the fresh property included in the data group corresponding to the leaf is set as the predicted value for that leaf. In one example, when the data group G41 contains m (m is an integer of 2 or more) data sets, the arithmetic mean of the m correct values of the fresh property is set as the predicted value for the leaf corresponding to the data group G41. Among multiple leaves, the predicted values of the fresh property set for those leaves are different.

[0060] The model construction unit 26 may perform the division (branching) up to the data group corresponding to the leaf while selecting the feature amount to be associated with the predicted value of the fresh property and setting the threshold value used when branching under any construction conditions. As described above, in the training phase, the training data TD is divided into two data groups according to the conditions for each stage, and the predicted value of the fresh property is set for each leaf (terminal node).

[0061] In the evaluation phase of obtaining the predicted value of the fresh property from the evaluation feature amounts F1 to FN using the prediction model M, it is evaluated which leaf the given evaluation feature amounts F1 to FN (combinations of the values of the feature amounts F1 to FN) are classified into according to the branching conditions for each stage. Since each leaf has a predicted value of the fresh property, the predicted value of the fresh property can be obtained according to the combination of the values of the feature amounts F1 to FN.

[0062] In the prediction model M, only some of the selected feature amounts among the feature amounts F1 to FN may be used for the branching conditions. That is, in the prediction model M, only some of the selected feature amounts among the feature amounts F1 to FN may be associated (related) with the predicted value of the fresh property. Even in such a case, since the part of the feature amounts associated with the predicted value of the fresh property is selected from among the feature amounts F1 to FN, the prediction model M still shows the relationship between the feature amounts F1 to FN and the predicted value of the fresh property.

[0063] The model construction unit 26 may construct a corresponding prediction model M (a prediction model M that predicts only one type of fresh property) for each type of fresh property. For example, the model construction unit 26 may separately construct a prediction model M for obtaining a predicted value of slump, a prediction model M for obtaining a predicted value of slump flow, and a prediction model M for obtaining a predicted value of air content.

[0064] Returning to FIG. 2, the model holding unit 28 holds (stores) the prediction model M constructed by the model construction unit 26. The prediction model M, which is a trained model, may be transferable between computers. Therefore, the prediction model M constructed in the quality prediction device 10 may be used in another manufacturing system different from the manufacturing system 1.

[0065] The prediction unit 30 uses the prediction model M constructed by the model construction unit 26 to obtain a predicted value of the fresh property corresponding to the evaluation feature amounts F1 to FN acquired by the feature amount acquisition unit 24. For example, the prediction unit 30 inputs the values of the evaluation feature amounts F1 to FN into the prediction model M and obtains the predicted value of the fresh property output from the prediction model M. The prediction unit 30 may obtain the predicted value using the corresponding prediction model M for each type of fresh property (for example, for each of slump, slump flow, and air content).

[0066] Assuming that the prediction model M illustrated in FIG. 3 is constructed, in the prediction model M, first, a comparison is made between the value of the evaluation feature amount F2 among the evaluation feature amounts F1 to FN and the threshold value Th1. If the value of the evaluation feature amount F2 is greater than or equal to the threshold value Th1, the calculation step proceeds to the node corresponding to the data group G22, and a comparison is made between the value of the evaluation feature amount F2 and the threshold value Th22. If the value of the evaluation feature amount F2 is smaller than the threshold value Th22, the calculation step proceeds to the node corresponding to the data group G33, and a comparison is made between the value of the evaluation feature amount F10 and the threshold value Th31.

[0067] And, if the value of the feature quantity F10 for evaluation is smaller than the threshold value Th31, the prediction model M outputs the predicted value of the fresh property set for the leaf (terminal node) corresponding to the data group G41. By executing the above-described operations in the prediction model M, the prediction unit 30 can obtain the predicted value of the fresh property of the green concrete to be evaluated from the combination of the values of the feature quantities F1 to FN for evaluation.

[0068] The output unit 32 outputs the predicted value of the fresh property obtained by the prediction unit 30 to the monitor 14. As a result, the predicted value of the fresh property of the green concrete to be evaluated is displayed on the monitor 14, and an operator such as a worker can grasp the predicted value of the fresh property of the green concrete.

[0069] <Specific Example of Decision Tree Algorithm> Subsequently, a specific example of machine learning executed by the model construction unit 26 when constructing the prediction model M based on the decision tree algorithm will be described with reference to FIGS. 4 to 7. The model construction unit 26 may construct the prediction model M by machine learning including ensemble learning. In machine learning including ensemble learning, a plurality of models may be fused to construct one prediction model M.

[0070] FIG. 4(a) schematically shows an analysis process for obtaining a predicted value using bagging, which is one method of ensemble learning. In one example, a first model M1 (weak learner) is generated from a part of the data randomly extracted from the training data TD, and a second model M2 (weak learner) is generated from another part of the data randomly extracted from the training data TD. A part of the data for generating the first model M1 and a part of the data for generating the second model M2 may overlap. In the prediction model M constructed by bagging, for example, the arithmetic mean of the predicted value in the first model M1 and the predicted value in the second model M2 is output as the final predicted value.

[0071] In FIG. 4(a), for the sake of simplicity, an example is illustrated in which the prediction model M is constructed from the fusion of two models (weak learners). However, the prediction model M that finally outputs one prediction value may also be constructed from three or more models (weak learners). A model (weak learner) generated from a part of the data such as the first model M1 may be a RandomForest (random forest) generated by a part of the features randomly selected for each model, and the prediction model M that finally outputs one prediction value may be constructed. Note that RandomForest can also be said to be a method of bagging.

[0072] The model construction unit 26 may construct the prediction model M by machine learning using boosting as machine learning including ensemble learning. In the machine learning using boosting, intermediate models Mtk (weak learners) based on the decision tree algorithm are sequentially generated. Here, k is an integer from 1 to K, and K is an integer of 2 or more. Assuming that the intermediate models Mtk are generated in order from k = 1 to K, in the machine learning using boosting, the intermediate model Mt1 is first generated. Then, paying attention to the error between the predicted value and the correct value in the intermediate model Mt1, the next intermediate model, the intermediate model Mt2, is generated. The one predicted value finally output by the prediction model M is a value obtained by aggregating the outputs of the intermediate models Mt1 to MtK by some operation.

[0073] The model construction unit 26 may construct the prediction model M composed of the intermediate models Mt1 to MtK, for example, by "Adaboost", which is a method of boosting. FIG. 4(b) schematically shows the process of constructing a model by Adaboost. The weight of the training data TD when generating the next intermediate model Mt2 is updated from the error between the predicted value (1) and the correct value in the previous intermediate model Mt1. For example, the larger the error of the data set, the larger the weight is updated. Thereafter, for example, the update of the weight and the generation of the intermediate model Mtk are repeated until K times, which is the set number of times.

[0074] The model construction unit 26 may construct the prediction model M by machine learning using gradient boosting as machine learning using boosting. That is, the model construction unit 26 may construct the prediction model M by machine learning using a gradient boosting decision tree. In a gradient boosting decision tree, boosting, gradient descent, and a decision tree are combined to construct the prediction model M.

[0075] FIG. 5 schematically shows the process of constructing a model by gradient boosting. In gradient boosting, first, an intermediate model Mt1 is generated, and the error between the predicted value (1) by the intermediate model Mt1 and the correct value is calculated. Then, an intermediate model Mt2 that can predict the error at the predicted value (1) is generated, and a correction value (2) for correcting the predicted value (1) is calculated from the intermediate model Mt2. Thereafter, for example, until the set number of times K, the calculation of the correction value (k) for correcting the predicted value (k) at that time is repeated. Note that the correction value (k) is also referred to as the gradient (k).

[0076] Examples of decision tree algorithms including gradient boosting include Catboost in addition to LightGBM and XGBoost described above. In XGBoost, as shown in FIG. 6(a), node branching (splitting) is performed level-wise. That is, the branching of all nodes is performed in order from the left. On the other hand, in LightGBM, as shown in FIG. 6(b), node branching (splitting) is performed leaf-wise. That is, branching is performed by narrowing down the nodes to be branched (for example, giving priority to nodes with a smaller loss). For nodes for which branching is no longer necessary, no further calculation is performed.

[0077] In addition, in LightGBM, the Histogram-based Algorithm is used when performing branching. Figure 7(a) schematically shows the judgment process when determining branching by the Pre-sorted Algorithm, which is different from the Histogram-based Algorithm. Figure 7(b) schematically shows the judgment process when determining branching by the Histogram-based Algorithm. In Figure 7(a) and Figure 7(b), each of "d1" to "d6" represents one piece of data (or one data set). In the Pre-sorted Algorithm, after checking each value in the data d1 to d6, it is determined where to perform the branching (division). On the other hand, in the Histogram-based Algorithm, the values in the data d1 to d6 are grouped by a histogram, and it is determined where to perform the branching (division) in units of groups (aggregates).

[0078] Furthermore, in LightGBM, the amount of data used is reduced during the training process by a method called GOSS (Gradient-based One-Side Sampling). Specifically, for data with large errors, all data are used as data that has not yet been trained, and for data with small errors, only a part is used assuming that it has been trained to a certain extent. In addition, in LightGBM, by a method called FEB (Exclusive Feature Bundling), among different types of features, multiple features that do not become zero at the same time (low correlation) and can be calculated together without problems are grouped and calculated as one.

[0079] <Specific examples of the load feature FL> Next, with reference to FIG. 8 as well, specific examples of one or more load characteristics FL included in the characteristic quantities F1 to FN will be described. The characteristic quantities F1 to FN include, for example, one or more load characteristics FL among the plurality of types of load characteristics FL exemplified below. FIG. 8(a) schematically shows time-series data related to the power load value of the mixer 114 in one batch of processing. The time-series data related to the power load value is, for example, data obtained by repeatedly measuring the power (kW) supplied to the mixer 114 at a predetermined sampling period.

[0080] In the graph representing the time-series data of FIG. 8(a), "time t1" indicates the point in time when the mixer 114 starts kneading, and "time t2" indicates the point in time when the mixer 114 finishes kneading. Time t1 corresponds to, for example, the point in time when the operation control unit 22 starts driving the stirring member 114a. Time t2 corresponds to, for example, the point in time when the operation control unit 22 determines that the conditions for ending the kneading by the mixer 114 are satisfied. In one example, the operation control unit 22 determines that the above conditions are satisfied when a predetermined time has elapsed since the start of driving the stirring member 114a, and stops driving the stirring member 114a.

[0081] One or more load characteristics FL may include the value at a predetermined point in time in the time-series data of the power load value. The predetermined point in time is a point in time preset by an operator or the like. One or more load characteristics FL may include one or more of the values of the initial value P1, the minimum value P2, the maximum value P3, and the final value P4. The initial value P1 is the power load value at the start of kneading in the time-series data. The minimum value P2 is the power load value at the point in time when it becomes the minimum after the time (time t1) when the initial value P1 is obtained. Note that the power load value (initial value P1) at the start of kneading may be the minimum. The maximum value P3 is the power load value at the point in time when it becomes the maximum in the time-series data. The final value P4 is the power load value at the end point of the time-series data.

[0082] When obtaining the load characteristic quantity FL, the period of the time-series data targeted is not limited to the period from time t1 to time t2. The period of the time-series data targeted when obtaining the load characteristic quantity FL may be a period from time t1 as the starting point to the point in time when an arbitrary elapsed time tp has elapsed from time t1. In the example shown in FIG. 8(a), the elapsed time tp corresponds to the difference between time t2 and time t1. In this case, the final value P4 corresponds to the power load value at the point in time when the kneading by the mixer 114 is completed.

[0083] Different from the time-series data shown in FIG. 8(a), as shown in FIG. 8(b), one or more values among the initial value P1, the minimum value P2, the maximum value P3, and the final value P4 may be obtained as the load characteristic quantity FL from the time-series data of the period from time t1 to a time before time t2. The starting point of the period of the time-series data targeted when obtaining the load characteristic quantity FL may be a time after time t1. For example, between the training phase and the evaluation phase, the starting point and the ending point of the period are set so that the period of the time-series data targeted when obtaining the load characteristic quantity FL is the same.

[0084] The one or more load characteristic quantities FL may include, instead of or in addition to the power load value at the above-mentioned predetermined point in time, a statistic obtained from the time-series data of the power load value. The statistic as the load characteristic quantity FL is one or more values selected from the group consisting of a variation range, a decline width, a total value, an average value, a standard deviation, a coefficient of variation, a median, a first quartile, a third quartile, kurtosis, and skewness. The variation range is obtained by the difference between the maximum value P3 and the minimum value P2. The decline width is obtained by the difference between the maximum value P3 and the final value P4.

[0085] Regarding statistics other than the range of variation and the rate of decline, they may be calculated from a data group of power load values included in at least a part of the time-series data (a set of power load values obtained for each sampling period). For example, the total value can be obtained by accumulating the power load values for each sampling period in at least a part of the time-series data. Note that, after thinning out a part of the data of the power load values for each sampling period, statistics such as the total value may be calculated. The time-series data for obtaining statistics etc. may be data representing the moving average of the power load values instead of the data representing the temporal change of the power load value itself. One or more load characteristics FL may include characteristics obtained from the waveform of the time-series data, or / and characteristics obtained from an image displaying the time-series data.

[0086] The sampling period when measuring the power load value may be between 0.01 and 1 second (for example, 0.1 second). The measurement of the power load value may start immediately after starting the mixing of the concrete material other than water, or immediately after starting the mixing after adding water to the concrete material other than water and starting the mixing. The number of power load values to be measured varies depending on the measurement interval of the power load value and the mixing time, but from the viewpoint of predicting the fresh properties with higher accuracy, it is preferably 1,000 or more, more preferably 2,000 or more, and particularly preferably 2,500 or more. The power load value can be measured by a device installed on the same power supply as the mixer 114.

[0087] One or more load characteristics FL may include the amount of change in the power load value, and may also include information (numerical data) representing the change pattern of the power load value. From the viewpoint of predicting the fresh properties with higher accuracy and at an earlier stage, after starting the mixing of water and the concrete material other than water, data regarding the power load value after the time when the power load value of the mixer 114 becomes stable (the time when the amount of change in the power load value becomes small) may be used as the load characteristic FL. In the mixing of concrete materials, the determination of the time when the power load value of the mixer 114 stabilizes varies depending on the water-cement ratio of the fresh concrete, etc. For example, if any of the following conditions (1) to (3) are satisfied, it may be determined that the power load value of the mixer has stabilized. (1) When, during a period in which the power load value has a decreasing tendency, the rate of change of the power load value at a predetermined interval (for example, 1 second) longer than the sampling period continues to be within the range of ±1% for a predetermined set time (for example, 3 seconds) or more. The rate of change of the above power load value is calculated, for example, by [P(t) - P(t - 1)] / P(t - 1), where P() is the power load value at each time, t is the current measurement time point, and t - 1 is the previous measurement time point. (2) If the water-cement ratio of the fresh concrete is about 50 to 70%, after about 30 seconds have elapsed since the start of mixing water and concrete materials other than water (for example, after dry mixing the concrete materials other than water in the mixer and then adding water to the mixer to start mixing). (3) When the water-cement ratio is reduced (to less than 50%) from the viewpoint of strength development, the above time becomes later, and in the case of high-strength concrete, after about 1 to 10 minutes have elapsed since the start of mixing water and concrete materials other than water.

[0088] <Specific examples of features other than the load feature quantity FL> The feature quantities F1 to FN may include, as information other than the load feature quantity FL, one or more feature quantities related to the time when the mixer 114 performs the mixing of concrete materials (hereinafter referred to as "time feature quantities"). For example, the one or more time feature quantities include at least one of the date and time when the mixer 114 performs the mixing of concrete materials, the cumulative number of batches in a day of mixing by the mixer 114, the time from an arbitrarily set reference time point to the time when the mixer 114 performs the mixing, and the cumulative number of batches of mixing by the mixer 114 from an arbitrarily set reference time point.

[0089] The date when the mixer 114 performs kneading represents the specific date when the mixer 114 produces fresh concrete, expressed as a combination of month and day. In this case, the specific month may be one feature quantity, and the specific day may be another feature quantity. The time when the mixer 114 performs kneading represents the specific time point when the mixer 114 produces fresh concrete, expressed as a combination of hour, minute, and second, or a combination of hour and minute. In this case, the specific hour may be one feature quantity, and the specific minute (or each of the specific minute and second) may be another feature quantity.

[0090] The above time (the specific time point when the mixer 114 produces fresh concrete) as a feature quantity may be the time when the concrete material is supplied to the mixer 114, or may be the time when the fresh concrete is discharged from the mixer 114. The above time (the specific time point when the mixer 114 produces fresh concrete) as a feature quantity may be the time when the driving of the stirring member 114a is started in the mixer 114, or may be the time when the driving of the stirring member 114a is ended.

[0091] Each of the month and day when the mixer 114 performs kneading, and each of the hour, minute, and second when the mixer 114 performs kneading may be specified using trigonometric functions. First, with reference to FIG. 9, an explanation will be given about specifying (representing) the "hour" among the times using trigonometric functions. Usually, the "hour" of a day returns to 0 o'clock when it reaches 24 o'clock, and is specified (represented) by 24 numerical values from 0 o'clock to 23 o'clock. When specifying using such normal numerical values, for example, between 0 o'clock and 23 o'clock, although there is only a difference of 1 hour without considering the date, numerically, they are far apart. Therefore, the "hour" may be specified (expressed) using trigonometric functions so as to have periodicity.

[0092] In one example, h o'clock (where h is an arbitrary integer from 0 to 23) is represented as a combination of cos{2π×(h / 24)} and sin{2π×(h / 24)}, as shown in FIG. 9. In this case, 0 o'clock is (1,0), 6 o'clock is (0,1), 12 o'clock is (-1,0), and 18 o'clock is (0,-1). In the input of the prediction model M, cos{2π×(h / 24)} may be one feature amount, and sin{2π×(h / 24)} may be another feature amount.

[0093] For m minutes (where m is an arbitrary integer from 0 to 59) and s seconds (where s is an arbitrary integer from 0 to 59), they may also be represented by the following combinations respectively. Also, in the input of the prediction model M, similar to the "hour", each of the two numerical values may be one feature amount. · m minutes: cos{2π×(m / 60)}, sin{2π×(m / 60)} · s seconds: cos{2π×(s / [60)}, sin{2π×(s / 60)}

[0094] When the "month" of the date is M months (where M is an arbitrary integer from 1 to 12) and the "day" is D days (where D is an arbitrary integer from 1 to Dm), M months and D days may be represented by the following combinations respectively. Dm is determined for each month and is an integer of either 28, 29, 30, or 31. In the input of the prediction model M, similar to the "hour", each of the two numerical values may be one feature amount. · M months: cos[2π×{(M - 1) / 12}], sin[2π×((M - 1) / 12)] · D days: cos[2π×{(D - 1) / Dm}], sin[2π×((M - 1) / Dm)]

[0095] The cumulative number of batches in a day regarding the kneading by the mixer 114 is information specifying which batch process was executed on that day (which fresh concrete was produced by the batch process executed at what order). Depending on which batch process of the day the fresh concrete being focused on was produced, it can represent at what timing the kneading by the mixer 114 was executed.

[0096] When the time from an arbitrarily set reference time point to the time when the mixer 114 performs kneading is used as a feature amount, the reference time point may be set by an operator such as a worker. The reference time point may be set at a time after cleaning inside the mixer 114 in the maintenance of the manufacturing apparatus 100 and before the manufacturing by the mixer 114 is restarted. Note that, in the case where the mixer 114 is cleaned after the operation is stopped every day, the time before the manufacturing by the mixer 114 is started for the first time in a day is also included in the time before the restart. The time between the reference time point and the time when the mixer 114 performs kneading (manufacture of fresh concrete) may be expressed in minutes or in seconds.

[0097] The time when the mixer 114 performs kneading (manufacture of fresh concrete) may be the time when the concrete material is supplied to the mixer 114, or may be the time when the fresh concrete after manufacture is discharged from the mixer 114. The time when the mixer 114 performs kneading (manufacture of fresh concrete) may be the time when the driving of the stirring member 114a is started in the mixer 114, or may be the time when the driving of the stirring member 114a is ended in the mixer 114. The quality prediction device 10 may measure the time from the reference time point to the time when the mixer 114 performs kneading.

[0098] The cumulative number of batches from an arbitrarily set reference time point is information for specifying which batch process is being executed from the reference time point (which batch process the fresh concrete is manufactured in). In the cumulative number of batches from the reference time point, the cumulative number of batches is reset to 0 at the reference time point. The reference time point when counting the cumulative number of batches may also be set at a time after cleaning inside the mixer 114 in the maintenance of the manufacturing apparatus 100 and before the manufacturing by the mixer 114 is restarted (including the time before the manufacturing by the mixer 114 is started for the first time in a day).

[0099] The characteristic quantities F1 to FN may include information related to kneading by the mixer 114 (information other than the power load value). As information related to kneading, in addition to the elapsed time tp, for example, the kneading amount of concrete for one batch, the kneading time, the time when the power load value reaches the maximum value, the time from when the power load value reaches the maximum value until the fresh concrete is discharged from the mixer 114, and the time from when the power load value reaches the maximum value to the measurement time point mp (final value P4) described later can be mentioned. The characteristic quantities F1 to FN may include the specified strength, information indicating the type of cement, information indicating the type of admixture, information indicating the type of fine aggregate, information indicating the type of coarse aggregate, and information indicating the addition amount of the additive. In the information indicating the type of cement or the like, the types of cement or the like may be classified in any way and then each type may be specified. The characteristic quantities F1 to FN may include numerical data obtained from an image of the fresh concrete during or immediately after production by the mixer 114, or numerical data obtained from an image of the fresh concrete after being discharged from the mixer 114. The characteristic quantities F1 to FN may include at least one of the target quality of the fresh concrete during use and the target quality of the fresh concrete at the time of shipment.

[0100] Other characteristic quantities other than the load characteristic quantity FL included in the characteristic quantities F1 to FN will be further exemplified. Note that some of the characteristic quantities also include those that overlap with the above-described examples. In addition to the load characteristic quantity FL, the characteristic quantities F1 to FN may include one or more types of data selected from data related to the mixing conditions of the fresh concrete, data related to the quality of the intended fresh concrete, data related to the cement as a concrete material, data related to concrete materials other than cement, data related to the kneading equipment for the concrete materials, and data related to the environment during the production (including transportation) of the fresh concrete.

[0101] (Data related to the mixing conditions of the fresh concrete) For the characteristic quantities F1 to FN, as data related to the mixing conditions of fresh concrete, (i) the types of materials used, and the place of origin (or product name), (ii) the amounts (mass, weight, volume) and at least one of the densities of cement, water, fine aggregate, coarse aggregate, various admixtures (ground granulated blast-furnace slag, fly ash, silica fume, expansive agent, fine powder of volcanic glass, crushed stone powder, metakaolin, etc.), various admixtures (AE agent, water reducing agent, AE water reducing agent, high-performance water reducing agent, high-performance AE water reducing agent, fluidizing agent, setting retarder, hardening accelerator, shrinkage reducing agent, etc.), and various fibers (steel fiber, glass fiber, carbon fiber, aramid fiber, nylon fiber, vinylon fiber, polyethylene fiber, polypropylene fiber, etc.), (iii) water-cement ratio, water-binder ratio, air content, fine aggregate ratio, bulk volume of coarse aggregate, maximum size of coarse aggregate, coarse grain ratio, actual volume ratio of grain shape determination, total alkali content in concrete, sludge solid content ratio, and recycled aggregate replacement ratio, and (iv) the amount of stabilizer used for adhering mortar and sludge water may be included.

[0102] (Data related to the quality of the target fresh concrete) For the characteristic quantities F1 to FN, as data related to the quality of the target fresh concrete, (i) the target slump, slump flow, air content, 500 mm flow arrival time, flow stop time of the flow, appearance (still image or moving image), temperature, unit water content, plastic viscosity, yield value, funnel flow time, compactability, deformability, fillability, and passing ability at the time of shipment, transportation, unloading, or placement of the fresh concrete, (ii) the strength of the concrete (design standard strength, durability design standard strength, quality standard strength, structural strength correction value (S value), mixing control strength, specified strength, etc.), and (iii) the static elastic modulus, coefficient of restitution, viscosity coefficient, chloride content in the concrete, length change rate, expansion rate, mass reduction rate, carbonation depth, chloride ion penetration depth, diffusion coefficient of chloride ions, porosity, air permeability coefficient, water permeability coefficient, air void spacing factor, electrical resistivity, dynamic elastic modulus, Poisson's ratio, creep coefficient, and color tone in design may include one or more types of data selected therefrom. The strength, slump, slump flow, air content, and chloride content in the concrete can be measured by the test methods described in JIS A 5308:2024 (Ready-mixed concrete).

[0103] (Data related to cement) For the characteristic quantities F1 to FN, as data related to cement, which is a concrete material, (a) data related to the whole cement, (b) data related to the raw materials of cement clinker, (c) data related to the firing conditions of cement clinker, (d) data related to the grinding conditions of cement, and (e) data related to cement clinker may include one or more types of data selected therefrom.

[0104] For the characteristic quantities F1 to FN, as data related to the whole cement in (a), (a1) the type, chemical composition, mineral composition, wet f.CaO, loss on ignition, Blaine specific surface area, particle size distribution, residue amount of sieve test, and color tone of the cement used as a concrete material, (a2) the mineralogical and crystallographic properties of each mineral contained in the cement, and (a3) the semi-hydration rate of gypsum contained in the cement may include one or more types of data selected therefrom.

[0105] For the characteristic quantities F1 to FN, as data related to the raw materials of cement clinker, (b1) the chemical composition, hydraulicity, residue on sieve test, Blaine specific surface area (powder fineness), loss on ignition, supply amount, supply amount of auxiliary raw materials (special raw materials such as waste), storage amount (remaining amount) in the blending silo, and storage amount (remaining amount) in the storage silo of the raw materials for blending cement clinker; (b2) the current value of the cyclone located between the raw material mill and the blending silo of the raw materials for blending (representing the rotation speed of the cyclone and having a correlation with the speed of the raw materials passing through the cyclone); (b3) the chemical composition and its hydraulicity of the raw materials of cement clinker at a point in time a predetermined time before charging into the kiln (for example, one point in time 5 hours before, or multiple points in time such as 3 hours before, 4 hours before, 5 hours before, and 6 hours before) (the raw materials for blending cement clinker from which fine particles etc. are extracted by the countercurrent air flow during transportation; hereinafter referred to as the raw materials for the kiln of cement clinker), and (b4) the chemical composition, hydraulicity, Blaine specific surface area, residue on sieve test, decarbonation rate, and moisture content of the raw materials obtained by mixing the raw materials for the kiln of cement clinker and auxiliary raw materials, one or more types of data selected from these may be included.

[0106] For the characteristic quantities F1 to FN, as data related to the firing conditions of cement clinker, (c1) the insertion amount of the raw materials of cement clinker into the kiln, kiln rotation speed, inlet temperature, firing zone temperature, temperature of cement clinker, average torque of the kiln, O2 concentration, and NO X concentration during the firing of cement clinker; (c2) the temperature of the clinker cooler; and (c3) the gas flow rate of the preheater (having a correlation with the temperature of the preheater), one or more types of data selected from these may be included.

[0107] For the characteristic quantities F1 to FN, as data related to the grinding conditions of cement, the grinding temperature, water spraying amount in the finishing mill, separator air volume, type of gypsum, addition amount of gypsum, input amount of cement clinker, rotation speed of the finishing mill, temperature of the powder discharged from the finishing mill, amount of the powder discharged from the finishing mill, and amount of the powder not discharged from the finishing mill, one or more types of data selected from these may be included.

[0108] Among the characteristic quantities F1 to FN, as data related to cement clinker, (e) one or more data selected from (e1) the mineral composition, chemical composition, wet f.CaO (free lime), and bulk density of cement clinker, (e2) the crystallographic properties (such as lattice constant and crystallite size) of each mineral contained in the cement clinker, and (e3) the ratio of two or more mineral compositions contained in the cement clinker may be included.

[0109] (Data related to concrete materials other than cement) Among the characteristic quantities F1 to FN, as data related to concrete materials other than cement, (i) for aggregates (at least one of fine aggregates and coarse aggregates), type, absolute dry density, apparent dry density, water absorption, moisture content, surface moisture ratio, loss mass fraction in the soundness test, abrasion loss, maximum size, particle size, coarse particle ratio, mass fraction of those remaining between consecutive sieves, particle shape determination actual volume ratio, fine particle content, clay lump content, organic impurities, chloride content, and classification by alkali-silica reactivity, (ii) for admixtures, type, density, specific surface area, residue on 45μm sieve, residue on 1.2mm sieve, flow value ratio, activity index, moisture content, expansibility (length change rate), and contents of various chemical components (silicon dioxide: SiO2, magnesium oxide: MgO, aluminum oxide: Al2O3, sulfur trioxide: SO3, free calcium oxide: f.CaO, free silicon: f.Si, etc.), (iii) for admixtures, type, components, density, appearance (color), chloride ion content, total alkali content, water reduction rate, flow value ratio, ratio of bleeding amount, difference in bleeding amount, difference in setting time (initial time, final time), compressive strength ratio, length change ratio, resistance to freeze-thaw (relative dynamic elastic modulus), and change amount over time, and (iv) for fibers, type, density, nominal diameter, nominal length, shape, fineness, tensile strength, tensile elastic modulus, attached moisture rate, melting temperature, and alkali resistance (strength retention rate), one or more data selected therefrom may be included.

[0110] (Data related to the mixing equipment of concrete materials) For the characteristic quantities F1 to FN, as data related to the concrete material mixing equipment, (i) for the mixer 114, data selected from the type, form, product name, manufacturer name, manufacturing year, rated capacity, total output, manufacturing capacity, body mass, no-load mass during operation, external dimensions, inclination angle (in the case of a drum type), rotation speed (drum, stirring blades, stirring shaft), and mixing performance (deviation rate of the air amount in the concrete, deviation rate of the mortar amount in the concrete, deviation rate of the coarse aggregate in the concrete, deviation rate of consistency (slump), deviation rate of compressive strength), and (ii) information regarding the material charging order, mixing amount, charging time, mixing time, discharging time, recharging time, and cycle time may be included as one or more types of data selected therefrom.

[0111] (Data related to the environment) For the characteristic quantities F1 to FN, as data related to the environment during the production (including transportation) of concrete, (i) outdoor temperature and humidity, atmospheric pressure, solar radiation amount, sunshine duration, rainfall amount, wind speed, wind direction, weather, climate, and weather conditions, (ii) temperature of each concrete material, temperature and humidity of the location where the material is stored (inside a container such as a silo or storage bottle), and temperature inside the mixer or mixer, and (iii) vehicle information of the truck agitator, loading capacity, transportation time, transportation distance, temperature and humidity inside the drum or drum, information regarding the road traffic conditions, and information regarding the vibration applied to the vehicle may be included as one or more types of data selected therefrom.

[0112] <Hardware configuration of the control device> As shown in FIG. 10, the quality prediction device 10 includes a circuit 50. The circuit 50 has a processor 51, a memory 52, a storage 53, an input / output port 54, and a timer 55. The storage 53 is composed of one or more non-volatile memory devices such as a flash memory or a hard disk. The storage 53 stores at least a quality prediction program for causing a computer to execute the construction process, acquisition process, and prediction process described later. The storage 53 stores a quality prediction program for configuring each functional block of the quality prediction device 10.

[0113] The memory 52 is composed of one or more volatile memory devices such as, for example, a random access memory. The memory 52 temporarily stores the quality prediction program loaded from the storage 53. The processor 51 is composed of one or more arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The processor 51 executes the quality prediction program loaded into the memory 52 to constitute each functional block of the quality prediction device 10. The calculation result by the processor 51 is temporarily stored in the memory 52. The input / output port 54 performs input / output of information with the input device 12, the monitor 14, the mixer 114, etc. in response to a request from the processor 51.

[0114] The timer 55 measures the elapsed time, for example, by counting reference pulses at a fixed period. Note that the circuit 50 is not necessarily limited to being configured by a program for each function. For example, the circuit 50 may configure at least part of the functions by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) integrating the same. The quality prediction program may be provided after being fixedly recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, the quality prediction program may be provided via a communication network as a data signal superimposed on a carrier wave.

[0115] [Method for manufacturing fresh concrete] Subsequently, an example of the method for manufacturing fresh concrete executed in the manufacturing system 1 will be described. The method for manufacturing fresh concrete includes a manufacturing process and a quality prediction process. The manufacturing process is a process for manufacturing fresh concrete. The quality prediction process is a process for predicting the fresh properties of fresh concrete after or during the execution of the manufacturing process. The quality prediction process may be executed during a period overlapping at least part of the period in which the manufacturing process is repeatedly executed.

[0116] The manufacturing process includes, for example, a conveying process, a metering process, a charging process, a mixing process, a discharging process, and a loading process. In the conveying process, various concrete materials are conveyed to the storage bottle 111 by the conveying device 104, and various materials are individually supplied to the storage bottle 111. In the metering process, various materials are individually supplied from the storage bottle 111 to the metering bottle 112, and various materials are metered in the metering bottle 112. In the metering process, for each material, when the measured amount reaches a predetermined set amount, the material is discharged into the collective hopper 113. In the charging process, after all types of materials are aggregated in the collective hopper 113, the materials in the collective hopper 113 are charged (supplied) into the mixer 114.

[0117] In the mixing process, a plurality of types of concrete materials are mixed in the mixer 114. In the mixing process, the quality prediction device 10 having a function of controlling the manufacturing device 100 may control the mixer drive unit 114b according to predetermined operating conditions. In the mixing process, the power supplied from the quality prediction device 10 to the mixer drive unit 114b may be adjusted so that the rotation speed of the mixer drive unit 114b follows the target rotation speed.

[0118] In the discharging process, after the mixing of the concrete materials in the mixer 114 is completed, the fresh concrete is discharged from the mixer 114 into the loading hopper 115. In the loading process, the fresh concrete discharged into the loading hopper 115 is loaded onto the transport vehicle 200.

[0119] The quality prediction process (method for predicting the quality of fresh concrete) includes a model construction process in the learning phase and a quality evaluation process in the evaluation phase. In the quality prediction process, the model construction process is executed before the quality evaluation process.

[0120] The model construction process includes a construction process. The construction process is a process of constructing a prediction model M by machine learning based on training data TD including a plurality of data sets. In the construction process, the prediction model M is constructed such that one or more features selected from among a plurality of features F1 to FN and the predicted value of the fresh property are associated with each other by a decision tree. The construction process may be executed by the model construction unit 26 of the quality prediction device 10. In each of the plurality of data sets in the training data TD, the training features F1 to FN and the correct value of the fresh property are associated with each other.

[0121] The quality evaluation process includes an acquisition process and a prediction process. The acquisition process is a process of acquiring evaluation features F1 to FN obtained when producing fresh concrete using the mixer 114. The evaluation features F1 to FN are the same type of data as the training features F1 to FN. The features F1 to FN include at least one or more load features FL related to the power load value of the mixer 114. The acquisition process may be executed by the feature acquisition unit 24 of the quality prediction device 10.

[0122] The prediction process is a process of acquiring a predicted value of the fresh property according to the evaluation features F1 to FN acquired in the acquisition process using the prediction model M constructed in the construction process. The acquisition process may be executed by the prediction unit 30 of the quality prediction device 10. The quality evaluation process may include an output process. The output process is a process of outputting the predicted value of the fresh property acquired in the prediction process to the monitor 14. The output process may be executed by the output unit 32 of the quality prediction device 10.

[0123] (Model construction process) FIG. 11(a) is a flowchart showing an example of a series of processes executed in the model construction process. This model construction process is executed before the production phase including the above-described production process in the manufacturing apparatus 100 is executed, or at the initial stage when the above-described production phase is started. In this model construction process, for example, the fresh concrete actually produced in the manufacturing apparatus 100 is used as the training fresh concrete.

[0124] In the model construction process, first, step S11 is executed. In step S11, for example, training data TD for performing machine learning is prepared by an operator such as a worker. The training data TD is composed of a plurality of data sets as described above. Each of the plurality of data sets in the training data TD includes feature quantities F1 to FN (input information for training) obtained when producing fresh concrete for training, and correct values of fresh properties (for example, slump) associated with the feature quantities F1 to FN.

[0125] The correct value of the fresh property may be a value obtained by actually measuring the fresh property of the fresh concrete for training. In one example, after the fresh concrete for training is loaded onto the transport vehicle 200, a part of the fresh concrete is extracted by an operator or the like. Then, the slump, slump flow, air content, etc. of the extracted fresh concrete are measured by an operator or the like, and at least a part of these measured values is used as the correct value in the training data TD.

[0126] Next, step S12 is executed. In step S12, for example, the model construction unit 26 of the quality prediction device 10 constructs a prediction model M by performing machine learning using the training data TD prepared in step S11. The model construction unit 26 may construct the prediction model M by machine learning using the decision tree algorithm. The model construction unit 26 may construct the prediction model M by performing machine learning involving ensemble learning in the machine learning using the decision tree algorithm. The model construction unit 26 may construct the prediction model M by performing machine learning using boosting (for example, gradient boosting) in the machine learning involving ensemble learning. In one example, the model construction unit 26 constructs the prediction model M using LightGBM developed by Microsoft (registered trademark).

[0127] The model construction unit 26 may construct one or more models among the prediction model M that outputs the predicted value of slump, the prediction model M that outputs the predicted value of slump flow, and the prediction model M that outputs the predicted value of air content. The model construction unit 26 may construct a prediction model M that outputs the ratio of slump flow to slump as the fresh property.

[0128] Next, step S13 is executed. In step S13, for example, the model holding unit 28 stores the prediction model M constructed in step S12. Thus, the model construction process ends.

[0129] (Quality evaluation process) FIG. 11(b) is a flowchart showing an example of a series of processes executed in the quality evaluation process. This quality evaluation process is performed, for example, during at least a part of the period in which the above-described manufacturing process for the fresh concrete to be evaluated is being executed.

[0130] In the quality evaluation process, first, the quality prediction device 10 executes step S21. In step S21, for example, the quality prediction device 10 waits until it becomes the evaluation timing, which is the timing for evaluating the quality of the fresh concrete to be evaluated. The above evaluation timing may be predetermined for a certain time zone within a day, or may be predetermined for the timing of executing a certain batch process within a day. The above evaluation timing may be the timing when an operator such as a worker receives an instruction to execute the evaluation.

[0131] Next, the quality prediction device 10 executes step S22. In step S22, for example, the feature amount acquisition unit 24 acquires the feature amounts F1 to FN when manufacturing the fresh concrete to be evaluated. The feature amounts F1 to FN acquired in step S22 are input data for evaluation in which the fresh property (for example, slump) is unknown.

[0132] Next, the quality prediction device 10 executes step S23. In step S23, for example, the prediction unit 30 predicts the fresh properties of the green concrete to be evaluated based on the evaluation feature amounts F1 to FN acquired in step S22 and the prediction model M held in the model holding unit 28. In one example, the prediction unit 30 inputs the evaluation feature amounts F1 to FN acquired in step S22 into the prediction model M and obtains the predicted value of the fresh properties output from the prediction model M.

[0133] Next, the quality prediction device 10 executes step S24. In step S24, for example, the output unit 32 causes the monitor 14 to display the predicted value of the fresh properties acquired in step S23. Thereby, an operator such as a worker can confirm the prediction result regarding the fresh properties of the green concrete to be evaluated.

[0134] Thus, the quality evaluation process ends. The quality prediction device 10 may execute a series of processes of steps S21 to S24 each time the production of green concrete for one batch is performed (for each single batch process). The quality prediction device 10 may execute a series of processes of steps S21 to S24 each time the production of green concrete for a plurality of batches is performed (for each plurality of batch processes).

[0135] [Modification Example] The series of processes shown in FIGS. 11(a) and 11(b) is an example and can be changed as appropriate. In the above series of processes, one step and the next step may be executed in parallel, and some steps may be executed in an order different from the example described above. Instead of at least some of the steps of the above series of processes, or in addition to the above series of processes, steps having contents different from the example described above may be executed.

[0136] In addition to predicting the quality of fresh concrete by the quality prediction device 10, in the manufacturing device 100, the fresh properties of fresh concrete may be periodically measured (for example, 1 to 100 times a day). In this case, the prediction model M may be updated based on the measured values of the fresh properties of fresh concrete and the feature amounts F1 to FN when the measured values are obtained. In the above prediction step, the fresh properties of fresh concrete may be predicted by the updated prediction model M. Even when the updated prediction model M is used, the step of predicting the fresh properties of fresh concrete based on the prediction model M and the feature amounts F1 to FN obtained in the above acquisition step will surely be performed.

[0137] In the above example, the quality prediction device 10 predicts the quality of fresh concrete after being obtained by the mixing by the mixer 114 of the manufacturing device 100 and before being shipped to the site. The timing at which the quality is predicted by the quality prediction device 10 is not limited to this example. The quality prediction device 10 may predict the quality of fresh concrete during the mixing by the mixer 114 (fresh concrete already obtained before the mixing is completed). The quality prediction device 10 may predict the quality of fresh concrete during transportation to the site where fresh concrete is used or the quality of fresh concrete at the time of arrival at the site (at the time of unloading). For example, the transport vehicle 200 is provided with a mixer 214 for stirring fresh concrete (see FIG. 1). The mixer 214 in the transport vehicle 200 such as an agitator truck is also referred to as a drum.

[0138] When the transport vehicle 200 is included in the manufacturing system 1, at least a part of the feature quantities F1 to FN may include the feature quantities obtained when producing (mixing) fresh concrete using the mixer 214. One or more load feature quantities FL may be one or more feature quantities related to the power load value of the mixer 214 when mixing fresh concrete. The feature quantity acquisition unit 24 of the quality prediction device 10 may calculate one or more load feature quantities FL from the time-series data up to the measurement time point mp during the operation of mixing by the mixer 214. The prediction unit 30 may predict the quality of the fresh concrete at the time after being discharged from the mixer 214 while the mixing operation by the mixer 214 is continuing.

[0139] The feature quantity acquisition unit 24 of the quality prediction device 10 may acquire, as at least a part of the feature quantities F1 to FN (one or more load feature quantities FL), the feature quantities related to the power load value of the mixer 214 when mixing fresh concrete. The feature quantities F1 to FN may include, as data related to the transportation of fresh concrete, (i) the capacity, thickness, mass, and temperature of the fresh concrete in the mixer 214 of the transport vehicle 200, (ii) the outside air temperature during transportation, and (iii) one or more types of data selected from the transportation time (the time from the end of mixing to the end of transportation (unloading time)) and the transportation distance. The feature quantities F1 to FN may include batch data (heating / cooling device type, stored water temperature, stored aggregate temperature, stored cement temperature, metering bottle type, discharge amount, maximum discharge pressure, etc.). The quality prediction device 10 may not have a function of controlling the manufacturing device 100.

[0140] In the above example, the fresh concrete is manufactured at a location separate from the construction site (manufacturing apparatus 100). The location where the fresh concrete is manufactured is not limited to this example. The concrete materials (for example, concrete materials excluding water) may be transported to the construction site by a transport vehicle, and then the fresh concrete may be manufactured at the construction site. In that case, water may be added to the concrete materials transported by the transport vehicle by a mixer provided on the transport vehicle, and then the materials may be kneaded to obtain fresh concrete. When the fresh concrete is manufactured at the construction site, the quality prediction device 10 may predict the quality such as the fresh properties of the fresh concrete at that construction site. The feature amount acquisition unit 24 of the quality prediction device 10 may acquire, as at least a part of one or more load feature amounts FL, a feature amount related to the power load value of the mixer provided on the transport vehicle when manufacturing the fresh concrete at the construction site.

[0141] In one of the various examples described above, at least a part of the matters described in other examples may be combined.

[0142] [Verification of Prediction Results by Prediction Model] Subsequently, using a data set with known correct values, the results of verifying the prediction results of the fresh properties by the prediction model M constructed by the decision tree algorithm with the feature amounts F1 to FN as inputs will be described. In this verification, using the same training data TD, a comparison was made with the prediction results of a comparative model constructed by a deep neural network (DNN: Deep Neural Network) different from the decision tree algorithm.

[0143] 120,552 training data sets were prepared as training data TD, and 13,383 test data sets with known correct values for evaluating prediction results were prepared. For each of the training data set and the test data set, feature quantities F1 to FN and the correct value of the fresh property are associated. As the feature quantities F1 to FN and the correct value (measured value) of the fresh property, data obtained from a fresh concrete manufacturing system configured in the same manner as the manufacturing system 1 was used. In addition, the fresh property of the fresh concrete before shipment was set as the prediction target by the prediction model.

[0144] As the feature quantities F1 to FN, the data items shown in Table 1 below were used. In Table 1, "manufacturing month (sin, cos)" means data representing the manufacturing month using the sine and cosine trigonometric functions, and the same meaning applies to the manufacturing date and the like. "Cement type (a1, a2, a3, a4, a5)" means data indicating which of the cement types a1 to a5 is included, and the same meaning applies to the admixture type and the like.

Table 1

[0145] (Example 1: LightGBM) A prediction model M was constructed by LightGBM, which is a type of decision tree algorithm. Specifically, the "LightGBM (version 4.3.0)" provided by Microsoft was used to construct the prediction model M. A model for predicting slump, a model for predicting slump flow, and a model for predicting air content were constructed individually. When constructing the prediction model M according to Example 1, the main hyperparameters were set as shown in Table 2 below.

Table 2

[0146] (Example 2: XGBoost) A prediction model M was constructed using XGBoost, which is a type of decision tree algorithm. Specifically, the prediction model M was constructed using the well-known library "XGBoost (version 1.5.2)". A model for predicting slump, a model for predicting slump flow, and a model for predicting air content were constructed individually. When constructing the prediction model M according to Example 2, the main hyperparameters were set as shown in Table 3 below. [Table 3]

[0147] (Comparative Example: DNN) A comparative prediction model was constructed using a deep neural network. When constructing the comparative prediction model, the main conditions (network configuration) were set as follows. · The Batch Normalization layer for normalizing the input data to the model was not used. · Dropout, which causes some connections in the intermediate layer of the neural network to be missing, was not used. · The intermediate layer of the neural network consisted of 8 fully connected layers. · Adam was used as the weight update formula, and HuberLoss was used as the loss function. · Machine learning was performed so that three types of data, namely slump, slump flow, and air content, were output as fresh properties from one prediction model.

[0148] (Verification Results) For each of Example 1, Example 2, and the Comparative Example, the predicted values obtained by the prediction model were compared with the correct values included in the test dataset using the test dataset. Specifically, when the predicted value obtained by the model was within the range obtained by adding a predetermined tolerance to the correct value in the test dataset, it was defined as correct, and the correct rate, which represents the ratio of the number of correctly predicted datasets to the total number of test datasets, was calculated as the evaluation index.

[0149] Fig. 12(a) shows the evaluation results of the correct answer rate regarding slump. In Fig. 12(a), when the tolerance is “±0.5 cm”, a data set in which the predicted value by the model is included in the range of ±0.5 cm of the correct value is regarded as correct, and the ratio of the data set regarded as correct is shown as the correct answer rate (%). Similarly, the correct answer rate (%) is calculated when the tolerances are “±1.0 cm”, “±1.5 cm”, “±2.0 cm”, and “±2.5 cm”. From the results shown in Fig. 12(a), it can be seen that the correct answer rate is significantly higher when predicting slump using the prediction models M according to Example 1 and Example 2 as compared with the comparative example.

[0150] Fig. 12(b) shows the evaluation results of the correct answer rate regarding slump flow. In Fig. 12(b), when the tolerance is “±2.5 cm”, a data set in which the predicted value by the model is included in the range of ±2.5 cm of the correct value is regarded as correct, and the ratio of the data set regarded as correct is shown as the correct answer rate (%). Similarly, the correct answer rate (%) is calculated when the tolerances are “±3.0 cm”, “±4.0 cm”, “±5.0 cm”, “±6.0 cm”, and “±7.5 cm”. From the results shown in Fig. 12(b), it can be seen that the correct answer rate is significantly higher when predicting slump flow using the prediction models M according to Example 1 and Example 2 as compared with the comparative example.

[0151] Fig. 13(a) shows the evaluation results of the correct answer rate regarding the air content. In Fig. 13(a), when the tolerance is “±0.3%”, a data set in which the predicted value by the model is included in the range of ±0.3% of the correct value is regarded as correct, and the ratio of the data set regarded as correct is shown as the correct answer rate (%). Similarly, the correct answer rate (%) is calculated when the tolerances are “±0.5%”, “±0.7%”, “±1.0%”, “±1.2%”, and “±1.5%”. From the results shown in Fig. 13(a), it can be seen that the correct answer rate is significantly higher when predicting the air content using the prediction models M according to Example 1 and Example 2 as compared with the comparative example.

[0152] As another evaluation index different from the correct answer rate, the "sum of the mean value and the standard deviation" regarding the error between the predicted value and the correct answer value was calculated. Using the sum of the mean value and the standard deviation as an evaluation index means evaluating the error taking into account the variation. That is, even if the error itself (the mean value of the errors) is small, if the standard deviation is large, it means that the variation of the errors is large, which is not preferable. Also, even if the standard deviation of the errors is small, if the mean value is large, it means that the original error is large, which is not preferable. In the calculation of the sum of the mean value and the standard deviation, for each test dataset, the error between the predicted value by the model and the correct answer value was calculated, and the mean value and the standard deviation were obtained from these errors.

[0153] Figure 13(b) shows the results of calculating the sum of the mean value and the standard deviation of the errors for each of slump, slump flow, and air content. From the results shown in Figure 13(b), it can be seen that in any of the fresh properties of slump, slump flow, and air content, when predicting by the prediction model M according to Example 1 and Example 2, the sum of the mean value and the standard deviation of the errors is significantly smaller than that of the comparative example.

[0154] From another perspective of prediction accuracy, after the preparation of the training data TD and the setting of various hyperparameters were completed, the "learning time" from when the computer started training until the construction of the prediction model was completed was evaluated. The learning time can be said to be the time required for the computer to construct the prediction model. The evaluation results of the learning time are shown in Table 4 below.

Table 4

[0155] In Examples 1 and 2 in Table 4, the learning time for constructing the prediction model M that only predicts slump is shown, and in the comparative example, the learning time for constructing the prediction model that simultaneously outputs slump, slump flow, and air content is shown. From the results shown in Table 4, it can be seen that in Examples 1 and 2, the learning time can be significantly shortened compared to the comparative example. Note that although the learning time in the comparative example is assumed to not differ significantly from the learning time when constructing a comparative model that only predicts slump using the same method as the comparative example, even if the learning time in Examples 1 and 2 is tripled, it is significantly shortened compared to the comparative example. Also, it can be seen that in Example 1, the learning time can be further shortened compared to Example 2. For example, by being able to shorten the learning time, it is possible to construct or update the prediction model M while manufacturing fresh concrete (it is possible to continue manufacturing fresh concrete without being restricted by the construction or update of the prediction model M).

[0156] (Consideration of the factors for improved prediction accuracy compared to the comparative example) [Prediction model based on decision tree] The prediction models M constructed in Examples 1 and 2 above are both constructed by the decision tree algorithm. In a decision tree, since branching is performed by focusing on the part with the most difference in features at each node, if there is a relationship between the features and the data to be predicted, even if the number of training data is limited, a certain degree of accuracy can be ensured. Also, since the structure of the model is simple, overfitting is less likely to occur even when the number of data is small. On the other hand, in the DNN according to the comparative example, complex learning is performed using all the data of the features in the training data, and it is good at capturing fine features and patterns of the data. A DNN having such features requires an enormous number of data to ensure accuracy. Also, since the structure of the model is complex, it is likely to overfit to the training data when the number of data is small. In machine learning, it is practically difficult to prepare training data having an enormous number of data, and in prediction using limited data, it is considered that a decision tree can construct a prediction model with higher accuracy.

[0157] [Ensemble learning using decision tree] The prediction models M constructed in the above-mentioned Example 1 and Example 2 are both constructed by a decision tree algorithm that includes ensemble learning. Ensemble learning using decision trees includes a method called "bagging" in which a plurality of decision trees with different characteristics are arranged in parallel and learned independently, or a method called "boosting" in which a plurality of decision trees with different characteristics are arranged in series and learned in order. Since these methods construct a prediction model using a plurality of decision trees, they have the effect of reducing the variation in prediction results and the error between the predicted value and the correct value. Therefore, it is presumed that the prediction accuracy has been improved by performing ensemble learning using a plurality of decision trees as compared with learning by a single DNN. Note that the ensemble learning of DNN may be a demerit in that it requires a great deal of effort to construct the model because the structure of the model becomes more complex and it may be difficult to appropriately combine models and adjust parameters.

[0158] 〔Gradient Boosting Decision Tree〕 The prediction model M constructed in the above-mentioned Example 1 is constructed by LightGBM, which is a type of machine learning including gradient boosting. (1) LightGBM is good at capturing the synergistic effect when numerical data between each parameter (between each feature amount) are combined. On the other hand, DNN is not suitable for capturing the mutual relationship between data. Considering that the fresh property, which is an example of quality, is determined by the combination of the concrete mix, date, load value, etc., it is presumed that the prediction accuracy has been improved by constructing the model with LightGBM, which can capture the interaction. (2) LightGBM can select important features of data for learning. Specifically, for features with large prediction errors, all data are used assuming they have important features, and for features with small prediction errors, data are randomly extracted as learned and re - learned to create a prediction model. On the other hand, DNN uses all data related to each feature quantity to perform complex learning and is good at capturing fine features and patterns in the data. Such a DNN with these features is effective for image, video, and audio data, but requires a huge amount of data for accuracy improvement. In prediction using limited numerical data, it is considered that a highly accurate prediction model can be created by selecting important features of the data for learning.

[0159] [Summary of the present disclosure] The fresh concrete quality prediction method described above is a quality prediction method for predicting the quality of fresh concrete. This quality prediction method includes a construction step of constructing a prediction model (M) by machine learning based on training data (TD) including a plurality of data sets, an acquisition step of acquiring a plurality of feature quantities (F1 to FN) obtained when producing fresh concrete using a mixer (114, 214), and a prediction step of using the prediction model (M) constructed in the construction step to obtain a predicted value of the quality corresponding to the plurality of feature quantities (F1 to FN) acquired in the acquisition step. In each of the plurality of data sets, a plurality of feature quantities (F1 to FN) and the correct value of the quality are associated. The plurality of feature quantities (F1 to FN) include one or more load feature quantities (FL) related to the power load value of the mixer (114, 214) when mixing concrete materials or stirring fresh concrete. In the construction step, the prediction model (M) is constructed such that one or more feature quantities selected from the plurality of feature quantities (F1 to FN) and the predicted value of the quality are associated by a decision tree.

[0160] As described in the above verification results, by using a prediction model (M) constructed such that one or more feature quantities selected from a plurality of feature quantities (F1 to FN) and a predicted value of quality are associated with each other by a decision tree, it has been confirmed that the prediction accuracy by the prediction model can be significantly improved. That is, the above quality prediction method for predicting the quality of fresh concrete using the prediction model (M) is useful for improving the prediction accuracy. Further, since a model by a decision tree is constructed, the time for constructing the prediction model can be shortened.

[0161] In the quality prediction method of fresh concrete described above, in the construction process, the prediction model (M) may be constructed by machine learning including ensemble learning. In this case, since a prediction model (M) is constructed in which a plurality of models are fused to obtain one final predicted value, it is more useful for improving the prediction accuracy.

[0162] In the quality prediction method of fresh concrete described above, in the construction process, the prediction model (M) may be constructed by machine learning using boosting. In this case, since a plurality of models generated in order taking into account the error are fused to construct a prediction model (M) that obtains one final predicted value, the prediction accuracy by the model using a decision tree can be improved.

[0163] In the quality prediction method of fresh concrete described above, in the construction process, the prediction model (M) may be constructed by machine learning using gradient boosting. In this case, since a plurality of models generated in order to reduce the error are fused to construct a prediction model (M) that obtains one final predicted value, the prediction accuracy by the model using a decision tree can be further improved.

[0164] In the method for predicting the quality of fresh concrete described above, the one or more load characteristics (FL) may include at least one of the value at a predetermined time point in the time-series data of the power load value of the mixer (114, 214) when mixing the concrete material or stirring the fresh concrete, and the statistic obtained from the above time-series data. The above statistic may be one or more values selected from the group consisting of the difference between the maximum value and the minimum value (range of variation), the difference between the maximum value and the final value (drop width), the total value, the average value, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness. The power load value when mixing or stirring by the mixer varies with time. By using at least one of the power load value at a specific time point and the statistic obtained from the time-series data as a characteristic as in the above method, information that more reflects the characteristics of the power load value that varies with time can be added to the input of the prediction model (M).

[0165] In the method for predicting the quality of fresh concrete described above, the plurality of characteristics (F1 to FN) may further include one or more time characteristics related to the time when the mixer (114) performs the mixing of the concrete material. The one or more time characteristics may include at least one of the date and time when the mixer (114) performs the mixing of the concrete material, the cumulative number of batches in one day of the mixing by the mixer (114), the time from an arbitrarily set reference time point to the time when the mixer (114) performs the mixing, and the cumulative number of batches of the mixing by the mixer (114) from an arbitrarily set reference time point. As a result of the verification by the present inventors, it was confirmed that even when the manufacturing conditions including the operating conditions of the mixer (114) are the same, the power load value itself can vary depending on the timing of manufacturing the fresh concrete. In the above method, by adding the above time characteristics to the input to the prediction model (M), the prediction model (M) can be constructed taking into account the variation in the power load value depending on the timing of manufacturing the fresh concrete. As a result, the variation in the prediction results by the prediction model (M) that may occur due to the difference in the manufacturing time can be reduced.

[0166] The quality prediction program described above is a program that causes a computer to execute the above quality prediction method. Since the above quality prediction method is executed by this quality prediction program, it is useful for improving prediction accuracy in the same manner as the above quality prediction method.

[0167] The method for manufacturing fresh concrete described above includes a manufacturing process for manufacturing fresh concrete and a quality prediction process for predicting quality by the above quality prediction method after or during the execution of the manufacturing process. In this method for manufacturing fresh concrete, since quality is predicted by the above quality prediction method, it is useful for improving prediction accuracy in the same manner as the above quality prediction method.

[0168] The fresh concrete quality prediction device (10) described above is a device for predicting the quality of fresh concrete. This quality prediction device (10) includes a model construction unit (26) that constructs a prediction model (M) by machine learning based on training data (TD) including a plurality of data sets, an acquisition unit (24) that acquires a plurality of feature quantities (F1 to FN) obtained when manufacturing fresh concrete using a mixer (114, 214), and a prediction unit (30) that uses the prediction model (M) constructed by the model construction unit (26) to acquire a predicted value of quality corresponding to the plurality of feature quantities (F1 to FN) acquired by the acquisition unit (24). In each of the plurality of data sets, a plurality of feature quantities (F1 to FN) and a correct value of quality are associated with each other. The plurality of feature quantities (F1 to FN) include one or more load feature quantities (FL) related to the power load value of the mixer (114, 214) when mixing concrete materials or stirring fresh concrete. The model construction unit (26) constructs the prediction model (M) such that one or more feature quantities selected from the plurality of feature quantities (F1 to FN) and the predicted value of quality are associated with each other by a decision tree. This quality prediction device is useful for improving prediction accuracy in the same manner as the above quality prediction method.

[0169] The fresh concrete manufacturing system (1) described above includes a manufacturing apparatus (100) for manufacturing fresh concrete and the quality prediction apparatus (10). The quality prediction apparatus (10) predicts the quality of the fresh concrete manufactured by the manufacturing apparatus (100). Since this manufacturing system includes the quality prediction apparatus, it is useful for improving the prediction accuracy in the same manner as the quality prediction apparatus.

Explanation of Signs

[0170] 1... Manufacturing system, 10... Quality prediction apparatus, 24... Feature quantity acquisition unit, F1 to FN... Feature quantities, FL... Load feature quantity, 26... Model construction unit, TD... Training data, 30... Prediction unit, M... Prediction model, 100... Manufacturing apparatus, 114... Mixer, 114a... Stirring member, 114b... Mixer drive unit, 214... Mixer.

Claims

1. A quality prediction method for predicting the quality of fresh concrete, comprising: a construction step of constructing a prediction model by machine learning based on training data including a plurality of data sets; an acquisition step of acquiring a plurality of feature quantities obtained when producing fresh concrete using a mixer; a prediction step of obtaining a predicted value of the quality corresponding to the plurality of feature quantities acquired in the acquisition step by using the prediction model constructed in the construction step, in each of the plurality of data sets, the plurality of feature quantities and the correct value of the quality are associated with each other; the plurality of feature quantities include one or more load feature quantities related to the power load value of the mixer when mixing concrete materials or stirring fresh concrete, and one or more time feature quantities related to the time when the mixer executes the mixing of concrete materials; in the construction step, the prediction model is constructed such that one or more feature quantities selected from the plurality of feature quantities and the predicted value of the quality are associated with each other by a decision tree; the one or more time feature quantities include the date and time when the mixer executes the mixing of concrete materials, the cumulative number of batches in one day of the mixing by the mixer, the time from an arbitrarily set reference time point to the time point when the mixer executes the mixing, and the cumulative number of batches of the mixing by the mixer from an arbitrarily set reference time point, including at least one of them; A quality prediction method for fresh concrete.

2. In the construction step, the prediction model is constructed by machine learning including ensemble learning. The quality prediction method for fresh concrete according to Claim 1.

3. In the construction step, the prediction model is constructed by machine learning using boosting. The quality prediction method for fresh concrete according to Claim 2.

4. In the construction step, the prediction model is constructed by machine learning using gradient boosting. The quality prediction method for fresh concrete according to Claim 3.

5. The one or more load feature quantities include the value at a predetermined time point in the time series data of the power load value of the mixer when mixing concrete materials or stirring fresh concrete, and at least one of the statistical quantities obtained from the time series data. The statistic is one or more values selected from the group consisting of the difference between the maximum value and the minimum value, the difference between the maximum value and the final value, the total value, the average value, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness. The method for predicting the quality of fresh concrete according to any one of claims 1 to 4.

6. A quality prediction program for causing a computer to execute the quality prediction method according to any one of claims 1 to 4.

7. A manufacturing process for manufacturing fresh concrete, A quality prediction process for predicting the quality by the quality prediction method according to any one of claims 1 to 4 after or during the execution of the manufacturing process, A method for manufacturing fresh concrete.

8. A quality prediction device for predicting the quality of fresh concrete, A model construction unit that constructs a prediction model by machine learning based on training data including a plurality of data sets, An acquisition unit that acquires a plurality of feature amounts obtained when manufacturing fresh concrete using a mixer, A prediction unit that acquires a predicted value of the quality corresponding to the plurality of feature amounts acquired by the acquisition unit using the prediction model constructed by the model construction unit, In each of the plurality of data sets, the plurality of feature amounts and the correct value of the quality are associated with each other, The plurality of feature amounts include one or more load feature amounts related to the power load value of the mixer when kneading concrete materials or stirring fresh concrete, and one or more time feature amounts related to the time when the mixer executes kneading of concrete materials, The model construction unit constructs the prediction model so that one or more feature amounts selected from the plurality of feature amounts and the predicted value of the quality are associated with each other by a decision tree, The one or more time feature amounts are The date and time when the mixer executes kneading of concrete materials, The cumulative number of batches in one day of kneading by the mixer, The time from an arbitrarily set reference time point to the time when the mixer executes kneading, and The cumulative number of batches of kneading by the mixer from an arbitrarily set reference time point, including at least one of them. A quality prediction device for fresh concrete.

9. A manufacturing device for manufacturing fresh concrete, The quality prediction device according to claim 8, The quality prediction device predicts the quality of the fresh concrete manufactured by the manufacturing device. A manufacturing system for fresh concrete.

Images