Method for predicting the quality of ready-mixed concrete, quality prediction program, method for manufacturing ready-mixed concrete, device for predicting the quality of ready-mixed concrete, and system for manufacturing ready-mixed concrete.

Abstract

To improve the convenience of predicting quality or using quality prediction results. [Solution] The method for predicting the quality of ready-mixed concrete includes: a first acquisition step of acquiring one or more load features from time-series data of the power load value of a mixer up to a measurement point while the mixer is still operating to mix concrete materials or stir ready-mixed concrete; a second acquisition step of acquiring input information including the one or more load features and the time from a predetermined reference point to the measurement point; and a prediction step of acquiring a predicted value of the quality according to the input information acquired in the second acquisition step, using a prediction model that has been pre-constructed by machine learning to output the quality of ready-mixed concrete in response to the input information.

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, a quality prediction device for 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, a quality prediction device for fresh concrete, and a manufacturing system for fresh concrete, which are useful for improving convenience.

Means for Solving the Problems

[0005] [1] A method for predicting the quality of ready-mixed concrete, comprising: a first acquisition step of acquiring one or more load features from time-series data of the power load value of a mixer up to a measurement point while the mixer is in operation for mixing concrete materials or stirring ready-mixed concrete; a second acquisition step of acquiring input information including the one or more load features and the time from a predetermined reference point to the measurement point; and a prediction step of acquiring a predicted value of the quality according to the input information acquired in the second acquisition step, using a prediction model that has been pre-constructed by machine learning to output the quality of ready-mixed concrete in response to the input information.

[0006] [2] The method for predicting the quality of ready-mixed concrete according to [1] above, wherein the quality for which the predicted value is obtained in the prediction step is the quality of ready-mixed concrete after the operation of the mixer has finished and the concrete has been discharged from the mixer.

[0007] [3] A method for predicting the quality of ready-mixed concrete according to [1] or [2] above, further comprising an output step of outputting the predicted quality value obtained in the prediction step, wherein the prediction step and the output step are performed while the mixer for ready-mixed concrete to be evaluated is in operation.

[0008] [4] A method for predicting the quality of ready-mixed concrete according to any one of [1] to [3] above, wherein in the first acquisition step, one or more load features are acquired from the time series data from the start of operation of the mixer to the measurement time, or one or more load features are acquired from the time series data from a predetermined time before the measurement time to the measurement time.

[0009] [5] A method for predicting the quality of ready-mixed concrete according to any one of [1] to [4] above, further comprising a construction step of constructing the prediction model by performing machine learning using training data composed of multiple training datasets, wherein in each of the multiple training datasets, the input information is associated with the correct value of the quality, and the multiple training datasets include two or more datasets in which the correct value of the quality is common and the measurement time points are different from each other.

[0010] [6] The method for predicting the quality of ready-mixed concrete according to any one of [1] to [5] above, wherein the one or more load features include at least one of the values ​​at a predetermined time in the time series data up to the measurement time, and a statistic obtained from the time series data up to the measurement time, the statistic being 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 sum, the mean, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness.

[0011] [7] A method for predicting the quality of ready-mixed concrete according to any one of [1] to [6] above, further comprising a construction step of constructing the prediction model by performing machine learning using training data composed of multiple training datasets, wherein the input information is a plurality of features including one or more load features and the time from the reference time to the measurement time, in each of the plurality of training datasets, the plurality of features and the correct value of the quality are associated, and in the construction step, the prediction model is constructed such that one or more features selected from the plurality of features and the predicted value of the quality are associated by a decision tree.

[0012] [8] A quality prediction program that causes a computer to execute one of the quality prediction methods described in [1] to [7] above.

[0013] [9] A method for manufacturing ready-mixed concrete, comprising a manufacturing step for manufacturing ready-mixed concrete, and a quality prediction step for predicting the quality of the ready-mixed concrete manufactured by the manufacturing step using the quality prediction method described in any one of [1] to [7] above.

[0014]

[10] A ready-mix concrete quality prediction device comprising: a first acquisition unit that acquires one or more load features from time-series data of the power load value of a mixer up to a measurement point while the mixer is in operation for mixing concrete materials or stirring ready-mix concrete; a second acquisition unit that acquires input information including the one or more load features and the time from a predetermined reference point to the measurement point; and a prediction unit that acquires a predicted value of the quality according to the input information acquired by the second acquisition unit, using a prediction model that has been constructed in advance by machine learning to output the quality of ready-mix concrete in response to the input information.

[0015]

[11] A ready-mix concrete manufacturing system comprising a manufacturing apparatus for manufacturing ready-mix concrete and a quality prediction apparatus as described in

[10] above, wherein the quality prediction apparatus predicts the quality of ready-mix concrete manufactured by the manufacturing apparatus. [Effects of the Invention]

[0016] According to this disclosure, a method for predicting the quality of ready-mixed concrete, a quality prediction program, a method for producing ready-mixed concrete, a device for predicting the quality of ready-mixed concrete, and a system for producing ready-mixed concrete are provided that are useful for improving convenience. [Brief explanation of the drawing]

[0017] [Figure 1] Figure 1 is a schematic diagram showing an example of a ready-mix concrete manufacturing system. [Figure 2] Figure 2 is a block diagram showing an example of the functional configuration of a quality prediction device. [Figure 3] Figures 3(a) and 3(b) illustrate data related to power load values. [Figure 4]FIG. 4 is a diagram for explaining an example of a process for calculating a load characteristic amount. [Figure 5] FIG. 5 is a diagram for explaining an example of a process for calculating a load characteristic amount. [Figure 6] FIG. 6 is a diagram for explaining an example of a prediction model. [Figure 7] FIG. 7 is a diagram for explaining an example of a process for constructing a prediction model. [Figure 8] FIG. 8 is a diagram for explaining an example of a prediction model. [Figure 9] FIG. 9 is a block diagram showing an example of a hardware configuration of a quality prediction device. [Figure 10] FIG. 10 is a flowchart showing an example of a processing flow executed in a training phase. [Figure 11] FIG. 11 is a flowchart showing an example of a processing flow executed in an evaluation phase. [Figure 12] FIGS. 12(a), 12(b), and 12(c) are graphs exemplifying results of verifying prediction accuracy by a prediction model. [Figure 13] FIGS. 13(a), 13(b), and 13(c) are graphs exemplifying results of verifying prediction accuracy by a prediction model. [Figure 14] FIGS. 14(a), 14(b), and 14(c) are graphs exemplifying results of verifying prediction accuracy by a prediction model.

MODE FOR CARRYING OUT THE INVENTION

[0018] 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 redundant descriptions are omitted. FIG. 1 schematically shows a production system for fresh concrete including a quality prediction device according to an embodiment.

[0019] [Production System for Fresh Concrete] First, we will explain the overview of the ready-mix concrete manufacturing system. Manufacturing System 1 (Ready-Mix Concrete Manufacturing System) shown in Figure 1 is a system for manufacturing ready-mix concrete. Manufacturing System 1 manufactures ready-mix concrete through at least a mixing process of concrete materials. The concrete materials used in Manufacturing System 1 include cement, admixtures, coarse aggregate, fine aggregate, water, and other admixtures.

[0020] Examples of coarse aggregate include natural coarse aggregate or artificial coarse aggregate. Other examples of coarse aggregate include gravel, crushed stone, slag coarse aggregate, lightweight coarse aggregate, recycled coarse aggregate, recovered coarse aggregate, or coarse aggregates that are mixtures of these. Gravel includes mountain gravel, land gravel, river gravel, or sea gravel. Slag coarse aggregate includes blast furnace slag aggregate, ferronickel slag aggregate, electric furnace oxidized slag aggregate, or coal gasification slag aggregate. Lightweight coarse aggregate includes natural lightweight aggregate, by-product lightweight aggregate, or artificial lightweight aggregate. Coarse aggregate may also include crushed rock or crushed limestone.

[0021] Examples of fine aggregate include natural aggregate and artificial aggregate. Other examples of fine aggregate include sand, crushed sand, slag fine aggregate, lightweight fine aggregate, recycled fine aggregate, recovered fine aggregate, or fine aggregates that are mixtures of these. Sand can be mountain sand, land sand, river sand, or sea sand. Slag fine aggregate can be blast furnace slag aggregate, ferronickel slag aggregate, copper slag aggregate, electric furnace oxidized slag aggregate, or coal gasification slag aggregate. Lightweight fine aggregate can be natural lightweight aggregate, by-product lightweight aggregate, or artificial lightweight aggregate.

[0022] Examples of rock types for crushed stone and crushed sand include igneous rocks, sedimentary rocks, metamorphic rocks, silica, limestone, domaloyte, or peridotite. Igneous rocks include granite, diorite, gabbro, porphyry, diorite, rhyolite, andesite, basalt, or serpentinite. Sedimentary rocks include conglomerate, sandstone, shale, slate, or tuff. Metamorphic rocks include gneiss or schist. A mixture of two or more of the above-mentioned examples may be used as coarse aggregate and fine aggregate.

[0023] At least a portion of the manufacturing system 1 is installed in a location different from the site where the ready-mixed concrete will be used (for example, a factory). The transport vehicle 200 transports the ready-mixed concrete to the site where it will be used (for example, a construction site) after it has been loaded. Examples of transport vehicles 200 include agitator trucks (mixer trucks) or dump trucks. The manufacturing system 1 may produce ready-mixed concrete from concrete materials to meet the target quality (required quality) set for each site. For example, an operator of the manufacturing system 1 determines the concrete mix to meet the target quality and inputs operation instructions to the manufacturing system 1.

[0024] In one example, ready-mixed concrete is manufactured in manufacturing system 1 to ensure that the target quality for use of the ready-mixed concrete, which is set for each site, is met, and the quality of the ready-mixed concrete is controlled (inspected, etc.) before shipment. Use of the ready-mixed concrete refers to the time when the ready-mixed concrete is received at the site. For quality control of the ready-mixed concrete before shipment, a target quality for shipment of the ready-mixed concrete may be determined based on the target quality for use of the ready-mixed concrete. The quality of ready-mixed concrete includes fresh properties. Specific examples of fresh properties include slump, slump flow, and air content.

[0025] Manufacturing system 1 includes, for example, a manufacturing device 100 and a quality prediction device 10. Manufacturing device 100 is a device that mixes concrete materials to produce ready-mixed concrete. Manufacturing device 100 may be installed in a factory or other location different from the site where the ready-mixed concrete is used. Manufacturing device 100 includes, for example, a material storage area 101, a transport device 104, a storage jar 111, a measuring jar 112, a collection hopper 113, a mixer 114, and a loading hopper 115.

[0026] 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 portion of the concrete materials, separated by material type. 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.

[0027] The transport device 104 is a device that transports concrete materials stored in multiple silos 102 to storage jars 111. The transport device 104 includes, for example, a belt conveyor for transporting concrete materials. The transport device 104 may transport concrete materials by type at different times. In one example, based on operation instructions from a control device provided in the manufacturing system 1, a specific material from among the various concrete materials is transferred to the transport device 104 and transported to the storage jars 111.

[0028] The storage jar 111 temporarily stores various concrete materials. Various concrete materials are transported to the storage jar 111 from the material storage area 101 by the transport device 104. The storage jar 111 is configured to store various concrete materials individually. Hereinafter, "concrete materials" may be simply referred to as "materials". The various materials stored in the storage jar 111 are supplied to the measuring jar 112 as needed.

[0029] The measuring bottle 112 is located below the storage bottle 111. The measuring bottle 112 operates based on operation instructions from a control device provided in the manufacturing system 1, and weighs various materials individually. When the measuring bottle 112 detects the target amount of material instructed by the control device, it supplies that material to the collection hopper 113. When water is supplied to the measuring bottle 112, an admixture may be mixed into the water. The collection hopper 113 is located below the measuring bottle 112. The collection hopper 113 collects the various materials discharged from the measuring bottle 112 and supplies the collected materials to the mixer 114. Note that the manufacturing apparatus 100 does not necessarily have a collection hopper 113, and the various materials may be supplied from the measuring bottle 112 to the mixer 114.

[0030] The mixer 114 is located below the collection hopper 113. The mixer 114 is a device for mixing concrete materials. The mixer 114 produces ready-mixed concrete by mixing aggregate, cement, water, and admixtures. The ready-mixed 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-axis mixer, a horizontal double-axis mixer, or a pan-type mixer. The mixer 114 includes, for example, two stirring members 114a and a mixer drive unit 114b.

[0031] The stirring members 114a are components that stir the various materials supplied to the mixer 114. Two stirring members 114a are arranged side by side inside the main body (container) of the mixer 114 and are rotatably mounted. Each of the two stirring members 114a includes a rotation axis extending in one horizontal direction. The mixer drive unit 114b rotates the rotation axes of each of the two stirring members 114a based on operation instructions from a control device provided in the manufacturing system 1. The mixer drive unit 114b includes a drive source, such as a motor, that provides driving force to the stirring members 114a. An opening is provided at the bottom of the main body of the mixer 114 for discharging the manufactured ready-mixed concrete into the loading hopper 115.

[0032] The loading hopper 115 is located below the mixer 114 and temporarily stores the ready-mix concrete. The loading hopper 115 then supplies the temporarily stored ready-mix concrete to the transport vehicle 200.

[0033] The manufacturing apparatus 100 described above is just one example of a ready-mix concrete manufacturing apparatus, and the ready-mix concrete manufacturing apparatus may be configured in any way as long as it can produce ready-mix concrete through a process of mixing concrete materials with a mixer. In this disclosure, the production of ready-mix concrete may include not only obtaining ready-mix concrete by mixing concrete materials, but also transporting the obtained ready-mix concrete and stirring the ready-mix concrete during transport. In one example, the manufacturing system 1 also includes a transport vehicle 200.

[0034] In this disclosure, a unit of ready-mixed concrete produced in one mixing cycle in the mixer 114 and loaded onto the transport vehicle 200 is defined as "1 batch". The process performed by the manufacturing system 1 to produce 1 batch of ready-mixed concrete is defined as "batch processing". The number of batch processing cycles performed from a certain reference point (cumulative number of executions) is called the "cumulative batch count". At the reference point, the cumulative batch count is reset to zero. For example, each day, when the manufacturing equipment 100 starts operation, the cumulative batch count is reset to zero. In this case, the cumulative batch count for the q-th batch processing (where q is an integer greater than or equal to 1) performed on a given day will be q.

[0035] One to three batches of ready-mix concrete may be loaded onto one transport vehicle 200. For example, if two batches of ready-mix concrete are loaded onto one transport vehicle 200, two batch processes, each following the same manufacturing conditions, will be carried out at different times (sequentially). Multiple batch processes, each following the same manufacturing conditions, may be carried out consecutively and sequentially during a certain period of time within a day.

[0036] <Quality Prediction Device> The quality prediction device 10 is a device that predicts the quality of ready-mixed concrete produced by the manufacturing device 100. The quality prediction device 10 is composed of one or more computers. If the quality prediction device 10 is composed of multiple computers, these computers are connected to each other in a way that allows them to communicate with one another. In addition to the function of predicting quality, the quality prediction device 10 may also have a function of controlling the manufacturing device 100. For example, the quality prediction device 10 may be configured as part of a control device that controls the manufacturing device 100. The quality prediction device 10 may control the manufacturing device 100 according to set operating conditions. At least a portion of the operating conditions may be determined by instructions from an operator such as a worker.

[0037] 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 instructions from workers, etc., into the quality prediction device 10. The input device 12 can be anything that can input the 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 workers, etc. The monitor 14 can be anything that can display graphics, and may be a liquid crystal display. The input device 12 and the monitor 14 may be integrated, such as a touch panel. The quality prediction device 10, the input device 12, and the monitor 14 may be integrated, such as a tablet computer (tablet terminal).

[0038] The quality of the ready-mixed concrete predicted by the quality prediction device 10 includes not only the quality of the ready-mixed concrete itself, but also the quality of the ready-mixed concrete after hardening (hardened concrete). Predicting the quality of ready-mixed concrete means calculating predicted values ​​for indicator values ​​that represent quality.

[0039] The quality prediction device 10 predicts, for example, the fresh properties of ready-mixed concrete as the quality of the ready-mixed concrete. The quality prediction device 10 may predict the fresh properties of the ready-mixed concrete as the quality of the ready-mixed concrete after it has been mixed in the manufacturing device 100 and before it is shipped from the factory where the manufacturing device 100 is installed to the site. Alternatively, the quality prediction device 10 may predict the fresh properties of the ready-mixed concrete after it has been shipped from the factory and before it is placed at the site. In other words, the quality prediction device 10 may predict the fresh properties of the ready-mixed concrete during transport, or it may predict the fresh properties of the ready-mixed concrete at the time of unloading.

[0040] The quality prediction device 10 may predict one or more indicator values ​​from among slump, slump flow, and air content as fresh concrete properties. The quality prediction device 10 may predict two or more indicator values ​​from among slump, slump flow, and air content as fresh concrete properties. The quality prediction device 10 may predict indicator values ​​other than slump, slump flow, and air content as fresh concrete properties. Examples of fresh concrete property indicators other than slump, slump flow, and air content include 500 mm flow arrival time, flow stop time, presence or absence of material segregation, temperature, bleeding amount, bleeding rate, setting time, unit water content, plastic viscosity, yield value, funnel flow time, compaction, deformability, filling ability, and void passage ability.

[0041] The quality prediction device 10 may predict at least one of the following as indicators of strength and durability as the quality of the fresh concrete after hardening. Indicators of strength include compressive strength, flexural strength, splitting tensile strength, bond strength, static elastic modulus, rebound degree, and toughness coefficient. Indicators of durability include length change rate, expansion rate, mass loss 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, "quality of fresh concrete" may be simply referred to as "quality."

[0042] To give an overview of the functions of the quality prediction device 10, the quality prediction device 10 first acquires feature quantities from time-series data of power load values ​​up to a certain point in time while the mixer 114 is operating. Then, the quality prediction device 10 predicts the quality of the ready-mixed concrete based on the input information including the feature quantities related to the power load values ​​and the time at which the feature quantities were acquired, and a prediction model constructed by machine learning. The quality prediction device 10 may also predict the quality of the ready-mixed concrete (the ready-mixed concrete after it has been discharged from the mixer 114 after its operation) at an intermediate stage while the mixer 114 is operating to obtain the ready-mixed concrete to be predicted. The quality prediction device 10 constructs a prediction model so that it can obtain a predicted value of the quality of the ready-mixed concrete while the mixer 114 is operating.

[0043] Figure 2 shows an example of the functional components (hereinafter referred to as "functional blocks") of the quality prediction device 10. The quality prediction device 10 includes, for example, an operation control unit 22, a first acquisition unit 24, a second acquisition unit 25, a model construction unit 26, a model holding unit 28, a prediction unit 30, and an output unit 32 as functional blocks. The processes performed by these functional blocks correspond to the processes performed by the quality prediction device 10.

[0044] The operation control unit 22 controls the manufacturing apparatus 100 to produce ready-mixed concrete according to predetermined operating conditions. At least a portion of the above operating conditions may be determined by an operator, such as a worker, each time ready-mixed concrete is produced. The operation control unit 22 may control the mixer drive unit 114b of the mixer 114 so that its rotational speed follows the target rotational speed specified in the operating conditions. When controlling the mixer drive unit 114b, the operation control unit 22 may adjust the power (e.g., current value) supplied to the mixer drive unit 114b. The power load tends to be higher when the ready-mixed concrete is hard, and lower when the ready-mixed concrete is soft.

[0045] The first acquisition unit 24 acquires one or more feature quantities from the time-series data of the power load value of the mixer 114 up to a point in the middle of the operation of the mixer 114 that mixes the concrete material. The middle of the operation of the mixer 114 means during the period in which mixing is continued to obtain the ready-mixed concrete for which quality is to be predicted (before that period ends). Hereinafter, the feature quantity obtained from the time-series data of the power load value will be referred to as "load feature quantity FL", and the point in time in the middle of obtaining the load feature quantity FL will be referred to as "measurement point mp".

[0046] The load feature FL is a feature related to the power load value of the mixer 114 when mixing concrete materials, for example. The power load value of the mixer 114 may be a value that represents the power (W) supplied to the mixer 114 itself, or a value that represents the current value (A) supplied to the mixer 114. Alternatively, the power load value of the mixer 114 may be a value that represents the load oil pressure (Pa) or a value that represents the torque (N·m) instead of power or current.

[0047] The time-series data of power load values ​​is a set of data consisting of continuous power load values ​​during the period in which the mixer 114 is operating during one batch of processing (a set of data showing the change in power load values ​​over time). Figure 3(a) schematically shows the time-series data (time-series data of power load values) for the entire period T in one batch of processing. The time-series data of power load values ​​can be obtained, for example, by repeatedly measuring the power (kW) supplied to the mixer 114 at a predetermined sampling period. The sampling period when measuring the power load value may be 0.01 to 1 second (for example, 0.1 seconds). In the following description, the time-series data of power load values ​​may be simply referred to as "time-series data".

[0048] In the graph representing the time-series data in Figure 3(a), "Time t1" indicates the time when the mixer 114 starts mixing, i.e., the start of the mixer 114's operation, and "Time t2" indicates the time when the mixer 114 finishes mixing. Time t1 corresponds, for example, to the time when the operation control unit 22 starts driving the agitator 114a. Time t2 corresponds, for example, to the time when the operation control unit 22 determines that the conditions for ending the mixing by the mixer 114 have been met. In one example, the operation control unit 22 determines that the above conditions have been met when a set time has elapsed since the start of driving the agitator 114a, and stops driving the agitator 114a. The above set time may be set for each batch of ready-mixed concrete produced, and in one example, it is set to about 15 to 120 seconds.

[0049] One or more loading features FL may include values ​​at a predetermined point in time in the time series data up to measurement point mp. The time series data up to measurement point mp refers to a portion of the time series data for the entire period T of one batch, from a point in time prior to measurement point mp (for example, the start of mixing: time t1) up to measurement point mp. Figure 3(b) shows an example of time series data up to measurement point mp. In the example shown in Figure 3(b), measurement point mp is the point in time after the elapsed time indicated by "tp" has elapsed from time t1, which represents the start of mixing. In other words, measurement point mp can be identified by the elapsed time tp from time t1 (a predetermined reference point).

[0050] The predetermined time point in the time series data up to measurement time point mp is a time point set in advance by the operator or the like. The predetermined time point can also be rephrased as a time point that satisfies the pre-set conditions. One or more load features FL may include one or more values ​​from among the initial value P1, minimum value P2, maximum value P3, and final value P4 as the value (power load value) at the predetermined time point.

[0051] The initial value P1 is the power load value at the start of the time series data up to measurement time mp (for example, at the start of mixing). The minimum value P2 is the power load value at the minimum point in the time series data up to measurement time mp. The maximum value P3 is the power load value at the maximum point in the time series data up to measurement time mp. The final value P4 is the power load value at the end of the time series data up to measurement time mp. Since the time series data of interest differs depending on the measurement time mp (elapsed time tp), the initial value P1, minimum value P2, maximum value P3, and final value P4 may also differ depending on the measurement time mp (elapsed time tp). Note that the initial value P1 may also be the minimum value P2 or the maximum value P3, and the final value P4 may also be the minimum value P2 or the maximum value P3.

[0052] One or more load features FL may include, in place of or in addition to, the power load value at the predetermined time point, statistics obtained from time-series data up to measurement time point mp. The statistics as load features FL are one or more values ​​selected from the group consisting of fluctuation range, decline range, sum, mean, standard deviation, coefficient of variation, median, first quartile, third quartile, kurtosis, and skewness. The fluctuation range is determined by the difference between the maximum value P3 and the minimum value P2. The decline range is determined by the difference between the maximum value P3 and the final value P4.

[0053] Other statistical quantities besides the range of fluctuation and the range of decrease may be calculated from the data set that constitutes the time series data up to the measurement time mp (the 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 the time series data up to the measurement time mp. In addition, statistical quantities such as the total value may be calculated after some of the power load values ​​for each sampling period have been decimated.

[0054] One or more load features FL may include the difference in power load values. The difference in power load values ​​is obtained by subtracting the power load value at a setting point prior to the measurement time mp from the power load value at the measurement time mp. The setting point can be determined in any way, but it may be the time corresponding to the previous sampling period, or, as described later, the measurement time mp may be changed stepwise to calculate the load features FL, etc., and may be the measurement time mp prior to the measurement time mp. One or more load features FL may include the slope of the power load values. The slope of the power load values ​​is obtained by dividing the difference in power load values ​​by the difference between the measurement time mp and the setting point.

[0055] One or more load features FL may include features obtained from the waveform of time-series data up to measurement time mp, or features obtained from an image displaying time-series data up to measurement time mp, or from image data related to said image. The time-series data related to the power load value may represent a moving average of the power load value instead of data representing the time change of the power load value itself.

[0056] The second acquisition unit 25 acquires input information including one or more load features FL acquired by the first acquisition unit 24 and the time from a predetermined reference time to the measurement time mp. The predetermined reference time is, for example, the time when mixing begins, and in this case, the time from the reference time to the measurement time mp is the elapsed time tp described above. The input information acquired by the second acquisition unit 25 may include one or more load features FL and the elapsed time tp. The input information may also include information other than one or more load features FL and the elapsed time tp, but specific examples of other information will be described later.

[0057] If the prediction model used by the quality prediction device 10 for predicting quality is constructed using a decision tree algorithm described later, the second acquisition unit 25 may acquire multiple features as input information, including one or more load features FL and elapsed time tp. Here, each of the multiple features acquired by the second acquisition unit 25 may be denoted as feature F1, feature F2, ..., and feature FN (where N is an integer of 2 or more). At least some of the feature quantities F1 to FN (multiple feature quantities) may be features that represent the result of the mixer 114's operation, or features that represent the state of the mixer 114 while it is operating. The feature quantities F1 to FN may include features that represent the manufacturing conditions using the mixer 114 (for example, the operating conditions described above).

[0058] Each of the feature vectors F1 to FN is data represented by a numerical value. Feature vectors F1 to FN may include data representing classification (type), in which case a different numerical value (for example, 0 or 1) is assigned to each classification. For example, a value of 0 could be used if the data does not belong to that classification, and a value of 1 if it does belong to that classification. Feature vectors F1 to FN do not include image data composed of a combination of coordinate information and pixel values. However, feature vectors F1 to FN may include such image data, or feature vectors (one-dimensional numerical data) obtained from images related to such image data.

[0059] The model building unit 26 constructs a predictive model by performing machine learning using training data composed of multiple training datasets. Hereinafter, the training data composed of multiple training datasets will be referred to as "training data TD," and the predictive model constructed by machine learning will be referred to as "predictive model M." Predictive model M may be a model for predicting the fresh properties of ready-mixed concrete obtained when the second acquisition unit 25 obtains input information (for example, features F1 to FN). The fresh properties predicted by predictive model M may be slump, slump flow, or air content.

[0060] Machine learning is a method in which a machine (computer) autonomously discovers laws or rules by iteratively learning based on given information. A predictive model M can be constructed using algorithms and data structures. The algorithm used to construct the predictive model M is not particularly limited. The stage in which the predictive model M is autonomously constructed corresponds to the training phase (learning phase). The training phase may be performed before the production phase in which ready-mixed concrete is manufactured, or it may be performed at the beginning of the production phase.

[0061] The evaluation phase corresponds to the stage in which a predictive model M is used to predict the quality of freshness, etc., from input information (e.g., features F1-FN) whose quality is unknown. In the following explanation, terms may be prefixed with "training data" or "evaluation data" to distinguish whether it is the training phase or the evaluation phase, or which phase the data is used in. A specific example of the predictive model M will be described later.

[0062] Here, with reference to Figures 4 and 5, an example of preparing training data TD will be described. In order to predict the quality of the fresh concrete discharged from mixer 114 while mixer 114 is operating and after mixer 114 has finished operating, it is conceivable to prepare training datasets by gradually varying the measurement time mp (elapsed time tp). In this case, M training time series data (M batches) are prepared, and then multiple training datasets are prepared for each single training time series data by gradually varying the measurement time mp. M is an integer of 2 or greater.

[0063] To facilitate understanding of the explanation, we assume that the total period T (see Figure 3(a)) of one batch of time-series data is 30 seconds. The total period T is the time from time t1 to time t2. In the training phase, unlike the evaluation phase in which predictions are made during the operation of mixer 114, time-series data for the entire period T is obtained. Therefore, in the training phase, the time-series data up to measurement time mp is extracted from the time-series data for the entire period T, and the training dataset is prepared.

[0064] The model building unit 26 performs the following acquisition process for each of the M training time series data (for each time series data). In the acquisition process, as shown in Figure 4, the model building unit 26 changes the measurement time mp in 1-second increments, starting from 3 seconds after time t1, and for each changed measurement time mp, it acquires one or more load features FL from the time series data up to that measurement time mp. In the example shown in Figure 4, the model building unit 26 calculates one or more load features FL from the time series data from the start of mixing by the mixer 114 (time t1) to the measurement time mp for each changed measurement time mp.

[0065] Specifically, the model building unit 26 calculates one or more load features FL from the time series data from time t1 (0 seconds) to 3 seconds, and saves the calculation result in association with the elapsed time tp (=3 seconds). Then, the model building unit 26 calculates one or more load features FL from the time series data from time t1 (0 seconds) to 4 seconds, and saves the calculation result in association with the elapsed time tp (=4 seconds). Next, the model building unit 26 calculates one or more load features FL from the time series data from time t1 (0 seconds) to 5 seconds, and saves the calculation result in association with the elapsed time tp (=5 seconds). Thereafter, the model building unit 26 repeats the same process in 1-second increments until the elapsed time tp reaches 30 seconds. The starting point for calculating one or more load features FL, and the range of variation for the measurement time mp (elapsed time tp), can be set arbitrarily. As shown in the example in Figure 4, the measurement time mp may include not only cases where it is before time t2, but also cases where it is time t2.

[0066] The model building unit 26 associates the correct quality data (e.g., slump) with each of the M training time-series data sets, for each combination of one or more load features FL obtained at each measurement point and the elapsed time tp. An example of the training data TD constructed in this manner is shown in Table 1 below. Note that while specific numerical values ​​are shown in Table 1 for clarity, the values ​​in Table 1 are just examples.

[0067] [Table 1]

[0068] In Table 1, "Batch No." indicates which batch of time-series data (out of M batches) the data belongs to. "FL(i,j)" represents the j-th loading feature FL obtained from the time-series data of the i-th batch. i is an integer from 1 to M, and j is an integer from 1 to J (where J is an integer greater than or equal to 2). In the example shown in Figure 4, J is 28.

[0069] In Table 1 above, focusing on a training dataset for any single batch (hereinafter referred to as "a training dataset for one batch"), this training dataset for one batch is data obtained from a single continuous operation of the mixer 114. A training dataset for one batch includes a first dataset and a second dataset that have the following relationship: The first dataset and the second dataset share the same ground truth value for quality (slump), and their measurement time mp are different. If a training dataset for one batch consists of three or more datasets, the first dataset and the second dataset represent a set of arbitrarily selected datasets. Furthermore, in all combinations of a set of datasets, the relationship holds that the ground truth value for quality (slump) is shared, and their measurement time mp are different. In at least some of the combinations of datasets, one or more loading features FL are different. As described above, a training dataset for one batch includes two or more datasets that share the same ground truth value for quality and have different measurement time mp.

[0070] Unlike the example shown in Figure 4, as shown in Figure 5, the model building unit 26 may calculate one or more load features FL in block units having a fixed interval during the acquisition process. In the acquisition process, the model building unit 26 changes the measurement time mp in 1-second increments, starting from a point 5 seconds after time t1, and for each changed measurement time mp, it acquires one or more load features FL from the time series data up to that measurement time mp. In the example shown in Figure 5, the model building unit 26 acquires one or more load features FL for each changed measurement time mp, based on the time series data from a point a predetermined time (for example, 2 to 10 seconds) before that measurement time mp to that measurement time mp. In one example, the model building unit 26 extracts time series data in 5-second increments for each changed measurement time mp and calculates one or more load features FL.

[0071] Specifically, the model building unit 26 calculates one or more load features FL from the time series data from time t1 (0 seconds) to 5 seconds, and saves the calculation result in association with the elapsed time tp (= 5 seconds). Then, the model building unit 26 calculates one or more load features FL from the time series data from 1 second to 6 seconds, and saves the calculation result in association with the elapsed time tp (= 6 seconds). Subsequently, the model building unit 26 calculates one or more load features FL from the time series data from 2 seconds to 7 seconds, and saves the calculation result in association with the elapsed time tp (= 7 seconds). From there, the model building unit 26 repeats the same process in 1-second increments until the elapsed time tp reaches 30 seconds. The starting point for calculating one or more load features FL, the range of variation for the measurement time mp (elapsed time tp), and the range for extracting time series data can be set arbitrarily.

[0072] The model building unit 26, similar to the example shown in Figure 4, associates the correct quality data (e.g., slump) with each of the M training time series data sets, for each combination of one or more load features FL obtained at each measurement time point mp and the elapsed time tp. This prepares the training data TD, which consists of multiple training datasets including the first dataset and the second dataset. As described above, for explanatory purposes, it was assumed that the total period T of one batch of time series data (see Figure 3(a)) is 30 seconds, but the total period T of one batch of time series data may be longer than 30 seconds, and a portion of that total period T (30 seconds) may be used to prepare the training data TD.

[0073] Returning to Figure 2, the model holding unit 28 holds (stores) the prediction model M constructed by the model building unit 26. The prediction model M, being a trained model, may be transferable between computers. Therefore, the prediction model M constructed in the quality prediction device 10 may be used in other manufacturing systems different from the manufacturing system 1. The model building unit 26 may construct a prediction model M for each type of freshness (for example, for slump, slump flow, and air content), in which case the model holding unit 28 holds a prediction model M for each type of freshness.

[0074] The prediction unit 30 uses the prediction model M constructed by the model construction unit 26 to obtain predicted quality values ​​corresponding to the evaluation input information obtained by the second acquisition unit 25. For example, the prediction unit 30 inputs the values ​​of evaluation features F1 to FN into the prediction model M and obtains the predicted freshness characteristics output from the prediction model M. The prediction unit 30 may also obtain predicted values ​​for each type of freshness characteristic using the corresponding prediction model M.

[0075] The prediction unit 30 may calculate a predicted quality value while the mixer 114 is still operating on the ready-mixed concrete to be evaluated (hereinafter referred to as "ready-mixed concrete to be evaluated"). In this case, the prediction unit 30 will predict the quality of the ready-mixed concrete at the time of completion of mixing (at the time of shipment) before the mixing of the concrete materials by the mixer 114 is completed.

[0076] The output unit 32 outputs the prediction results from the prediction unit 30. For example, the output unit 32 outputs the predicted quality values, such as fresh concrete properties, calculated by the prediction unit 30, to the monitor 14. As a result, the predicted quality values ​​of the ready-mixed concrete being evaluated are displayed on the monitor 14, allowing operators such as workers to understand the predicted quality values ​​of the ready-mixed concrete.

[0077] The output unit 32 may display the predicted quality value on the monitor 14 while the mixer 114 is still operating with the ready-mixed concrete being evaluated. In this case, operators such as workers can understand the quality of the ready-mixed concrete at the time of mixing completion (at the time of shipment) before the mixing of the concrete materials by the mixer 114 is finished.

[0078] <Specific example of a predictive model M: Decision tree> Next, with reference to Figures 6 and 7, an example of a prediction model M for predicting the fresh properties of ready-mixed concrete will be explained. Prediction model M is implemented, for example, using a decision tree, which is a method of analyzing data using a tree structure (tree diagram). The prediction model M constructed by the decision tree algorithm outputs predicted values ​​by sequentially setting conditions on the given input data and branching out to the expected results.

[0079] When a prediction model M constructed by a decision tree algorithm is used to predict freshness characteristics, the input information to the prediction model M consists of features F1 to FN, which include one or more loading features FL and a time (e.g., elapsed time tp) that identifies the measurement point. Elapsed time tp is one of the features F1 to FN. In one example, the model construction unit 26 constructs a prediction model M using machine learning that shows the relationship between the features F1 to FN and freshness characteristics (predicted values). The prediction model M is constructed to output predicted values ​​of freshness characteristics in response to the input of features F1 to FN. The features F1 to FN can be called explanatory variables, and freshness characteristics can be called the target variable.

[0080] The model building unit 26 constructs a prediction model M such that one or more features selected from the feature vectors F1 to FN are associated with the predicted values ​​of fresh characteristics using a decision tree (decision tree algorithm). The one or more features used for associating with fresh characteristics in the prediction model M are selected from the feature vectors F1 to FN by machine learning. LightGBM (Light Gradient Boosting Machine) may be used as the decision tree algorithm. Alternatively, XGBoost (eXtreme Gradient Boosting) or RandomForest may be used instead of LightGBM as the decision tree algorithm.

[0081] The model building unit 26 may autonomously construct a predictive model M by performing machine learning using the data provided as input to machine learning and the ground truth data (ground truth values ​​for fresh concrete properties) from the output of machine learning. The input to machine learning is features F1 to FN. The output of machine learning is data (numerical values) indicating the fresh concrete properties. The model building unit 26 iteratively learns a model that outputs predicted values ​​for fresh concrete properties using multiple combinations of features F1 to FN and ground truth values ​​for fresh concrete properties (the training data TD mentioned above). In each of the multiple datasets that make up the training data TD, features F1 to FN are associated with ground truth values ​​for fresh concrete properties. The same types of features F1 to FN are used between the training phase and the evaluation phase.

[0082] Figure 6 shows a tree diagram that visualizes an example of the analysis process in the predictive model M. For the sake of easier understanding of this disclosure, a simplified example will be used for explanation. In the decision tree algorithm exemplified below, one data set is divided into two data sets in stages according to certain conditions. These stages are referred to as "Stage 1," "Stage 2," and "Stage 3," respectively, from highest to lowest. The points where division or branching occurs based on conditions are called "nodes (condition nodes)." In Figure 6, combinations of "G" and numerical values, such as "G1," represent the names of the data sets. Combinations of "Th" and numerical values, such as "Th1," represent thresholds.

[0083] The model building unit 26 prepares training data TD, which consists of multiple training datasets in which the combinations of feature values ​​F1 to FN are different from each other. Data group G1 is, for example, all the training datasets included in training data TD. In the first stage, data group G1 is divided into two data groups, data group G21 and data group G22, based on one feature (feature F2 in the example shown in Figure 6) selected by machine learning from among the features F1 to FN. In the first stage, data group G1 is divided into two data groups based on the condition that feature F2 is less than or equal to the threshold Th1.

[0084] Dataset G21 consists of multiple training datasets from dataset G1 where the feature F2 is less than the threshold Th1, and dataset G22 consists of the remaining training datasets from dataset G1 where the feature F2 is greater than or equal to the threshold Th1. The number of training datasets in dataset G21 and the number of training datasets in dataset G22 may be the same or different. In the first stage, the data is divided into two sets using the feature F2, but the values ​​of features other than feature F2 are also retained in datasets G21 and G22. The threshold Th1 is also set autonomously by machine learning.

[0085] In the second stage, data group G21 and data group G22 are each divided into two data groups using features and thresholds selected by machine learning, similar to the first stage. The features selected in the second stage may be different from or the same as those selected in the first stage. The features selected to divide each data group into two may be different from or the same between data group G21 and data group G22. In the example shown in Figure 6, feature F1 is selected to divide data group G21 into two, and feature F2 is selected to divide data group G22 into two. In the second stage, data group G21 is divided into data group G31 and data group G32, and data group G22 is divided into data group G33 and data group G34.

[0086] In the example shown in Figure 6, in the third stage, data groups G31, G32, and G34 are not divided into two data groups. That is, each of data groups G31, G32, and G34 corresponds to a "leaf" (or "end node") in the tree diagram. In the third stage, with respect to data group G33, which is not a leaf, it is divided into two data groups by a feature and threshold selected by machine learning, as in the previous stage. Feature F10 is selected to divide data group G33 into two data groups, and data group G33 is divided into data group G41 and data group G42. Each of data groups G41 and G42 corresponds to a leaf. As illustrated in Figure 6, the leaf depth (level) may differ among at least some of the leaves in the tree diagram. Alternatively, unlike the example shown in Figure 6, the depth of each leaf may be the same.

[0087] For each data set corresponding to a leaf, the predicted value of the fresh characteristics can be determined by the ground truth values ​​of the fresh characteristics included in that data set. For example, the arithmetic mean of the ground truth values ​​of the fresh characteristics included in the data set corresponding to a leaf is set as the predicted value for that leaf. In one example, if data set G41 contains p data sets (where p is an integer greater than or equal to 2), the arithmetic mean of the p ground truth values ​​of the fresh characteristics is set as the predicted value for the leaf corresponding to data set G41. The predicted values ​​of the fresh characteristics set for each leaf will differ among multiple leaves.

[0088] The model building unit 26 may perform the division (branching) up to the data group corresponding to the leaves, while selecting features to be associated with the predicted values ​​of fresh characteristics and setting thresholds used when branching, according to various construction conditions. As described above, in the training phase, the training data TD is divided into two data groups according to the conditions of each stage, and a predicted value of fresh characteristics is set for each leaf (end node).

[0089] In the evaluation phase, where a prediction model M is used to obtain a predicted value for freshness from evaluation features F1 to FN, the prediction unit 30 evaluates which leaf a given evaluation feature F1 to FN (a combination of the values ​​of feature F1 to FN) belongs to, according to the branching conditions for each stage. Since each leaf has a predicted value for freshness, the predicted value for freshness can be obtained according to the combination of the values ​​of feature F1 to FN.

[0090] Assuming that a simplified prediction model M, as illustrated in Figure 6, is constructed, the prediction model M first compares the value of evaluation feature F2 (one of the evaluation features F1-FN) with the threshold Th1. If the value of evaluation feature F2 is greater than or equal to the threshold Th1, the calculation step proceeds to the node corresponding to data group G22, where the value of evaluation feature F2 is compared with the threshold Th22. If the value of evaluation feature F2 is less than the threshold Th22, the calculation step proceeds to the node corresponding to data group G33, where the value of evaluation feature F10 is compared with the threshold Th31.

[0091] Furthermore, if the value of the evaluation feature F10 is smaller than the threshold Th31, the prediction model M outputs a predicted value for the fresh properties set for the leaf (end node) corresponding to the data group G41. By performing the above calculations in the prediction model M, the prediction unit 30 can obtain a predicted value for the fresh properties of the ready-mixed concrete to be evaluated from the combination of the values ​​of the evaluation feature F1 to FN.

[0092] In the prediction model M, only a selection of features from features F1 to FN may be used in the branching conditions. That is, in the prediction model M, only a selection of features from features F1 to FN may be associated with the predicted values ​​of fresh characteristics. Even in such cases, the features associated with the predicted values ​​of fresh characteristics are selected from features F1 to FN, so the prediction model M still represents the relationship between features F1 to FN and the predicted values ​​of fresh characteristics.

[0093] As illustrated using Figure 4 or Figure 5, if training data TD is prepared by gradually varying the measurement time mp (elapsed time tp) for each of the M batches of time series data, then the elapsed time tp may be selected as a feature used for branching in the prediction model M.

[0094] The model building unit 26 may construct the prediction model M using machine learning, including ensemble learning, when constructing the prediction model M using the decision tree algorithm. In machine learning including ensemble learning, multiple models may be merged to construct a single prediction model M.

[0095] The model building unit 26 may construct a prediction model M for obtaining predicted values ​​using bagging, a method of ensemble learning. In one example, a first model M1 (weak learner) is generated from a portion of the data randomly extracted from the training data TD, and a second model M2 (weak learner) is generated from another portion of the data randomly extracted from the training data TD. The portion of the data used to generate the first model M1 and the portion of the data used to generate the second model M2 may overlap. In the prediction model M constructed by bagging, for example, the arithmetic mean of the predicted value from the first model M1 and the predicted value from the second model M2 is output as the final predicted value.

[0096] A prediction model M that ultimately outputs a single predicted value may be constructed from three or more models (weak learners). Alternatively, a prediction model M that ultimately outputs a single predicted value may be constructed by using a Random Forest, which is generated by a random selection of features from a portion of the data, such as the first model M1 (weak learner). Note that Random Forest can also be considered a type of bagging technique.

[0097] The model building unit 26 may construct a predictive model M using machine learning that utilizes boosting, as a form of machine learning that includes ensemble learning. In machine learning that utilizes boosting, intermediate models Mtk (weak learners) are sequentially generated using a decision tree algorithm. Here, k is an integer from 1 to K, and K is an integer greater than or equal to 2. If intermediate models Mtk are generated in the order of k from 1 to K, then in machine learning that utilizes boosting, intermediate model Mt1 is generated first. Then, focusing on the error between the predicted value and the correct value in intermediate model Mt1, the next intermediate model, intermediate model Mt2, is generated. The single predicted value that is ultimately output by the predictive model M is a value that is aggregated by some calculation from the outputs of each intermediate model Mt1 to MtK.

[0098] The model building unit 26 may construct a predictive model M consisting of intermediate models Mt1 to MtK using "Adaboost," a boosting technique. In the predictive model M constructed by Adaboost, the weights of the training data TD used to generate the next intermediate model Mt2 are updated based on the error between the predicted value (1) of the previous intermediate model Mt1 and the actual value. For example, the weights are updated to be larger for datasets with larger errors. Subsequently, the weight updates and the generation of intermediate model Mtk are repeated for a set number of times, for example, K times.

[0099] The model building unit 26 may construct a predictive model M using machine learning that utilizes boosting, specifically by using a Gradient Boosting Decision Tree. In machine learning using a Gradient Boosting Tree, the predictive model M is constructed by combining a decision tree as a weak learner with boosting and gradient descent (gradient boosting).

[0100] Figure 7 schematically illustrates the process of building a model using machine learning with gradient boosting trees. In gradient boosting trees, an intermediate model Mt1 is first generated, and the error between the predicted value (1) from intermediate model Mt1 and the actual value is calculated. Then, an intermediate model Mt2 is generated that can predict the error in the predicted value (1), and a correction value (2) is calculated from intermediate model Mt2 to correct the predicted value (1). Subsequently, for example, the calculation of the correction value (k) to correct the predicted value (k) at that point is repeated up to a set number of times, such as K times. The correction value (k) is also called the gradient (k).

[0101] Examples of decision tree algorithms that include gradient boosting (i.e., gradient boosting trees) include LightGBM and XGBoost mentioned above, as well as Catboost. In XGBoost, node branching (splitting) is performed level-wise; that is, all nodes are branched from left to right. On the other hand, in LightGBM, node branching (splitting) is performed leaf-wise; that is, branching is performed only on nodes that should be branched (for example, prioritizing nodes that result in a smaller loss). No further calculations are performed on nodes that no longer need to be branched.

[0102] Furthermore, LightGBM uses a histogram-based algorithm when performing branching. For the sake of explanation, let's assume there are six data points (or datasets) to be branched, and we will label these data points "d1" to "d6". In a pre-sorted algorithm, which is different from a histogram-based algorithm, each value in data points d1 to d6 is examined to determine where to branch (split). On the other hand, in a histogram-based algorithm, the values ​​in data points d1 to d6 are grouped by a histogram, and the branching (splitting) is determined on a group (cluster) basis.

[0103] Furthermore, LightGBM uses a technique called GOSS (Gradient-based One-Side Sampling) to reduce the amount of data used during the training process. Specifically, for data with large errors, all of it is used as untrained data, while for data with small errors, only a portion is used as it is considered to have been trained to some extent. In addition, LightGBM uses a technique called FEB (Exclusive Feature Bundling) to combine multiple features of different types that do not simultaneously become zero (low correlation) and can be combined without problems into a single feature.

[0104] <Specific example of a predictive model M: Neural network> The prediction model M may be implemented using a neural network, which is an information processing model that mimics the structure of the human brain. A neural network, for example, has an input layer, one or more hidden layers, and an output layer. By including one or more hidden layers, a more complex prediction model M can be constructed, improving prediction accuracy. A deep neural network (DNN) with multiple (many) hidden layers may be used as the neural network.

[0105] Figure 8 schematically shows a prediction model M constructed by the model construction unit 26 performing machine learning using a neural network. The prediction model M shown in Figure 8 is a model that outputs a quality prediction value in response to input information including one or more load features FL and a time (e.g., elapsed time tp) that identifies the measurement time mp. The input information to the prediction model M may be the features F1 to FN described above, or it may be information that combines the features F1 to FN with at least one of waveform data or image data. The prediction model M may be constructed to output a single quality prediction value, or it may be constructed to output two or more types of quality prediction values. When the prediction model M is constructed to output a single quality prediction value, the model construction unit 26 may construct a prediction model M for each type of quality.

[0106] Here, an example of a predictive model M constructed by a neural network will be explained using simplified mathematical formulas for ease of understanding. The predictive model M constructed by the model construction unit 26 can be simply expressed as shown in equations (1) and (2) below.

number

number

[0107] In equation (2), Y represents the output value of quality. When a predictive model M is constructed that outputs two or more types of quality, Y represents the output value of one of the two or more types of quality, and equations (1) and (2) are calculated for each type of quality. I is an integer greater than or equal to 2 and represents the number of input data. x represents the various input values ​​included in the input data, and the input data corresponds to the above input information, and the input data includes at least one loading feature FL and the time that identifies the measurement time mp (e.g., elapsed time tp). wi is the weight (coefficient), and b is the bias term (coefficient). f(U) represents the activation function. The activation function is a linear function (identity function) or a nonlinear function such as a polynomial.

[0108] The model building unit 26 may repeatedly evaluate the error (loss) between Y (predicted value) obtained by equation (2) and the correct quality value using the training data TD, and determine the weight wi and bias term b in equation (1) so as to minimize the error. The model building unit 26 may use any type of loss function as the function for evaluating the error between Y (output value from the intermediate model during the construction of the prediction model M) obtained by equation (2) and the correct quality value. Examples of loss functions include the Huber loss function, the mean absolute error (MAE), and the ε-insensitive loss function.

[0109] The model building unit 26 may repeatedly update the weights wi using the gradient method so that the error evaluated by the loss function is minimized. The model building unit 26 may use any type of update formula (weight update formula) when updating the weights wi. Examples of weight update formulas include Adam, AdamBelief, Adamax, AdaBound, Adagrad, AMSGRAD, AMSBound, RMSprop, SgdW, Momentum, and Nesterov. The weight update formula is also called an optimization algorithm or optimization method. An operation to normalize the input information of the prediction model M may or may not be performed. An operation to lose some of the connections in the fully connected layer of the prediction model M may or may not be performed.

[0110] <Specific examples of other information in the input information> As described above, the input information or features F1-FN may include information other than one or more loading features FL and the time (e.g., elapsed time tp) that identifies the measurement time mp. If the other information exemplified in the following description is one-dimensional data, that information can be one of the features F1-FN, as well as part of the input information. When training data TD is prepared as shown in Table 1, the other information may be a single common value across multiple training datasets in one batch.

[0111] The input information to the prediction model M may include information relating to the timing of when the mixer 114 performs mixing of concrete materials (hereinafter referred to as "production time information"). The production time information may include, for example, at least one of the following: the date and time when the mixer 114 performs mixing of concrete materials, the cumulative number of batches mixed by the mixer 114 in one day, the time from an arbitrarily set reference point to the time when the mixer 114 performs mixing, and the cumulative number of batches mixed by the mixer 114 from an arbitrarily set reference point.

[0112] The date on which mixer 114 performs mixing may be specified by representing the specific day on which mixer 114 produced the ready-mixed concrete as a combination of month and day. In this case, the specific month may be one feature and the specific day may be another feature. The time on which mixer 114 performs mixing may be specified by representing the specific point in time on which mixer 114 produced the ready-mixed concrete as a combination of hour, minute, and second, or as a combination of hour and minute. In this case, the specific time may be one feature and the specific minute (or each of the specific minute and second) may be another feature.

[0113] The above time, as information relating to the manufacturing period (the specific point in time when the mixer 114 manufactured ready-mixed concrete), may be the time when concrete material was supplied to the mixer 114, or the time when ready-mixed concrete was discharged from the mixer 114. The above time, as information relating to the manufacturing period (the specific point in time when the mixer 114 manufactured ready-mixed concrete), may be the time when the mixing member 114a was started to operate in the mixer 114, or the time when the mixing member 114a was stopped to operate. Unlike the elapsed time tp, the time in the information relating to the manufacturing period may be a constant value across multiple training datasets for one batch.

[0114] The cumulative number of batches mixed by mixer 114 in a day is information that identifies which batch process was executed on that day (which batch process produced the ready-mixed concrete). The batch process in which the ready-mixed concrete of interest was produced also indicates when the mixing by mixer 114 was performed. In other words, information related to the production period, unlike elapsed time tp, indicates when mixer 114 operated over a longer period.

[0115] If the time from an arbitrarily set reference point to the time when the mixer 114 performs mixing is used as part of the input information (one feature) as information related to the manufacturing period, the reference point may be set by an operator such as a worker. The reference point may be set to the time after the inside of the mixer 114 has been cleaned during maintenance of the manufacturing equipment 100, and before manufacturing by the mixer 114 is resumed. In the case where the mixer 114 is cleaned every day after being shut down, the time before manufacturing by the mixer 114 is started for the first time in a day is also included in the time before manufacturing by the mixer 114 is started for the first time in a day. The time between the reference point and the time when the mixer 114 performs mixing (production of ready-mixed concrete) may be expressed in minutes or in seconds.

[0116] The time when mixer 114 performs mixing (production of ready-mixed concrete) may be the time when concrete materials are supplied to mixer 114, or the time when the ready-mixed concrete is discharged from mixer 114. The time when mixer 114 performs mixing (production of ready-mixed concrete) may be the time when the agitator 114a is started to drive, or the time when the agitator 114a is stopped from driving. The quality prediction device 10 may measure the time from the reference time to the time when mixer 114 performs mixing.

[0117] The cumulative batch count from an arbitrarily set reference point is information that identifies which batch process is being executed from the reference point (which batch process from the reference point produced the ready-mixed concrete). In the cumulative batch count from the reference point, the cumulative batch count is reset to 0 at the reference point. The reference point for counting the cumulative batch count may also be set to the point after cleaning of the mixer 114 during maintenance of the manufacturing equipment 100, and before manufacturing by the mixer 114 is resumed (including the point before manufacturing by the mixer 114 is started for the first time in a day).

[0118] The input information may include information related to mixing by the mixer 114 (information other than the power load value). In addition to the elapsed time tp mentioned above, information related to mixing may include, for example, the amount of concrete mixed for one batch, the time at which the power load value reached its maximum value in the time series data up to the measurement time mp, and the time from the time at which the power load value reached its maximum value to the measurement time mp (final value P4).

[0119] The input information may include information indicating the specified strength, type of cement, type of admixture, type of fine aggregate, type of coarse aggregate, and amount of additives added. In the information indicating the type of cement, etc., the types of cement, etc. may be classified in any way before each type is specified. The input information may also include image data relating to images of ready-mixed concrete being produced in mixer 114.

[0120] The input information includes the mix strength, target slump, target slump flow, target air content, water-binder ratio, fine aggregate ratio, and the unit amounts [kg / m³] of various materials (water, cement, admixtures, fine aggregate, coarse aggregate, and other admixtures). 3 ] or unit volume [L / m³ 3 The input information may include one or more of the following: brand name, manufacturing date, and density [g / cm³] of each material. 3 ], and information on one or more of the following for aggregates (fine aggregates and coarse aggregates): water absorption rate, moisture content, surface moisture content, actual volume, coarse particle size ratio, particle size distribution, and maximum size.

[0121] The input information (feature quantities F1-DN) may include numerical data obtained from images of ready-mixed concrete being manufactured in mixer 114. The input information (feature quantities F1-DN) may include at least one of the target quality of the ready-mixed concrete when used and the target quality of the ready-mixed concrete when shipped. One of the other types of information exemplified above may be included in the input information, and two or more types of information may be included in the input information.

[0122] <Hardware configuration of the quality prediction system> As shown in Figure 9, the quality prediction device 10 includes a circuit 50. The circuit 50 includes 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 flash memory or a hard disk. The storage 53 stores at least a quality prediction program that causes the computer to execute the construction process, the first acquisition process, the second acquisition process, and the prediction process, which will be described later. The storage 53 stores quality prediction programs for configuring each functional block of the quality prediction device 10.

[0123] Memory 52 is composed of one or more volatile memory devices, such as random access memory. Memory 52 temporarily stores the quality prediction program loaded from storage 53. Processor 51 is composed of one or more computing devices, such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). Processor 51 executes the quality prediction program loaded into memory 52 to configure each functional block of the quality prediction device 10. The calculation results by processor 51 are temporarily stored in memory 52. ​​Input / output ports 54 perform input and output of information to and from input devices 12, monitors 14, mixers 114, etc., in response to requests from processor 51.

[0124] Timer 55 measures elapsed time, for example, by counting reference pulses of a fixed period. Note that circuit 50 is not necessarily limited to having each function configured by a program. For example, circuit 50 may have at least some functions configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates such a circuit. The quality prediction program may be provided by being permanently recorded on a tangible recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory. Alternatively, the quality prediction program may be provided via a communication network as a data signal superimposed on a carrier wave.

[0125] [Method of manufacturing ready-mixed concrete] Next, an example of a ready-mix concrete manufacturing method performed in manufacturing system 1 will be described. The ready-mix concrete manufacturing method includes a manufacturing process and a quality prediction process. The manufacturing process is the process of manufacturing ready-mix concrete. The quality prediction process is the process of predicting the quality (e.g., fresh properties) of the ready-mix concrete produced by the manufacturing process. The quality prediction process may be performed for a period that overlaps with at least a portion of the period during which the manufacturing process is repeatedly executed.

[0126] The manufacturing process includes, for example, a conveying process, a weighing process, a loading process, a mixing process, a discharge 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 each material is supplied to the storage bottle 111 individually. In the weighing process, each material is supplied individually from the storage bottle 111 to the weighing bottle 112, and each material is weighed in the weighing bottle 112. In the weighing process, when the measured amount of each material reaches a predetermined set amount, that material is discharged into the collection hopper 113. In the loading process, after all types of materials have been collected in the collection hopper 113, the materials in the collection hopper 113 are loaded (supplied) into the mixer 114.

[0127] In the mixing process, multiple types of concrete materials are mixed in the mixer 114. In the mixing process, a quality prediction device 10, which has the function of controlling the manufacturing apparatus 100 according to predetermined operating conditions, may control the mixer drive unit 114b. 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 rotational speed of the mixer drive unit 114b follows a target rotational speed.

[0128] In the discharge process, after the mixing of the concrete materials is completed in the mixer 114, the ready-mixed concrete is discharged from the mixer 114 to the loading hopper 115. In the loading process, the ready-mixed concrete discharged into the loading hopper 115 is loaded onto the transport vehicle 200.

[0129] The quality prediction process (method for predicting the quality of ready-mixed concrete) includes a model building process in the training phase and a quality evaluation process in the evaluation phase. In the quality prediction process, the model building process is performed before the quality evaluation process.

[0130] The model building process includes a preparation process and a building process. The preparation process is the process of preparing training data TD, which consists of multiple training datasets. In the preparation process, for example, an operator such as a worker inputs M batches of time-series data into the quality prediction device 10. Then, in the preparation process, the model building unit 26 may generate multiple training datasets from the M batches of time-series data using the method illustrated in Figure 4 or Figure 5. At least a part of the preparation process may be performed by the model building unit 26 of the quality prediction device 10.

[0131] The construction process involves building a predictive model M by performing machine learning using training data TD, which consists of multiple training datasets. The construction process may be performed by the model construction unit 26 of the quality prediction device 10. Each of the multiple training datasets that make up training data TD may be associated with input information including one or more loading features FL and elapsed time tp, and with the correct value of quality such as slump. As shown in Table 1, when M batches of time series data are obtained, multiple (for example, about 15 to 120) training datasets may be prepared for each batch of time series data.

[0132] In the construction phase, the prediction model M may be constructed using a decision tree algorithm (e.g., LightGBM). In this case, the input information to the prediction model M may be multiple features (features F1 to FN as described above) including one or more loading features FL and elapsed time tp. In each of the multiple training datasets in the training data TD, the training features F1 to FN may be associated with the correct values ​​of quality such as freshness. In the construction phase, the prediction model M may be constructed such that one or more features selected from the features F1 to FN are associated with the predicted values ​​of quality such as freshness using a decision tree.

[0133] The quality evaluation process includes a first acquisition process, a second acquisition process, and a prediction process. The first acquisition process is a process of acquiring one or more load feature quantities FL from time-series data of the power load value of the mixer 114 up to measurement point mp while the mixer 114 is still in operation. The first acquisition process may be performed by the first acquisition unit 24 of the quality prediction device 10.

[0134] In the first acquisition step, one or more loading features FL may be obtained from time-series data from the start of mixing in the mixer 114 to measurement time mp (see also Figure 4). Alternatively, in the first acquisition step, one or more loading features FL may be obtained from time-series data from a predetermined time before measurement time mp to measurement time mp (see also Figure 5). The range of time-series data to be focused on when obtaining one or more loading features FL between the training phase (model building step) and the evaluation phase (quality evaluation step) is set under the same conditions.

[0135] The second acquisition step is a step in which input information for evaluation is acquired, which includes one or more load feature quantities FL obtained in the first acquisition step and the time from a predetermined reference time to measurement time mp. The second acquisition step may be performed by the second acquisition unit 25 of the quality prediction device 10. In the input information for evaluation, the time from a predetermined reference time to measurement time mp may be the elapsed time tp (elapsed time from the start of mixing) as described above. In the evaluation phase, the measurement time mp may be set in accordance with the timing of executing the prediction step, as will be described later.

[0136] The prediction process is a process of obtaining a predicted quality value corresponding to the evaluation input information obtained in the second acquisition process, using a prediction model M that was previously constructed by machine learning in the construction process to output the quality of the ready-mixed concrete in response to the input information. The prediction process may also be performed by the prediction unit 30 of the quality prediction device 10. In the prediction process, for example, the evaluation input information obtained in the second acquisition process is input to the prediction model M, and a predicted value of the fresh concrete properties output from the prediction model M is obtained. The quality, such as the fresh concrete properties, for which the predicted value is obtained in the prediction process may be the quality of the ready-mixed concrete after the operation of the mixer 114 has finished and the concrete has been discharged from the mixer 114.

[0137] The output process is a process of outputting the predicted quality values ​​obtained in the prediction process. The output process may be executed by the output unit 32 of the quality prediction device 10. In the output process, for example, the predicted quality values ​​such as fresh properties obtained in the prediction process are displayed on the monitor 14. The prediction process and the output process may be executed while the operation of the mixer 114 related to the ready-mixed concrete to be evaluated (the mixing operation by the mixer 114 to obtain the ready-mixed concrete) is continuing. In this case, the first acquisition process and the second acquisition process may also be executed while the operation of the mixer 114 related to the ready-mixed concrete to be evaluated is continuing. The measurement time point mp, which is the timing for acquiring input information, may be a time corresponding to the timing when the prediction process (or the first acquisition process) is executed.

[0138] The measurement time point mp, which represents the timing for quality evaluation, may be set to match the stage in which the measurement time point mp is changed in the training data TD (for example, any of the time points among 3 seconds, 4 seconds, ..., 30 seconds). A series of steps including the first acquisition step, the second acquisition step, the prediction step, and the output step may be repeatedly executed for each stage in which the measurement time point mp is changed in the training data TD.

[0139] <Processing flow in the model building process> Figure 10 illustrates the processing flow performed in the model building process. This model building process is performed before the production phase, which includes the above-mentioned manufacturing process, is executed in the manufacturing equipment 100, or in the early stages after the production phase has started. In this model building process, for example, the ready-mixed concrete actually produced in the manufacturing equipment 100 is used as ready-mixed concrete for training.

[0140] In the processing flow shown in Figure 10, step S11 is executed first. In step S11, for example, an operator such as a worker prepares M batches (M data points) of time-series data, which are input to the quality prediction device 10 via operation by the input device 12. Each of the M batches of time-series data is time-series data of the power load value over the entire period T during which the mixer 114 operated to obtain training concrete. Each of the M batches of time-series data is associated with the correct value of the quality (e.g., fresh properties) obtained by evaluating the training concrete. When predicting two or more types of quality, each of the M batches of time-series data may be associated with the correct values ​​of two or more types of quality.

[0141] Next, the quality prediction device 10 executes steps S12 and S13. In step S12, for example, the model building unit 26 sets i, which represents a variable, to 1. In step S13, the elapsed time tp is set to an initial value. The initial value of the elapsed time tp may be set in advance by the operator, and in the method illustrated in Figure 4 and Table 1, the initial value of the elapsed time tp is set to 3 seconds.

[0142] Next, the quality prediction device 10 executes step S14. In step S14, for example, the model building unit 26 calculates one or more load features FL from the time series data up to the measurement time mp corresponding to the current elapsed time tp. At this time, the model building unit 26 links the calculation result of one or more load features FL with the current elapsed time tp.

[0143] Next, the quality prediction device 10 performs step S15. In step S15, for example, the model building unit 26 updates the elapsed time tp by adding the change amount Δt to the current elapsed time tp. The change amount Δt may be set in advance by the operator, and in the method illustrated in Figure 4 and Table 1, the change amount Δt is set to 1 second.

[0144] Next, the quality prediction device 10 executes step S16. In step S16, for example, the model building unit 26 determines whether the elapsed time tp updated in step S15 is greater than the total period T (the time from time t1 to time t2). If the model building unit 26 prepares training data TD by focusing on a period shorter than the total period T, it may also determine whether the elapsed time tp updated in step S15 is greater than the period it focused on. If it is determined in step S16 that the elapsed time tp is less than or equal to the total period T (step S16: NO), the quality prediction device 10 returns to step S14. In step S14, which is executed for the second time, the elapsed time tp is set to a different value than the first time, and a loading feature FL of 1 or more is calculated.

[0145] On the other hand, in step S16, if it is determined that the elapsed time tp after the update in step S15 is greater than the total period T (step S16: YES), the quality prediction device 10 proceeds to step S17. In this way, unless it is determined in step S16 that the elapsed time tp is greater than the total period T, steps S14 and S15 are repeatedly executed. As a result, from one batch of time-series data, multiple training datasets (J training datasets) are generated in which at least one or more load features FL, elapsed time tp, and correct values ​​of quality such as freshness are associated with each other. J represents the number of times steps S14 and S15 are repeated. In step S17, for example, the model building unit 26 increments the variable i.

[0146] Next, the quality prediction device 10 executes step S18. In step S18, for example, the model building unit 26 determines whether the variable i is greater than or equal to M. If it is determined in step S18 that the variable i is less than or equal to M (step S18: NO), the quality prediction device 10 returns to step S13. As a result, a training dataset is generated from the time series data of the next batch.

[0147] On the other hand, if it is determined in step S18 that the variable i is greater than M (step S18: YES), the quality prediction device 10 proceeds to step S19. In this way, the series of processes including steps S13 to S18 are repeatedly executed unless it is determined that the variable i is greater than M. As a result, multiple training datasets (J × M training datasets) are generated from each of the M batches of time-series data, with at least one or more load features FL, elapsed time tp, and correct values ​​for quality such as freshness being associated with each other.

[0148] In step S19, for example, the model building unit 26 constructs a predictive model M by performing machine learning based on multiple training datasets generated up to the execution of step S18. The model building unit 26 may construct the predictive model M using a neural network (for example, a deep neural network).

[0149] In step S19, the model building unit 26 may build a predictive model M using machine learning with a decision tree algorithm. The model building unit 26 may build the predictive model M by performing machine learning with ensemble learning in the machine learning with a decision tree algorithm. In this case, the model building unit 26 may build the predictive model M by performing machine learning using boosting (for example, gradient boosting) in the machine learning with ensemble learning using a decision tree algorithm. In one example, the model building unit 26 builds the predictive model M using LightGBM developed by Microsoft (registered trademark). After the model building unit 26 builds the predictive model M, the model storage unit 28 stores the built predictive model M.

[0150] In the processing flow illustrated in Figure 10, after M batches of time-series data are obtained (i.e., after the mixing of M batches is completed), the calculation of loading features FL is performed for each time-series data. However, each time mixing is performed, a series of processes including steps S13 to S16 may be executed for the time-series data involved in that mixing. Furthermore, a series of processes including steps S13 to S16 may be performed in parallel with the execution of mixing, in which case, once one mixing is completed, the process returns to step S13.

[0151] <Processing flow in the quality evaluation process> Figure 11 illustrates a processing flow performed in the quality evaluation process. The processing flow illustrated in Figure 11 is performed during a period that overlaps with at least a portion of the period during which the above manufacturing process for the ready-mixed concrete being evaluated is carried out.

[0152] First, the quality prediction device 10 executes step S31 with all the materials necessary to obtain the ready-mixed concrete to be evaluated loaded into the mixer 114. In step S31, for example, the operation control unit 22 controls the mixer drive unit 114b to start mixing by the mixer 114 (starts supplying power to the mixer drive unit 114b). After the execution of step S31, the first acquisition unit 24 may repeatedly measure the power load value at a predetermined measurement cycle.

[0153] Next, the quality prediction device 10 executes step S32. In step S32, for example, the first acquisition unit 24 waits until the evaluation start timing is reached. The evaluation start timing is the timing at which the quality of the ready-mixed concrete produced in the processing flow begins to be evaluated, and is set in advance by an operator such as a worker. The evaluation start timing may also correspond to the initial value of the elapsed time tp in the training phase.

[0154] Next, the quality prediction device 10 executes steps S33 and S34. In step S33, for example, the first acquisition unit 24 calculates one or more load features FL from the time-series data up to measurement time mp, with the current time being the measurement time mp (in other words, the time from the start of mixing by the mixer 114 to the present as elapsed time tp). In step S34, for example, the second acquisition unit 25 acquires input information including one or more load features FL calculated in step S33 and the elapsed time tp corresponding to the measurement time mp in which the load features FL were calculated. As described above, the input information acquired in step S34 may include other information besides one or more load features FL and elapsed time tp.

[0155] Next, the quality prediction device 10 executes steps S35 and S36. In step S35, for example, the prediction unit 30 inputs the input information obtained in step S34 into the prediction model M stored in the model holding unit 28 to obtain predicted values ​​for the quality of the fresh concrete to be evaluated, such as its fresh properties. In step S36, for example, the output unit 32 outputs the predicted quality values ​​calculated in step S35. In one example, the output unit 32 displays the predicted quality values ​​on the monitor 14. In this case, the output unit 32 may also display the predicted quality values ​​on the monitor 14 along with the elapsed time tp.

[0156] Next, the quality prediction device 10 executes step S37. In step S37, for example, the quality prediction device 10 waits until time Δte has elapsed from the time it started executing step S33. Time Δte is set in advance by an operator, such as a worker. Time Δte may be the same value as the change range Δt, which represents the range by which the measurement time mp is changed in the training phase. For the sake of explanation, the processing flow of the training phase has been described first, but the initial value of the elapsed time tp and the change range Δt in the training phase may be set according to the timing at which quality evaluation is to be started in the evaluation phase and the interval to be left when continuing the evaluation.

[0157] Next, the quality prediction device 10 executes step S38. In step S38, for example, the operation control unit 22 determines whether the conditions for ending the operation of the mixer 114 are met. In one example, the operation control unit 22 determines whether a predetermined set time has elapsed since the execution of step S31. In this case, the condition for ending the operation corresponds to the set time elapsed from the time the mixer 114 started operating. If it is determined in step S38 that the conditions for ending the operation are not met (step S38: NO), the process executed by the quality prediction device 10 returns to step S33. After the process returns to step S33, in the next steps S33 and S34, input information is acquired at different elapsed time timings tp, and in step S36, the quality of the ready-mixed concrete to be evaluated is predicted from the input information obtained at different elapsed time timings tp.

[0158] On the other hand, if it is determined in step S38 that the conditions for the end of operation of the mixer 114 have been met (step S38: YES), the quality prediction device 10 proceeds to step S39. In this way, the series of processes including steps S33 to S38 are repeatedly executed unless it is determined that the conditions for the end of operation of the mixer 114 have been met. As a result, the series of processes including acquisition of input information, quality prediction, and output of prediction results are repeated, with the elapsed time tp being varied sequentially. Note that even if the elapsed time tp is at different timings, the quality to be predicted is the quality of the ready-mixed concrete after it has been discharged from the mixer 114 (the same quality).

[0159] In step S39, for example, the operation control unit 22 controls the mixer drive unit 114b to terminate the operation of the mixer 114 (by stopping the power supply to the mixer drive unit 114b). Subsequently, the operation control unit 22 controls the mixer 114, etc., so that the ready-mixed concrete to be evaluated is discharged from the mixer 114. The quality prediction device 10 may repeat the series of processes including steps S31 to S39 for each batch, or for each batch.

[0160] [Differentiation] The processing flow shown in Figures 10 and 11 is an example and can be modified as appropriate. In the above processing flow, one step and the next step may be executed in parallel, and some steps may be executed in a different order than the example above. In place of at least some of the steps in the above processing flow, or in addition to all of the steps in the above processing flow, steps with content different from the example above may be executed.

[0161] The specific machine learning algorithm used to construct the predictive model M is not particularly limited. The model building unit 26 may construct the predictive model M by performing machine learning using algorithms other than decision trees and neural networks.

[0162] When training data TD is prepared by gradually changing the measurement time mp (elapsed time tp), the model construction unit 26 may construct a predictive model capable of outputting a predicted quality value for each measurement time mp by machine learning based on a portion of the training data TD at that measurement time mp. This predictive model for each measurement time mp (hereinafter referred to as "sub-predictive model") is a model that outputs a predicted quality value according to input information that does not include the time used to identify the measurement time mp. In this case, the predictive model M constructed by the model construction unit 26 consists of a plurality of sub-predictive models and a part that selects a sub-predictive model from among the plurality of sub-predictive models according to the measurement time mp at the time of evaluation.

[0163] In the example described above, the prediction process is performed while the mixer 114 is operating. However, the quality prediction device 10 may predict the quality, such as the fresh concrete properties, based on the measurement time mp during the operation of the mixer 114, after the fresh concrete has been discharged from the mixer 114. In the example described above, during the evaluation phase, the quality is predicted by gradually changing the measurement time mp in accordance with the training data TD. However, the quality prediction device 10 may calculate a representative predicted value indicating the quality of the fresh concrete by using the statistical value (e.g., arithmetic mean) of the predicted values ​​at multiple measurement time mp. When the average of the predicted values ​​at each of the multiple measurement time mp is used as the representative predicted value, predicted values ​​that deviate significantly from other predicted values ​​(e.g., deviating from ±1σ, ±2σ, or ±3σ when "σ" is the standard deviation) may be excluded from the calculation of the average. The quality prediction device 10 may also use the final (last obtained) predicted value among the predicted values ​​at multiple measurement time mp as the representative predicted value.

[0164] The output unit 32 may output a predicted quality value to the operation control unit 22 included in the quality prediction device 10, or to a control device for controlling the manufacturing device 100. In this case, the operation control unit 22 or the control device may control devices such as the mixer 114 included in the manufacturing device 100 based on the predicted quality value.

[0165] In addition to predicting the quality of the ready-mixed concrete using the quality prediction device 10, the manufacturing device 100 may periodically measure the quality of the ready-mixed concrete, such as its fresh properties, at a rate of 1 to 100 times per day. In this case, the prediction model M may be updated based on the measured quality of the ready-mixed concrete and the input information (e.g., feature quantities F1 to FN) obtained at the time the measured value was acquired. In the above prediction process, the quality of the ready-mixed concrete may be predicted using the updated prediction model M. Even when the updated prediction model M is used, the process of predicting the quality of the ready-mixed concrete based on the prediction model M and the evaluation input information acquired in the second acquisition process remains unchanged.

[0166] In the example described above, the quality prediction device 10 predicts the quality of the ready-mixed concrete after it has been mixed by the mixer 114 of the manufacturing device 100 and before it is 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 also predict the quality of the ready-mixed concrete during transport or upon arrival at the site (at the time of unloading). For example, the transport vehicle 200 is equipped with a mixer 214 for agitating the ready-mixed concrete (see Figure 1). The mixer 214 in the transport vehicle 200, such as an agitator vehicle, is also called a drum.

[0167] If the manufacturing system 1 includes a transport vehicle 200, at least a portion of the input information to the prediction model M (e.g., feature quantities F1 to FN) may include feature quantities obtained when mixing the ready-mixed concrete using the mixer 214. One or more load feature quantities FL may be one or more feature quantities relating to the power load value of the mixer 214 when mixing the ready-mixed concrete. The first acquisition unit 24 of the quality prediction device 10 may calculate one or more load feature quantities FL from time-series data up to measurement time mp during the mixing operation by the mixer 214. The prediction unit 30 may predict the quality of the ready-mixed concrete at the time after it has been discharged from the mixer 214 while the mixing operation by the mixer 214 is continuing. The input information to the prediction model M may include information related to the transportation of the ready-mixed concrete (e.g., elapsed time from departure, planned transportation time, transportation distance, or ambient temperature and vibration during transportation). In this case, two or more batches of ready-mixed concrete may be mixed in the mixer 214. In this case as well, the same quality predictions as those for one batch of ready-mixed concrete can be made.

[0168] The input information for the prediction model M may include one or more pieces of information selected from the following for mixers 114 and 214: type, model, product name, manufacturer name, year of manufacture, rated capacity, total output, production capacity, main unit mass, operating no-load mass, external dimensions, inclination angle (in the case of drum type), rotational speed (drum, stirring blades, stirring shaft), and mixing performance (deviation rate of air content in concrete, deviation rate of mortar content in concrete, deviation rate of coarse aggregate in concrete, deviation rate of consistency (slump), deviation rate of compressive strength), as well as information regarding the order in which materials are added, mixing amount, addition time, mixing time, discharge time, re-addition time, and cycle time.

[0169] The input information for the prediction model M may include one or more types of information selected from the following: outdoor temperature and humidity, atmospheric pressure, solar radiation, sunshine duration, rainfall, wind speed, wind direction, weather, climate, temperature of each concrete material, temperature and humidity of the place where the materials are stored (in containers such as silos and storage bottles), and temperature of the mixer or inside the mixer; as well as vehicle information of the truck agitator, load capacity, transport time, transport distance, temperature and humidity of the drum or inside the drum, information on road traffic conditions, and information on vibrations affecting the vehicle.

[0170] In the example described above, the ready-mixed concrete is manufactured at a location other than the construction site (manufacturing equipment 100). The location where the ready-mixed concrete is manufactured is not limited to this example. Concrete materials (for example, concrete materials excluding water) may be transported to the construction site by a transport vehicle, and the ready-mixed concrete may be manufactured at the site. In this case, water may be added to the concrete materials transported by the transport vehicle using a mixer installed on the transport vehicle, and the materials may be mixed to obtain the ready-mixed concrete. The quality prediction device 10 may predict the quality of the ready-mixed concrete, such as its fresh properties, at the site when the ready-mixed concrete is manufactured there. The first acquisition unit 24 of the quality prediction device 10 may acquire feature quantities from time-series data during the operation of the mixer installed on the transport vehicle when the ready-mixed concrete is manufactured at the site, as at least a part of one or more load feature quantities FL.

[0171] The quality prediction device 10 does not necessarily have a function to control the manufacturing device 100. In one of the various examples described above, at least some of the matters described in the other examples may be combined.

[0172] [Verification of prediction results using predictive models] Next, we will explain the results of verifying the freshness property prediction results of a machine learning-based prediction model M using a dataset with known ground truth values. In this verification, we compared the prediction results of prediction model M, which can predict quality during mixing in mixer 114, with the prediction results of a comparative prediction model that predicts quality based on information up to the point when mixing in mixer 114 is completed. The comparative prediction model can be said to be a model that cannot make predictions until mixing is complete. For the sake of explanation, below we will refer to the prediction by prediction model M as "prediction during mixing" and the prediction by the comparative prediction model as "prediction after mixing". In addition, the comparative prediction model corresponding to prediction model M will simply be referred to as the "comparative model".

[0173] (Example 1: Predicted DNN during mixing) First, we prepared N1 batches of time-series data where the ground truth values ​​for slump, slump flow, and air volume were known, and the mixing time set by mixer 114 was 30 seconds. Specifically, N1 is 5982. Then, for each time-series data in the N1 batches, we sequentially calculated one or more loading features FL by gradually changing the elapsed time tp from 3 seconds in 1-second increments up to 30 seconds. This resulted in N2 (N1 × 28) datasets, where each dataset had features F1 to FN, each containing one or more loading features FL and elapsed time tp, associated with the ground truth values ​​for slump, slump flow, and air volume. Of the N2 datasets, 92% (N3, N2 × 0.92) were used as multiple training datasets to constitute the training data TD, and 8% (N4, N2 × 0.08) were used as multiple test datasets for evaluating prediction results, where the ground truth values ​​were known. By splitting the original N2 datasets into batches, we obtained training and test datasets.

[0174] A prediction model M was constructed using a deep neural network. In constructing the prediction model M, the neural network's hidden layers consisted of eight fully connected layers, and normalization of the input data and partial loss of connections in the hidden layers were not performed. Adam was used as the weight update formula, and HuberLoss was used as the loss function. Machine learning was performed so that a single prediction model M would output three types of data as freshness characteristics: slump, slump flow, and air volume.

[0175] Using N4 test datasets, we compared the predicted values ​​obtained by the prediction model M with the ground truth values ​​of the quality included in the dataset for each quality type. Specifically, a ground truth was defined as a value that falls within the range obtained by adding a predetermined tolerance to the ground truth values ​​in the test dataset. The accuracy rate, which represents the ratio of the number of ground truth datasets to the total number of test datasets, was calculated as an evaluation metric.

[0176] For slump, tolerances were set to ±0.5cm, ±1.0cm, ±1.5cm, ±2.0cm, and ±2.5cm, and the accuracy rate was calculated for each tolerance. For slump flow, tolerances were set to ±2.5cm, ±3.0cm, ±4.0cm, ±5.0cm, ±6.0cm, and ±7.5cm, and the accuracy rate was calculated for each tolerance. For air volume, tolerances were set to ±0.3%, ±0.5%, ±0.7%, ±1.0%, ±1.2%, and ±1.5%, and the accuracy rate was calculated for each tolerance. The calculation results for the accuracy rate of each quality will be described later.

[0177] (Comparative Example 1: Predicted DNN after mixing) N1 datasets were prepared from the N1 batch of time-series data used in Example 1. In each of the N1 datasets, multiple features that serve as input to the comparison model were associated with the ground truth values ​​of slump, slump flow, and air volume. The above multiple features were the same as the features F1 to FN used in Example 1, except that the elapsed time tp was fixed to the mixing setting time (30 seconds) and one or more loading features FL were calculated from the entire time-series data. Of the N1 datasets, 92% (N5, N1 × 0.92) were used as multiple training datasets to constitute the training data, and 8% (N6, N1 × 0.08) were used as multiple test datasets for evaluating the prediction results, as the ground truth values ​​were known.

[0178] A comparison model was constructed by performing machine learning under the same conditions as in Example 1, except that the training data was different. For each of the slump, slump flow, and air volume, the accuracy rate as an evaluation metric was calculated by comparing the predicted values ​​from the comparison model with the ground truth values ​​from the test datasets, using N6 test datasets, as in Example 1.

[0179] (Example 2: Prediction during mixing - LightGBM) In this study, 72% of the N2 datasets used in Example 1, specifically N7 (N2 × 0.72) datasets, were used as multiple training datasets to constitute the training data TD. A prediction model M was constructed using LightGBM, a type of decision tree algorithm. Specifically, the prediction model M was constructed using "LightGBM (version 4.3.0)" provided by Microsoft. Models for predicting slump, slump flow, and air volume were constructed separately (i.e., three types of prediction models M were constructed). For each of slump, slump flow, and air volume, similar to Example 1, N4 test datasets were used to compare the predicted values ​​from the prediction model M corresponding to the quality type with the ground truth values ​​in the test datasets, and the accuracy rate as an evaluation metric was calculated.

[0180] (Comparative Example 2: Prediction after mixing - LightGBM) Of the N1 datasets used in Comparative Example 1, 72% (N8, or N1 × 0.72) were used as multiple training datasets to constitute the training data TD, and a predictive model M was constructed using LightGBM. A comparative model was constructed by performing machine learning under the same conditions as in Example 2, except that the training data was different. The hyperparameter settings were the same when constructing the predictive models between Example 2 and Comparative Example 2. For each of slump, slump flow, and air volume, as in Comparative Example 1, N6 test datasets were used, and the predicted values ​​from the comparative model were compared with the ground truth values ​​of the test datasets to calculate the accuracy as an evaluation metric.

[0181] (Example 3: Prediction during mixing · XGBoost) Using the same N7 training datasets as in Example 2 as training data TD, a prediction model M was constructed using XGBoost, a type of decision tree algorithm. Specifically, the prediction model M was constructed using the well-known library "XGBoost (version 1.5.2)". Similar to Example 2, models for predicting slump, slump flow, and air volume were constructed separately (i.e., three types of prediction models M were constructed). For each of slump, slump flow, and air volume, similar to Examples 1 and 2, the prediction values ​​from the prediction model M corresponding to the quality type were compared with the ground truth values ​​in the test dataset using N4 test datasets to calculate the accuracy as an evaluation metric.

[0182] (Comparative Example 3: Prediction after mixing - XGBoost) A predictive model M was constructed using XGBoost, utilizing the same N8 training datasets used in Comparative Example 2. A comparative model was constructed by performing machine learning under the same conditions as in Example 3, except for the difference in training data. The hyperparameter settings were the same for constructing the predictive models in both Example 3 and Comparative Example 3. For each of slump, slump flow, and air volume, the accuracy rate as an evaluation metric was calculated by comparing the predicted values ​​from the comparative model with the ground truth values ​​from the test datasets, using N6 test datasets, similar to Comparative Examples 1 and 2.

[0183] (Verification results) Figure 12 shows the verification results for slump in Examples 1-3 and Comparative Examples 1-3. The graph in Figure 12(a) shows the verification results for Example 1 and Comparative Example 1, respectively, with "During Mixing_DNN" corresponding to Example 1 and "After Mixing_DNN" corresponding to Comparative Example 1. The graph in Figure 12(b) shows the verification results for Example 2 and Comparative Example 2, respectively, with "During Mixing_LGBM" corresponding to Example 2 and "After Mixing_LGBM" corresponding to Comparative Example 2. The graph in Figure 12(c) shows the verification results for Example 3 and Comparative Example 3, respectively, with "During Mixing_XGBoost" corresponding to Example 3 and "After Mixing_XGBoost" corresponding to Comparative Example 3.

[0184] The results shown in Figures 12(a), 12(b), and 12(c) indicate that there is no significant difference in the accuracy of slump prediction between Example 1 and Comparative Example 1, between Example 2 and Comparative Example 2, and between Example 3 and Comparative Example 3. In other words, predicting the slump of the ready-mixed concrete after the operation is completed based on time-series data up to a point in the operation before the mixing by the mixer 114 is completed yields the same prediction accuracy as predicting the slump based on time-series data up to the completion of the operation of the mixer 114.

[0185] Comparing Example 1 with Examples 2 and 3, it can be seen that the accuracy of predicting slump is higher in Examples 2 and 3 compared to Example 1. For example, focusing on the accuracy with a tolerance of ±1.0 cm, the accuracy in Example 1 is approximately 68%, while the accuracy in Example 2 is approximately 82%, and the accuracy in Example 3 is approximately 84%. In other words, it can be seen that the accuracy of predicting slump can be improved when using a decision tree algorithm compared to when using a deep neural network.

[0186] Figure 13 shows the verification results for slump flow in Examples 1-3 and Comparative Examples 1-3. The graph shown in Figure 13(a) corresponds to the graph shown in Figure 12(a), except for the quality type and tolerance settings. The graph shown in Figure 13(b) corresponds to the graph shown in Figure 12(b), except for the quality type and tolerance settings. The graph shown in Figure 13(c) corresponds to the graph shown in Figure 12(c), except for the quality type and tolerance settings.

[0187] The results shown in Figures 13(a), 13(b), and 13(c) indicate that there is no significant difference in the accuracy of the slump flow prediction between Example 1 and Comparative Example 1, between Example 2 and Comparative Example 2, and between Example 3 and Comparative Example 3. In other words, predicting the slump flow of the ready-mixed concrete after the operation is completed based on time-series data up to a point in the operation before the mixing by the mixer 114 is completed yields the same level of prediction accuracy as predicting the slump flow based on time-series data up to the completion of the operation of the mixer 114.

[0188] Comparing Example 1 with Examples 2 and 3, it can be seen that the accuracy of predicting slump flow is higher in Examples 2 and 3 compared to Example 1. For example, focusing on the accuracy with a tolerance of ±3.0 cm, the accuracy in Example 1 is approximately 64%, while the accuracy in Example 2 is approximately 82%, and the accuracy in Example 3 is approximately 83%. In other words, it can be seen that the prediction accuracy of slump flow can be improved when using a decision tree algorithm compared to when using a deep neural network.

[0189] Figure 14 shows the verification results for air volume in Examples 1-3 and Comparative Examples 1-3. The graph in Figure 14(a) corresponds to the graph in Figure 12(a), except for the quality type and tolerance settings. The graph in Figure 14(b) corresponds to the graph in Figure 12(b), except for the quality type and tolerance settings. The graph in Figure 14(c) corresponds to the graph in Figure 12(c), except for the quality type and tolerance settings.

[0190] The results shown in Figures 14(a), 14(b), and 14(c) indicate that there is no significant difference in the accuracy of predicting the amount of air between Example 1 and Comparative Example 1, between Example 2 and Comparative Example 2, and between Example 3 and Comparative Example 3. In other words, predicting the amount of air in the fresh concrete after the operation is completed based on time-series data up to a point in the operation before the mixing by mixer 114 is completed yields the same level of prediction accuracy as predicting the amount of air based on time-series data up to the completion of the operation of mixer 114.

[0191] Comparing Example 1 with Examples 2 and 3, it can be seen that the accuracy of predicting air volume is higher in Examples 2 and 3 compared to Example 1. For example, focusing on the accuracy with a tolerance of ±0.5%, the accuracy in Example 1 is approximately 56%, while the accuracy in Example 2 is approximately 73%, and the accuracy in Example 3 is approximately 75%. In other words, it can be seen that the accuracy of predicting air volume can be improved when using a decision tree algorithm compared to when using a deep neural network.

[0192] (others) Similar results were obtained when multiple training datasets (training data TD) were prepared using the method illustrated in Figure 5, and comparative verifications were performed in the same manner as in Examples 1-3 and Comparative Examples 1-3.

[0193] [Summary of this disclosure] The ready-mix concrete quality prediction method described above includes: a first acquisition step of acquiring one or more load features (FL) from time-series data of the power load value of a mixer (114, 214) that mixes concrete materials or stirs ready-mix concrete up to a measurement point (mp) while the mixer is still in operation; a second acquisition step of acquiring input information including one or more load features (FL) and the time (tp) from a predetermined reference point to a measurement point (mp); and a prediction step of acquiring a predicted value of the quality according to the input information acquired in the second acquisition step, using a prediction model (M) that has been pre-built by machine learning to output the quality of ready-mix concrete in response to the input information.

[0194] As described above, the inventors' verification results confirmed that there is no substantial difference in quality prediction accuracy between a prediction model that includes load features obtained from the entire time-series data of power load values ​​after the mixing operation of the mixer (114,214) and a prediction model (M) that includes load features (FL) obtained from the time-series data up to the middle of the mixing operation of the mixer (114,214) as input data. By the above method, the quality of ready-mixed concrete can be accurately predicted based on the time-series data up to the middle of the operation of the mixer (114,214). This improves the convenience when predicting quality or when using the quality prediction results. In other words, the above quality prediction method is useful for improving convenience.

[0195] In the ready-mix concrete quality prediction method described above, the quality for which the predicted value is obtained in the prediction process may be the quality of the ready-mix concrete after the operation of the mixer (114,214) has finished and the concrete has been discharged from the mixer (114,214). In this case, it is possible to predict the quality of the ready-mix concrete obtained after the operation of the mixer using data up to a certain point in the time series data, or to predict the quality of the ready-mix concrete obtained after the operation of the mixer at an intermediate stage when all the time series data during operation is not yet available. Therefore, this method is useful for improving the convenience when predicting quality or when using the quality prediction results.

[0196] The ready-mix concrete quality prediction method described above may further include an output step that outputs the predicted quality value obtained in the prediction step. The prediction step and the output step may be executed while the mixer (114, 214) for the ready-mix concrete to be evaluated is in operation. In this case, the predicted quality value can be obtained earlier while the mixer for the ready-mix concrete to be evaluated is in operation. For example, an operator such as a worker can grasp the prediction result while the mixer is in operation, and if the device uses the predicted quality result, the device can perform processing using the prediction result while the mixer is in operation.

[0197] In the concrete quality prediction method described above, in the first acquisition step, one or more load features (FL) may be acquired from the above time series data from the start of operation of the mixer (114, 214) to the measurement time (mp), or one or more load features (FL) may be acquired from the above time series data from a predetermined time before the measurement time (mp) to the measurement time (mp). Even when one or more load features (FL) are acquired from a portion of the time series data in this way, quality prediction results can be obtained with the same level of accuracy as when load features are acquired from the entire time series data.

[0198] The ready-mix concrete quality prediction method described above may further include a construction step in which a prediction model (M) is constructed by performing machine learning using training data (TD) composed of multiple training datasets. In each of the above multiple training datasets, the above input information may be associated with the above ground truth values ​​for quality. The above multiple training datasets may include two or more datasets in which the above ground truth values ​​for quality are common, but the measurement time (mp) is different from each other. In this case, in the evaluation phase, the ready-mix concrete quality can be predicted at each stage in which the measurement time (mp) is changed.

[0199] In the ready-mix concrete quality prediction method described above, one or more load features (FL) may include at least one of the values ​​at a predetermined point in time in the time series data up to the measurement point (mp), and a statistic obtained from the time series data up to the measurement point (mp). The statistic may be one or more values ​​selected from the group consisting of the difference between the maximum value (P3) and the minimum value (P2), the difference between the maximum value (P3) and the final value (P4), the sum, the mean, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness. By using the power load value at a specific point in time and at least one of the statistic obtained from some of the time series data as features, as in the method described above, information that better reflects the characteristics of the power load value that fluctuates over time can be added to the input of the prediction model (M).

[0200] The ready-mix concrete quality prediction method described above may further include a construction step in which a prediction model (M) is constructed by performing machine learning using training data (TD) composed of multiple training datasets. The above input information may consist of multiple features (F1~FN) including one or more load features (FL) and the time (tp) from the above reference time to the measurement time (mp). In each of the above multiple training datasets, the multiple features (F1~FN) may be associated with the above correct quality values. In the construction step, the prediction model (M) may be constructed such that one or more features selected from the multiple features (F1~FN) are associated with the above predicted quality values ​​using a decision tree. As explained in the verification results above, it was confirmed that the prediction accuracy of the prediction model (M) can be improved by using a prediction model (M) constructed such that one or more features selected from the multiple features (F1~FN) are associated with the predicted quality values ​​using a decision tree. In other words, this method, which uses a decision tree-based predictive model (M) to predict the quality of ready-mixed concrete, is useful for improving prediction accuracy.

[0201] The quality prediction program described above is a program that causes a computer to execute the above quality prediction method. This quality prediction program, like the quality prediction method described above, is useful for improving convenience.

[0202] The ready-mix concrete manufacturing method described above includes a manufacturing process for manufacturing ready-mix concrete and a quality prediction process for predicting the quality of the ready-mix concrete manufactured in the manufacturing process using the quality prediction method described above. This manufacturing method, like the quality prediction method described above, is useful for improving convenience.

[0203] The ready-mixed concrete quality prediction device (10) described above comprises: a first acquisition unit (24) that acquires one or more load features (FL) from time-series data of the power load value of a mixer (114,214) that mixes concrete materials or stirs ready-mixed concrete up to a measurement point (mp) while the mixer (114,214) is in operation; a second acquisition unit (25) that acquires input information including one or more load features (FL) and the time (tp) from a predetermined reference point to the measurement point (mp); and a prediction unit (30) that acquires a predicted value of the quality according to the input information acquired by the second acquisition unit (25) using a prediction model (M) that has been pre-constructed by machine learning to output the quality of ready-mixed concrete in response to the input information. This quality prediction device (10), like the quality prediction method described above, is useful for improving convenience.

[0204] The ready-mix concrete manufacturing system (1) described above comprises a manufacturing apparatus (100) for manufacturing ready-mix concrete and the quality prediction apparatus (10). The quality prediction apparatus (10) predicts the quality of the ready-mix concrete manufactured by the manufacturing apparatus (100). This manufacturing system (1), like the quality prediction method described above, is useful for improving convenience. [Explanation of Symbols]

[0205] 1...Manufacturing system, 10...Quality prediction device, 24...First acquisition unit, 25...Second acquisition unit, F1~FN...Feature quantities, FL...Load feature quantities, 26...Model construction unit, TD...Training data, 30...Prediction unit, M...Prediction model, 100...Manufacturing equipment, 114...Mixer, 114a...Agitation member, 114b...Mixer drive unit, 214...Mixer.

Claims

1. A first acquisition step involves obtaining one or more load features from time-series data of the power load value of a mixer up to a measurement point while the mixer is continuously operating to mix concrete materials or stir ready-mix concrete. A second acquisition step involves acquiring input information including one or more load features and the time from a predetermined reference point to the measurement point. A prediction step includes obtaining a predicted value of the quality corresponding to the input information obtained in the second acquisition step, using a prediction model that has been pre-built by machine learning to output the quality of ready-mixed concrete in response to the input of the aforementioned input information. A method for predicting the quality of ready-mixed concrete.

2. The quality for which the predicted value is obtained in the prediction process is the quality of the ready-mixed concrete after the mixer has finished operating and has been discharged from the mixer. The method for predicting the quality of ready-mixed concrete according to claim 1.

3. The process further includes an output step that outputs the predicted quality value obtained in the prediction step, The prediction step and the output step are performed while the mixer is continuing to operate with respect to the ready-mixed concrete to be evaluated. The method for predicting the quality of ready-mixed concrete according to claim 1.

4. In the first acquisition step, one or more load features are acquired from the time-series data from the start of operation of the mixer to the measurement time, or one or more load features are acquired from the time-series data from a predetermined time before the measurement time to the measurement time. The method for predicting the quality of ready-mixed concrete according to claim 1.

5. The process further includes a construction step in which the predictive model is constructed by performing machine learning using training data composed of multiple training datasets. In each of the aforementioned training datasets, the input information is associated with the correct value of the quality. The aforementioned multiple training datasets include two or more datasets in which the correct values ​​for quality are common and the measurement times are different from each other. A method for predicting the quality of ready-mixed concrete according to any one of claims 1 to 4.

6. The one or more loading features mentioned above are, The value at a predetermined point in time in the time series data up to the measurement point, and Includes at least one of the statistics obtained from the time-series data up to the aforementioned measurement point, The aforementioned statistics are one or more values ​​selected from the group consisting of the difference between the maximum and minimum values, the difference between the maximum and final values, the sum, the mean, the standard deviation, the coefficient of variation, the median, the first quartile, the third quartile, kurtosis, and skewness. A method for predicting the quality of ready-mixed concrete according to any one of claims 1 to 4.

7. The process further includes a construction step in which the predictive model is constructed by performing machine learning using training data composed of multiple training datasets. The input information includes one or more load features and a plurality of features including the time from the reference time to the measurement time. In each of the aforementioned training datasets, the aforementioned features are associated with the aforementioned quality ground truth values. In the construction step, the prediction model is constructed such that one or more features selected from the plurality of features are associated with the predicted quality values ​​using a decision tree. A method for predicting the quality of ready-mixed concrete according to any one of claims 1 to 4.

8. A quality prediction program that causes a computer to execute the quality prediction method described in any one of claims 1 to 4.

9. The manufacturing process for producing ready-mixed concrete, A quality prediction step includes predicting the quality of the ready-mixed concrete produced by the manufacturing step using the quality prediction method described in any one of claims 1 to 4. A method for manufacturing ready-mixed concrete.

10. A first acquisition unit acquires one or more load features from time-series data of the power load value of a mixer up to a measurement point while the mixer is continuously operating to mix concrete materials or stir ready-mix concrete. A second acquisition unit acquires input information including one or more load feature quantities and the time from a predetermined reference time to the measurement time, The system includes a prediction unit that uses a prediction model, pre-built by machine learning to output the quality of ready-mixed concrete in response to the input information, to acquire a predicted value of the quality corresponding to the input information acquired by the second acquisition unit. A device for predicting the quality of ready-mixed concrete.

11. A manufacturing apparatus for producing ready-mixed concrete, The quality prediction device according to claim 10 comprises, The quality prediction device predicts the quality of the ready-mixed concrete produced by the manufacturing device. A ready-mix concrete manufacturing system.

Images