Aggregate quality prediction method, quality prediction program, ready-mix concrete manufacturing method, aggregate quality prediction device, and ready-mix concrete manufacturing system

Abstract

This invention provides a method, program, and apparatus for predicting aggregate quality, as well as a method and system for manufacturing ready-mixed concrete, all utilizing a predictive model built using machine learning. [Solution] In the quality prediction device, the calculation device for predicting the quality of aggregate includes a model construction unit that constructs a prediction model that outputs physical property values ​​in response to input information by machine learning using training data which associates input information including input image data based on image data obtained by imaging aggregate with physical property values ​​of aggregate, an input data acquisition unit that acquires input information related to the aggregate to be evaluated, and a prediction unit 72 that predicts the quality of the aggregate to be evaluated based on the prediction model and the input information. When the image resolution representing the size of one pixel in the image data is A mm / pixel, and the minimum dimension of the aggregate included in the image data is B mm, the value A / B[1 / pixel] obtained by dividing A by B is 0.20 or more and 2.93 or less.

Description

Technical Field

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

Background Art

[0002] Patent Document 1 discloses a method for estimating the particle size distribution of soil.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The present disclosure provides a method for predicting the quality of aggregates, a quality prediction program, a method for manufacturing fresh concrete, an apparatus for predicting the quality of aggregates, and a manufacturing system for fresh concrete, which enable the quality of aggregates to be managed using a prediction model constructed by machine learning.

Means for Solving the Problems

[0005] [1] A method for predicting the quality of aggregate, comprising: a construction step of constructing a predictive model that outputs the physical properties in response to input of input information by machine learning using training data which associates input information including input image data based on image data obtained by imaging aggregate with physical properties of aggregate; an acquisition step of acquiring the input information relating to aggregate to be evaluated; and a prediction step of predicting the quality of aggregate to be evaluated based on the predictive model constructed in the construction step and the input information acquired in the acquisition step, wherein the image resolution representing the size of one pixel in the image data is A [mm / pixel] and the minimum dimension of aggregate included in the image data is B [mm], and the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less.

[0006] [2] The aggregate quality prediction method according to [1] above, wherein the training data is composed of a plurality of training datasets, each dataset containing the input information and the physical property values ​​associated with the input information, and when the number of the plurality of training datasets is C, the value obtained by dividing C by A (C / A) is 600 or more and 10000 or less.

[0007] [3] The aggregate quality prediction method according to [1] or [2] above, wherein the input image data is grayscale image data.

[0008] [4] The aggregate quality prediction method according to any one of [1] to [3] above, wherein B, which represents the minimum dimension of the aggregate included in the imaged data, is 5.0 or less.

[0009] [5] A quality prediction program for causing a computer to execute one of the aggregate quality prediction methods described in any one of [1] to [4] above.

[0010] [6] A method for producing ready-mixed concrete, comprising: a manufacturing step of mixing materials containing aggregate to produce ready-mixed concrete; and a quality prediction step of predicting the quality of at least a portion of the aggregate used in the manufacturing step using the aggregate quality prediction method described in any one of [1] to [4] above.

[0011] [7] A device for predicting the quality of aggregate, comprising: a model construction unit that constructs a predictive model that outputs the physical properties in response to input of the input information by machine learning using training data which associates input information including input image data based on image data obtained by imaging aggregate with physical properties of the aggregate; an input data acquisition unit that acquires the input information relating to the aggregate to be evaluated; and a prediction unit that predicts the quality of the aggregate to be evaluated based on the predictive model constructed by the model construction unit and the input information acquired by the input data acquisition unit, wherein the image resolution representing the size of one pixel in the image data is A [mm / pixel] and the minimum dimension of the aggregate included in the image data is B [mm], and the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less.

[0012] [8] A ready-mix concrete manufacturing system comprising a manufacturing apparatus for mixing materials containing aggregate to produce ready-mix concrete, and a quality prediction apparatus as described in [7] above, wherein the quality prediction apparatus predicts the quality of at least a portion of the aggregate used in the production of ready-mix concrete by the manufacturing apparatus. [Effects of the Invention]

[0013] This disclosure provides a method for predicting aggregate quality, a quality prediction program, a ready-mix concrete manufacturing method, an aggregate quality prediction device, and a ready-mix concrete manufacturing system that enable the management of aggregate quality using a predictive model constructed by machine learning. [Brief explanation of the drawing]

[0014] [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 hardware configuration of the arithmetic unit included in the quality prediction device. [Figure 3] Figure 3 is a block diagram showing an example of the functional configuration of the arithmetic unit. [Figure 4] Figure 4 is a diagram illustrating the arithmetic process when predicting quality using the prediction model. [Figure 5] Figure 5 is a schematic diagram for explaining the image resolution. [Figure 6] Figure 6 is a diagram showing an example of the imaging image data for training. [Figure 7] Figure 7 is a flowchart illustrating the processing flow executed in the training phase. [Figure 8] Figure 8 is a flowchart illustrating the processing flow executed in the evaluation phase. [Figure 9] Figures 9(a) and 9(b) are graphs illustrating the results of comparing the predicted values and the correct values in the test dataset.

Mode for Carrying Out the Invention

[0015] 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.

[0016] [Fresh Concrete Manufacturing System] In FIG. 1, a fresh concrete manufacturing system according to an embodiment is schematically shown. The manufacturing system 1 shown in FIG. 1 is a system for manufacturing fresh concrete. The manufacturing system 1 is installed, for example, in a factory for manufacturing fresh concrete. The manufacturing system 1 manufactures (produces) fresh concrete by kneading concrete materials.

[0017] The concrete material used in the manufacturing system 1 includes cement, admixtures, coarse aggregates, fine aggregates, water, admixtures, and the like. Examples of the coarse aggregates include natural coarse aggregates or artificial coarse aggregates. Further, examples of the coarse aggregates include gravel, crushed stone, slag coarse aggregates, lightweight coarse aggregates, recycled coarse aggregates, recovered aggregates, or coarse aggregates obtained by mixing these. The gravel is mountain gravel, land gravel, river gravel, or sea gravel, etc. The slag coarse aggregates are blast furnace slag aggregates, ferronickel slag aggregates, electric furnace oxidized slag aggregates, or coal gasification slag aggregates, etc. The lightweight coarse aggregates are natural lightweight aggregates, by-product lightweight aggregates, or artificial lightweight aggregates, etc. The coarse aggregates may include crushed rock or crushed limestone.

[0018] Examples of the fine aggregates include natural aggregates or artificial aggregates. Further, examples of the fine aggregates include sand, crushed sand, slag fine aggregates, lightweight fine aggregates, recycled fine aggregates, recovered aggregates, or fine aggregates obtained by mixing these. The sand is mountain sand, land sand, river sand, or sea sand, etc. The slag fine aggregates are blast furnace slag aggregates, ferronickel slag aggregates, copper slag aggregates, electric furnace oxidized slag aggregates, or coal gasification slag aggregates, etc. The lightweight fine aggregates are natural lightweight aggregates, by-product lightweight aggregates, or artificial lightweight aggregates, etc.

[0019] Examples of the rock types of the crushed stone and crushed sand include igneous rocks, sedimentary rocks, metamorphic rocks, silica, limestone, dolomite, or calcareous shale, etc. The igneous rocks are granite, diorite, porphyry, rhyolite, diabase, rhyolite, andesite, basalt, or serpentine, etc. The sedimentary rocks are conglomerate, sandstone, shale, slate, or tuff, etc. The metamorphic rocks are gneiss or schist, etc. As the coarse aggregates and fine aggregates, those obtained by mixing two or more of the above-exemplified ones may be used.

[0020] In addition to the function of manufacturing fresh concrete, the manufacturing system 1 has a function of predicting the quality of the aggregates contained in the concrete material. In the present disclosure, the coarse aggregates and fine aggregates may be collectively referred to as "aggregates". In this case, "aggregates" means coarse aggregates, fine aggregates, or both coarse aggregates and fine aggregates.

[0021] Manufacturing system 1 loads the manufactured ready-mix concrete onto transport vehicle C. After loading the ready-mix concrete, transport vehicle C transports the ready-mix concrete to the site where it will be used (for example, a construction site). Examples of transport vehicle C include an agitator truck (mixer truck) or a dump truck. Manufacturing system 1 may also manufacture ready-mix concrete from concrete materials to meet the target quality (required quality) set for each site. For example, the operator of manufacturing system 1 determines the concrete material mix to meet the quality control at the time of shipment based on the target quality at the site, and inputs operation instructions to manufacturing system 1.

[0022] The manufacturing system 1 includes, for example, a material storage area 2, a transport device 8, a manufacturing device 10, and a quality prediction device 50.

[0023] Material storage area 2 is a place for storing concrete materials. Material storage area 2 includes multiple silos 4. The multiple silos 4 are containers for storing at least a portion of the concrete materials, separated by material type. The multiple silos 4 include a silo 4 for storing coarse aggregate, a silo 4 for storing fine aggregate, and a silo 4 for storing cement.

[0024] The transport device 8 is a device that transports concrete materials stored in multiple silos 4 to the manufacturing device 10. The transport device 8 includes, for example, a belt conveyor for transporting concrete materials. The transport device 8 may transport concrete materials by type at different times. In one example, based on operation instructions from a control device included in the manufacturing system 1, a specific material from among the various concrete materials is transferred to the transport device 8 and transported to the manufacturing device 10.

[0025] The manufacturing apparatus 10 is a device that mixes concrete materials including aggregate to produce ready-mixed concrete. The manufacturing apparatus 10 operates based on operation instructions from a control device included in the manufacturing system 1. The manufacturing apparatus 10 includes, for example, a storage jar 12, a measuring jar 14, a collection hopper 16, a mixer 20, and a loading hopper 30.

[0026] The storage jar 12 temporarily stores various concrete materials. Various concrete materials are transported (conveyed) to the storage jar 12 from the material storage area 2 by the transport device 8. The storage jar 12 is configured to store each type of concrete material individually. Hereinafter, "concrete materials" may be simply referred to as "materials." The various materials stored in the storage jar 12 are supplied to the measuring jar 14 as needed.

[0027] The measuring bottle 14 is located below the storage bottle 12. The measuring bottle 14 operates based on operation instructions from the control device of the manufacturing system 1 and weighs various materials individually. When the measuring bottle 14 detects the target amount of material instructed by the control device, it supplies that material to the collection hopper 16. When water is supplied to the measuring bottle 14, an admixture may be mixed into the water. The collection hopper 16 is located below the measuring bottle 14. The collection hopper 16 collects the various materials discharged from the measuring bottle 14 and supplies the collected materials to the mixer 20.

[0028] Mixer 20 is located below the collection hopper 16. Mixer 20 is a device for mixing concrete materials. Mixer 20 produces ready-mixed concrete by mixing aggregate, cement, water, and admixtures. Ready-mixed concrete is discharged from the bottom of mixer 20 into loading hopper 30. Loading hopper 30 is located below mixer 20 and temporarily stores the ready-mixed concrete. Loading hopper 30 supplies the temporarily stored ready-mixed concrete to transport vehicle C.

[0029] The manufacturing apparatus 10 described above is just one example of a ready-mix concrete manufacturing apparatus. A ready-mix concrete manufacturing apparatus can be configured in any way as long as it can mix concrete materials and produce ready-mix concrete. For example, the manufacturing apparatus 10 does not need to have a collection hopper 16, and various materials measured in measuring bottles 14 may be supplied from the measuring bottles 14 to the mixer 20.

[0030] (Quality prediction device) The quality prediction device 50 (aggregate quality prediction device) is a device that predicts the quality of aggregates contained in concrete materials. The quality prediction device 50 may predict the quality of at least some of the aggregates used by the manufacturing device 10. It is known that the quality of ready-mixed concrete produced by the manufacturing device 10 is affected by the quality of the concrete materials. Specifically, the particle size of the aggregates used in the production of ready-mixed concrete affects the fluidity of the ready-mixed concrete (e.g., slump and slump flow). For example, with the same unit water content, the larger the aggregate particle size, the smaller the surface area of ​​the aggregate, which reduces the amount of water it can retain, and the greater the fluidity of the ready-mixed concrete. Conversely, the smaller the aggregate particle size, the larger the surface area of ​​the aggregate, which increases the amount of water it can retain, and the less fluid the ready-mixed concrete may be. Thus, in order to stabilize the quality, including the fluidity, of ready-mixed concrete, it is necessary to control the aggregate particle size.

[0031] The particle size of aggregate can be expressed as the coarseness ratio, and in order to stabilize the quality of ready-mixed concrete, it is necessary to manage, for example, whether the coarseness ratio of the aggregate is within a desired range. The quality prediction device 50 may predict the coarseness ratio of the aggregate as a quality of the aggregate. The quality prediction device 50 may be configured to predict the coarseness ratio of the aggregate from information including image data based on image data obtained by imaging the aggregate, using a prediction model constructed by machine learning.

[0032] At least a portion of the quality prediction device 50 is comprised of one or more computers. The computers comprising at least a portion of the quality prediction device 50 may be personal computers, tablet computers (tablet terminals), smartphones, wearable devices, workstations, server computers, or general-purpose computers.

[0033] If the quality prediction device 50 includes two or more computers, these computers may be connected to each other in a way that allows them to communicate with one another. The quality prediction device 50 may be connected to a control device provided in the manufacturing system 1 in a way that allows it to communicate with one another. The quality prediction device 50 may be a part of the control device provided in the manufacturing system 1 (i.e., it may be included in the control device).

[0034] The quality prediction device 50 includes, for example, a calculation device 52, an input / output device 54, and an imaging device 58. The calculation device 52 is a device that performs calculations to predict the quality of aggregate. The calculation device 52 may also be the main computer body (the main part of the computer) that performs the main functions of the computer that constitutes the quality prediction device 50.

[0035] The input / output device 54 is connected to the arithmetic unit 52. The input / output device 54 is a device that has the function of inputting information indicating instructions from a user, such as a worker, to the arithmetic unit 52, and the function of outputting information from the arithmetic unit 52 to the user, such as a worker. The input / output device 54 may include a keyboard, an operation panel, or a mouse as an input device, and may include a monitor (e.g., a liquid crystal display) as an output device. The input / output device 54 may be a touch panel in which the input and output devices are integrated. The arithmetic unit 52 and the input / output device 54 may be integrated.

[0036] The imaging device 58 is a device (camera) capable of imaging aggregates. The imaging device 58 generates color image data by, for example, imaging aggregates. Hereinafter, the color image data generated by the imaging device 58 will be referred to as "image data P1". The imaging device 58 is a digital camera or video camera that images within the field of view (imaging range) based on visible light. The color image data P1 generated by the imaging device 58 is used when predicting the quality of aggregates. If the imaging device 58 generates video data, still image data included in that video may be acquired as image data P1.

[0037] When the total number of pixels in the captured image data P1 is denoted as "Np", the number of pixels Np may be defined as "number of pixels in the horizontal direction × number of pixels in the vertical direction". The imaging device 58 may be a camera that allows the setting of the number of pixels Np to be changed. The range of the number of pixels Np is not particularly limited, but for example, the number of pixels in the vertical or horizontal direction may be selectable from multiple ranges such as 50 to 3500, 100 to 2500, or 100 to 1500. The aspect ratio, which represents the ratio of the number of pixels in the horizontal direction to the number of pixels in the vertical direction, is also not particularly limited, but for example, it may be 4:3, 3:2, 16:9, or 1:1.

[0038] In the quality prediction device 50, the computing unit 52, input / output device 54, and imaging device 58 may be integrated into a single unit, such as a smartphone with a built-in camera or a tablet computer with a built-in camera. In this case, the quality prediction device 50 may be portable by workers or other personnel. In one example, the quality prediction device 50 is configured by installing an application for predicting aggregate quality on a smartphone or tablet computer.

[0039] The imaging device 58 may be positioned in the manufacturing system 1 to image aggregates at specific locations. For example, the imaging device 58 may be positioned to image aggregates being transported by the transport device 8. The imaging device 58 may be mounted on a mobile device such as a drone, and the imaging device 58 may perform imaging of the aggregates by operating the mobile device by a user. The imaging device 58 may perform imaging of the aggregates in a room where the imaging environment, such as illumination, has been adjusted.

[0040] As shown in Figure 2, the arithmetic unit 52 includes a circuit 91. The circuit 91 includes a processor 92, a memory 93, a storage 94, and an input / output port 95. The storage 94 is composed of one or more non-volatile memory devices such as flash memory or a hard disk. The storage 94 stores at least a quality prediction program for causing the computer to execute the quality prediction process described later. The storage 94 stores quality prediction programs for configuring each of the functional blocks of the arithmetic unit 52 described later.

[0041] Memory 93 is composed of one or more volatile memory devices, such as random access memory. Memory 93 temporarily stores the quality prediction program loaded from storage 94. Processor 92 is composed of one or more computing devices, such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). Processor 92 executes the quality prediction program loaded into memory 93 to constitute each functional block of the computing unit 52. The calculation results by processor 92 are temporarily stored in memory 93. Input / output ports 95 perform information input and output with input / output devices 54 and imaging devices 58, etc., in response to requests from processor 92.

[0042] The circuit 91 is not necessarily limited to those whose functions are configured by a program. For example, the circuit 91 may have at least some functions configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates such circuits. 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.

[0043] Figure 3 shows an example of the functional components (referred to as "functional blocks" in this disclosure) of the computing device 52. The computing device 52 includes, as functional blocks, an imaging data acquisition unit 62, a preprocessing unit 64, an input data acquisition unit 66, a model construction unit 68, an imaging condition holding unit 69, a model holding unit 70, a prediction unit 72, and an output unit 74. The processing performed by these functional blocks corresponds to the processing performed by the computing device 52 (quality prediction device 50).

[0044] The imaging data acquisition unit 62 is configured to acquire imaging image data P1, which is color image data obtained by imaging the aggregate to be evaluated for quality. Hereinafter, the aggregate to be evaluated for quality will be referred to as "aggregate to be evaluated". The imaging data acquisition unit 62 acquires, for example, imaging image data P1 obtained by imaging by the imaging device 58 from the imaging device 58.

[0045] In color image data P1, each pixel has a defined pixel value for its red, green, and blue color components. If we denote the pixel values ​​for the red, green, and blue components of each pixel as "R," "G," and "B," respectively, the pixel value for each pixel is expressed as (R,G,B). Pixel values ​​are also called luminance values, density values, tonal values, gradation values, or color values. Pixel values ​​are typically expressed as numerical values ​​within the range of 0 to 255.

[0046] The preprocessing unit 64 generates preprocessed image data by applying predetermined preprocessing to the captured image data P1. The preprocessing unit 64 generates preprocessed image data (hereinafter referred to as "input image data P2") by applying a grayscale conversion process to the captured image data P1. The preprocessing unit 64 may perform other processing as part of the preprocessing at least before or after the grayscale conversion process, while maintaining the state in which the input image data P2 is grayscale image data. The preprocessing unit 64 may also perform only the grayscale conversion process as preprocessing.

[0047] The input image data P2 generated by the preprocessing unit 64 is image data based on the captured image data P1. If the ambient brightness of the object being captured differs, the image features may differ even if the object is the same aggregate. However, by performing a grayscale conversion as a preprocessing step, differences in image features due to brightness can be reduced.

[0048] Figure 4 schematically shows the calculation process in prediction using a prediction model for predicting quality. The preprocessing unit 64 generates input image data P2 by applying preprocessing (for example, only grayscale conversion) to the captured image data P1. In the preprocessing, the preprocessing unit 64 converts the combination of red, green, and blue pixel values ​​(R, G, B) for each pixel into a single pixel value. When the pixel value of each pixel in the input image data P2 (grayscale image data) is denoted as "Y", the preprocessing unit 64 may perform grayscale conversion for each pixel using the following equation (1). Y=0.299×R+0.587×G+0.114×B (1)

[0049] Returning to Figure 3, the input data acquisition unit 66 acquires input information including input image data P2. The input data acquisition unit 66 may acquire only input image data P2 as input information. The input data acquisition unit 66 may also acquire other information that may affect the aggregate particle size in addition to the input image data P2 as input information. This other information may include information that may affect the predicted value regarding the aggregate particle size when predicting quality based on image data.

[0050] Examples of other information mentioned above include the type of aggregate (e.g., rock type and origin) and the brightness of the aggregate (e.g., L). *Examples of other information include the color saturation of the aggregate, and the surface moisture content of the aggregate. Specific examples of the above-mentioned other information include environmental conditions at the time of imaging (e.g., illuminance, temperature, humidity, weather, and location), the distance between the camera or lens and the object being imaged, point cloud data, calculated values ​​based on point cloud data (e.g., angle, step, and cross-sectional area), dimensions of objects in the image, acoustic data (e.g., frequency, amplitude, wavelength, and sound pressure), and data on vibration, pressure, and friction or resistance during aggregate transport. Specific examples of the other information mentioned above include lens performance (e.g., focal length, 35mm equivalent focal length, F-number (aperture value), and sensor size), lens type (e.g., standard, wide-angle, and telephoto), lens model (e.g., manufacturer, product name, and year of manufacture), lens drive system (e.g., motor type: DC coreless motor, linear motor, and stepping motor), camera performance (e.g., shutter speed, ISO sensitivity, exposure compensation, and shooting distance), camera model (e.g., manufacturer, product name, and year of manufacture), image stabilization system (e.g., optical and electronic), and shutter control system (e.g., rolling shutter, global shutter, mechanical shutter, and electronic shutter). Other information may include the examples listed above, and may include two or more types of information.

[0051] The model building unit 68 constructs a model for predicting the physical properties of aggregates (hereinafter referred to as "prediction model M") by machine learning using training data that associates input information, including input image data P2, with the physical properties of aggregates. Prediction model M is a model constructed by machine learning to output the physical properties of aggregates according to the input information, including input image data P2. The physical properties of aggregates are physical properties related to the particle size distribution of the aggregates. Specific examples of physical properties related to aggregate particle size include the mass fraction remaining between each of the consecutive sieves (mass fraction for each particle size category), the mass fraction remaining on each sieve, the mass fraction passing through each sieve, and the coarse particle content of the aggregate, as specified in JIS A 1102:2014 "Test Method for Sieving Aggregates"; the fine particle content of the aggregate, as specified in JIS A 1102:2014 "Test Method for Fine Particle Content of Aggregates"; the excess particle content of the aggregate (amount of fine aggregate remaining on a 5mm sieve) and the under-particle content of the aggregate (amount of coarse aggregate passing through a 5mm sieve); and the actual packing ratio of the aggregate, as described in JIS A 1104:2019 "Test Method for Unit Volume Mass and Actual Packing Ratio of Aggregates".

[0052] The prediction unit 72 constructs a prediction model M by machine learning based on input information including input image data P2 and the correct information of the mass fractions remaining between each of the consecutive sieves associated with the input information.

[0053] Machine learning is a technology that allows machines to learn patterns and rules from dozens or more data points and use them to make judgments and predictions. For example, it refers to 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. A predictive model M can be realized, for example, using a neural network, which is an information processing model that mimics the structure of the human brain. The specific machine learning algorithm used when constructing a predictive model M is not particularly limited. A neural network has an input layer, one or more hidden layers, and an output layer. By including one or more hidden layers, a more complex predictive model M can be constructed, and the prediction accuracy can be improved.

[0054] The model building unit 68 may construct a prediction model M by machine learning using a convolutional neural network (CNN) in at least part of the calculation process from input information to outputting physical property values. By using a convolutional neural network, it becomes possible to more favorably capture the features of the image. The model building unit 68 may also construct a prediction model M by machine learning based on input information including input image data P2 and correct information on the mass fractions remaining between each of the consecutive sieves. The prediction model M constructed by the model building unit 68 may output a value indicating the mass fractions remaining between each of the consecutive sieves, depending on the input information including the input image data P2 (for example, depending on the input image data P2). That is, the prediction model M may output multiple values, each indicating a mass fraction for multiple particle size classifications.

[0055] The model building unit 68 may autonomously construct a predictive model M by performing machine learning using data provided as input to machine learning and ground truth information from the output of machine learning (such as the ground truth values ​​of the mass fractions remaining between each of the consecutive sieves). The input to machine learning is various datasets of input information, including input image data P2. The output of machine learning is data (numerical values) showing physical properties such as the mass fractions remaining between each of the consecutive sieves. The model building unit 68 iteratively trains a model that outputs physical properties related to the particle size of aggregate using multiple combinations of the input data datasets and the ground truth values ​​of the physical properties.

[0056] The stage in which the predictive model M is autonomously constructed corresponds to the training phase. The training phase is also called the learning phase. The same (corresponding) information and aggregates are used in the training phase and the evaluation phase. Hereafter, the information and aggregates used in the training phase and those used in the evaluation phase may be labeled "for training" or "for evaluation."

[0057] The imaging condition retention unit 69 retains information representing the conditions (imaging conditions) for imaging training aggregates during the training phase. The information representing the imaging conditions retained (stored) by the imaging condition retention unit 69 includes information for identifying the image resolution. Image resolution refers to the physical size (distance) per pixel. Physical size refers to the actual size of the object captured in the image. That is, when focusing on a single mass of object, the physical size of that object in terms of its dimensions (size) remains unchanged regardless of how the object is captured in the image.

[0058] When physical size is expressed in units of [mm], the unit of image resolution is [mm / pixel]. Image resolution is sometimes also called "pixel resolution," "resolution," "resolution," or "dot pitch." Even if the unit of image resolution is expressed in [dpi (dots per inch)] or [ppi (pixels per inch)], it can be converted to [mm / pixel (pixel)]. The lower the image resolution, the more detailed information can be seen in the image, but the larger the image file size (data size). On the other hand, the higher the image resolution, the smaller the image file size (data size), but the finer details on the image are removed.

[0059] Figure 5 shows a schematic diagram illustrating the difference in image resolution. In Figure 5, a portion of the image P1, which is imaged of aggregate, is magnified and schematic. The squares defined by the dashed grid lines represent each pixel of the image, and the pixels with diagonal lines indicate the presence of aggregate grains.

[0060] The "[d0,Np0]" near the image on the left side of Figure 5 indicates that d0 is the distance between the imaging device 58 and a reference position near the aggregate to be imaged (hereinafter referred to as the "imaging distance"), and Np0 is the (total) number of pixels in the image. The reference position may be the position on the surface of the tray on which the aggregate to be imaged is placed, and the imaging distance may be defined as the shortest distance between the reference position and the camera. Ds(r) represents the physical size described above. Assuming that Ds(r) corresponds to two pixels in the image, the image resolution is calculated as Ds(r) / 2 = 0.5·Ds(r).

[0061] The image in the upper right of Figure 5 shows an image obtained by changing the number of pixels to "Np1", which is greater than Np0, while maintaining the imaging distance at d0. When the number of pixels is Np1, the number of pixels is doubled in both the horizontal and vertical directions compared to when the number of pixels is Np0. In this case, Ds(r) corresponds to four pixels in the image. Therefore, the image resolution is calculated as "Ds(r) / 4 = 0.25·Ds(r)", resulting in a smaller image resolution. Conversely, if the number of pixels is made smaller than Np0, the image resolution will be greater than 0.5·Ds(r).

[0062] The image at the bottom right of Figure 5 shows an image obtained by changing the imaging distance to "d1", which is greater than d0, while maintaining the number of pixels at Np0. When the imaging distance is d1, Ds(r) corresponds to one pixel, compared to when the imaging distance is d0. In this case, the image resolution is calculated as "Ds(r) / 1 = Ds(r)", and the image resolution is increased. Conversely, if the imaging distance is made smaller than d0, the image resolution will be smaller than 0.5·Ds(r). Note that the image resolution can also be changed by the zoom function of the imaging device 58 instead of the imaging distance.

[0063] Returning to Figure 3, the model holding unit 70 holds the prediction model M constructed by the model construction unit 68. The prediction model M, being a trained model, may be transferable between computers. The prediction model M constructed in the quality prediction device 50 may also be used in other devices.

[0064] The prediction unit 72 predicts the quality of the aggregate to be evaluated during the evaluation phase. The prediction unit 72 predicts the quality of the aggregate to be evaluated based on the prediction model M constructed by the model construction unit 68 and the input information acquired by the input data acquisition unit 66. The prediction unit 72 may calculate a predicted quality value by performing a predetermined calculation on the physical property values ​​output from the prediction model M. Alternatively, the prediction unit 72 may acquire the physical property values ​​output from the prediction model M as they are and use them as the predicted quality value.

[0065] The prediction unit 72, for example, inputs evaluation input information into the prediction model M and then obtains a value from the prediction model M indicating the mass fraction remaining between each of the consecutive sieves. The prediction unit 72 may then calculate a predicted value for the coarseness of the aggregate to be evaluated by converting the value indicating the mass fraction remaining between each of the consecutive sieves into a coarseness ratio (see also Figure 4). The prediction unit 72 may correct the multiple mass fraction values ​​output from the prediction model M before converting to a coarseness ratio.

[0066] The output unit 74 outputs the quality prediction results from the prediction unit 72 to the monitor of the input / output device 54. The output unit 74 may also display the predicted value of the coarse particle percentage calculated by the prediction unit 72 on the monitor. In addition to the predicted value of the coarse particle percentage, the output unit 74 may also display the value obtained from the prediction model M regarding the mass fraction remaining between each of the consecutive sieves on the monitor. The output unit 74 may also display a graph of the particle size distribution on the monitor based on the predicted value of the mass fraction remaining between each of the consecutive sieves.

[0067] In modifications of this embodiment, image analysis may be used as the machine learning method. In one example of the image analysis method, an image of the aggregate is converted to grayscale (or a grayscale image of the aggregate is obtained), the area ratio of each particle is calculated, and then the area ratio is multiplied by a coefficient to convert it to a mass ratio, which is then converted to a mass ratio for each particle size category. Here, the coefficient multiplied by the area ratio is calculated from machine learning using images of the aggregate and measured mass fraction ground truth information as training data.

[0068] <Predictive Model M> Next, we will further explain the prediction model M used to predict the coarseness ratio. In the example shown in Figure 4, the input to the prediction model M is only the input image data P2, and the output to the prediction model M is the mass fraction remaining between each consecutive sieve. Unlike the example shown in Figure 4, in addition to the input image data P2, other features that may affect the aggregate quality may be used as input to the prediction model M.

[0069] In the training phase for constructing the prediction model M, the same type of aggregate that may be evaluated is used as the training aggregate. When constructing the prediction model M, training data for machine learning (hereinafter referred to as "training data TD") is prepared. Training data TD includes training input information and ground truth information for physical properties associated with the training input information. Training data TD consists of multiple datasets, each of which includes, for example, training input image data P2 and ground truth information for the mass fractions remaining between each consecutive sieve associated with that training input information. The training input image data P2 differs among the multiple datasets.

[0070] When preparing various training input image data P2, for example, various aggregates are prepared in which the mass fraction remaining between each of the consecutive sieves is known. The above-mentioned various aggregates do not mean different types of aggregates, but rather aggregates of the same type but with different particle sizes. In one example, the various training aggregates are individually imaged by the imaging device 58 while stacked on a flat tray. Stacked state (overlapping state) refers to a state in which, when viewed from vertically above, some of the particles in the aggregate overlap other particles.

[0071] The imaging device 58 images the area including the training materials on the tray from a predetermined distance away from the tray (for example, a predetermined position above it). The imaging by the imaging device 58 may be performed indoors or outdoors. The amount of training materials placed on the tray at the time of imaging is not particularly limited, but may be around 0.1 kg to 10 kg. At least a portion of the training materials being transported on the belt conveyor included in the transport device 8 may be used as training materials. In this case, the imaging device 58 may image the training materials being transported on the belt conveyor of the transport device 8. The imaging device 58 may also be positioned so that the amount of transporting training materials included in the imaging field of the imaging device 58 is around 0.1 kg to 10 kg.

[0072] In one example, various samples are prepared as training aggregates under the conditions shown in Table 1 below, and training image data P1 is obtained by imaging the prepared samples. The mass fraction and coarseness ratio that remain between each consecutive sieve for the various samples prepared under the conditions shown in Table 1 are known, and at least some of these known values ​​are used as ground truth information in machine learning. [Table 1]

[0073] The particle size ratio (FM) can be determined, for example, by the following formula (2) specified in JIS A 1102:2014 "Test Method for Sieving Aggregates". Note that in "Standard Specifications for Building Construction and Commentary JASS5 Reinforced Concrete Construction", nine sieves are used, excluding the 80 mm sieve in formula (2). The particle size ratio predicted in this disclosure may be determined by either method.

[0074]

number

[0075] Table 1 indicates that aggregates with 10 different particle sizes (coarseness ratios) are prepared. Each aggregate size is sieved sequentially using sieves with increasing mesh size, and classified into categories I to IV. Category I represents the particle size classification for particles that are 10 mm or larger but smaller than 20 mm. Particles that make up Category I are those that remain within a 10 mm sieve but pass through a 20 mm sieve. Category II represents the particle size classification for particles that are 5 mm or larger but smaller than 10 mm. Particles that make up Category II are those that remain within a 5 mm sieve but pass through a 10 mm sieve. Category III represents the particle size classification for particles that are 2.5 mm or larger but smaller than 5 mm. Particles that make up Category III are those that remain within a 2.5 mm sieve but pass through a 5 mm sieve. Category IV represents the particle size category for particles that are 1.2 mm or larger but smaller than 2.5 mm. Particles that make up Category IV are those that remain within a 1.2 mm sieve but pass through a 2.5 mm sieve. The mass fractions remaining in each successive sieve, using Level 1 in Table 1 as an example, mean that 85 wt% of the particles belonged to Category I, 15 wt% belonged to Category II, 0 wt% belonged to Category III, and 0 wt% belonged to Category IV.

[0076] The image data T1 to T10 shown in Figure 6 are examples of image data obtained by imaging training aggregate prepared according to the conditions of each of the levels 1 to 10 in Table 1. The image data illustrated in Figure 6 was obtained by imaging with the training aggregate placed on a sheet that simulates the transport of aggregate on a belt conveyor belt. Instead of such a sheet, the training aggregate may be placed on a flat tray, and image data relating to the training aggregate prepared according to the conditions of each of the levels 1 to 10 in Table 1 may be obtained. Level k (where k is an integer of 1 or more) corresponds to the image data Tk. For example, image data T1 is obtained by imaging training aggregate prepared according to level 1. The training aggregate according to each level k contains grains of multiple particle size classifications.

[0077] For each level k, multiple image data Tk may be prepared, each with a different state (more specifically, deposition state) of the training aggregate. In one example, one image data Tk is obtained by taking the first image with the training aggregate placed on a tray. Then, the training aggregate is temporarily removed from the tray, and the same training aggregate (same sample) is placed back on the tray so that the deposition state of the numerous particles is different from that of the first image. After that, a second image is taken. By repeating the above imaging process, multiple image data Tk may be prepared for each level k.

[0078] In addition to the conditions shown in Table 1, various samples may be prepared as training aggregates and training imaging image data P1 (image data Tk) may be prepared under the conditions shown in Tables 2 and 3 below.

[0079] [Table 2]

[0080] [Table 3]

[0081] In one example, one dataset in the training data TD is composed of one image data Tk and the ground truth information of the mass fraction at level k associated with that image data Tk. Multiple datasets may be prepared as the training data TD, each consisting of a combination of image data Tk and ground truth information of the mass fraction. Then, by performing machine learning using such training data TD, a prediction model M may be constructed that outputs predicted values ​​of four mass fractions for four granularity divisions. The mass fractions remaining between each consecutive sieve at each level in the training data TD, the number of levels (value of k), and the method of setting the granularity divisions are not limited to the examples shown in Tables 1 to 3, and may be set arbitrarily.

[0082] An example of a predictive model M will be explained using simplified mathematical formulas for easier understanding. The predictive model M constructed by the model building unit 68 can be expressed simply, for example, as shown in equations (3) and (4) below.

number

number

[0083] In equation (4), Y represents the output value of the physical property, for example, the output value of the mass fraction remaining between each consecutive sieve. In this case, equations (3) and (4) are calculated for each particle size category. N is an integer greater than or equal to 2 and represents the number of data sets in the input information. xl represents the various input values ​​included in the input information, and the input information includes at least input image data P2 based on the captured image data P1.

[0084] wl 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, absolute value, step function, sigmoid function, hardsigmoid function, logsigmoid function, softmax function, logsoftmax function, softmin function, softplus function, softsine function, tanh function, tanhShrink function, hardtanh function, tanhexp function, ReLU function, ReLU6 function, Leaky-ReLU function, PReLU function, ELU function, SELU function, CELU function, Switch function, Mish function, ACON function, etc.

[0085] The model building unit 68 may repeatedly evaluate the error (loss) between Y (predicted value) obtained by equation (4) and the correct information of physical properties using the training data TD, and determine the weights wl and bias term b in equation (3) so as to minimize the error. The model building unit 68 may use any type of loss function as a function to evaluate the error between Y (output value from an intermediate model in the intermediate stage of building the prediction model M) obtained by equation (4) and the correct information of physical properties. The loss function is a function that calculates the loss value based on the error between the predicted value and the correct value, and the weights wl are updated so as to minimize the loss value. Specific examples of loss functions include the Huber loss function, mean absolute error, mean squared error, and ε-tolerance loss function.

[0086] The model building unit 68 may repeatedly update the weights wl using the gradient method so that the loss value evaluated by the loss function is minimized. The model building unit 68 may use any type of update formula (weight update formula) when updating the weights wl. Specific examples of weight update formulas include Adam, AdamW, AdamBelief, Adamax, AdaBound, Adagrad, AMSGRAD, AMSBound, Lars, RMSprop, Adadelta, Sgd, SgdW, Momentum, and Nesterov. A weight update formula is also called an optimization algorithm or optimization method.

[0087] <Ratio of image resolution to minimum aggregate size> In quality prediction using the prediction model M by the quality prediction device 50, conditions regarding the image resolution when imaging the aggregate to obtain the imaged image data P1 are defined in both the training phase and the evaluation phase. Here, the image resolution is defined as "A", the minimum size of the aggregate as "B", and the image resolution relative to the minimum size of the aggregate as "A / B", and "A", "B", and "A / B" are defined as follows. • Image resolution A [mm / pixel]: The size (physical size) of one pixel in the captured image data P1. • Minimum dimension B [mm]: The minimum dimension of the aggregate included in the captured image data P1 (the measurement method will be described later). • Ratio A / B [1 / pixel]: A value obtained by dividing the image resolution A by the minimum dimension B.

[0088] Assuming that the minimum dimension B is constant, a small ratio A / B means a small image resolution A, and a large ratio A / B means a large image resolution A. The ratio A / B should be 0.20 or greater to avoid the possibility of being unable to construct a prediction model M due to an excessive amount of data, or to avoid situations where a huge amount of computation is required on the computer when constructing the prediction model M. To more reliably avoid problems caused by an excessive amount of data, the ratio A / B may be 0.25 or greater, 0.30 or greater, 0.35 or greater, or 0.40 or greater.

[0089] In quality control of aggregates, the ratio A / B is 2.93 or less, from the perspective of avoiding a decrease in prediction accuracy to the point where the prediction results using the prediction model M cannot be used. From the perspective of improving the prediction accuracy of the prediction model M, the ratio A / B may be 2.5 or less, 2.0 or less, or 1.6 or less. The ratio A / B is between 0.20 and 2.93 (0.20 to 2.93). The lower and upper limits of the ratio A / B may be arbitrarily selected and combined from the various values ​​described above. Image data P1 (for training and evaluation, respectively) may be acquired such that the ratio A / B matches between the training and evaluation phases.

[0090] The image resolution A is measured, for example, as follows: A reference member of known physical size (e.g., a scale with markings) is prepared. The reference member is then imaged by the imaging device 58 under the same conditions as when imaging the aggregate to be imaged. Subsequently, the physical size per pixel is calculated from the image in which the reference member is captured. For example, if the physical size indicated by the reference member is 10 mm, and it is assumed that this 10 mm portion is captured in 10 pixels in the image, the image resolution A is calculated to be 1 [mm / pixel].

[0091] The minimum dimension B may be 5.0 mm or less. The minimum dimension B does not mean the minimum physical size of the particles contained in the aggregate being imaged, and in this disclosure, the minimum dimension B is defined as follows. Minimum dimension B: Nominal size of sieves specified in Annex JA (Normative) of JIS A 5308:2024 "Ready-mixed concrete" for ready-mixed concrete aggregates, or JIS A 5005:2020 "Crushed stone and crushed sand for concrete". (=100,80,60,50,40,25,20,15,13,10,5,2.5,1.2,0.6,0.3,0.15mm) Of these, the minimum size is defined as one size smaller than the nominal size of the smallest sieve. However, if the nominal size of the smallest sieve is 0.15 mm, then 0.15 mm shall be considered the minimum size.

[0092] The minimum sieve nominal size mentioned above is determined by the type (classification) of aggregate. For example, for gravel with a maximum sieve nominal size of 20 mm, the minimum sieve nominal size is 2.5 mm, so according to the above definition, the minimum dimension B is 1.2 mm. For crushed stone 4005, the minimum sieve nominal size is 5 mm, so according to the above definition, the minimum dimension B is 2.5 mm. For sand and crushed sand, the minimum sieve nominal size is 0.15 mm, which is the smallest nominal size specified in JIS, so according to the above definition, the minimum dimension B is 0.15 mm. Note that when the minimum sieve nominal size is 0.3 mm or larger, the reason why the minimum dimension B is set to one size smaller is because the mass fraction (%) of what passes through the smallest nominal size sieve is specified to be 0-5%, meaning that there is a high possibility that some particles of one size smaller are also included. On the other hand, the reason why 0.15 mm is set as the minimum value for minimum dimension B is that there is no need to focus on particle sizes smaller than that in aggregate particle size control.

[0093] <Amount of training data> As described above, the training data TD used in machine learning to construct the predictive model M consists of multiple training datasets. Each of the multiple training datasets contains input information, including input image data P2, and the physical properties of the aggregate associated with that input information. Here, the number of multiple training datasets is denoted as "C", and the ratio of the number of multiple training datasets to the image resolution A is denoted as "C / A", with the ratio C / A defined as follows. • Ratio C / A: A value obtained by dividing the number C of multiple training datasets by the image resolution A.

[0094] To avoid the possibility of being unable to construct a predictive model M due to an excessive amount of data, or to avoid situations where a massive amount of computation is required on a computer when constructing a predictive model M, the ratio C / A may be 10,000 or less. To more reliably avoid problems caused by an excessive amount of data, the ratio C / A may be 9,500 or less, 9,000 or less, 8,500 or less, or 8,000 or less.

[0095] In managing the quality of aggregates, the ratio C / A is set to 600 or higher to avoid a decrease in prediction accuracy to the point where the prediction results using the prediction model M cannot be used. From the perspective of improving the prediction accuracy of the prediction model M, the ratio C / A may be 800 or higher, 1000 or higher, 1400 or higher, or 1800 or higher. The ratio C / A may also be between 600 and 10000 (600 to 10000). The lower and upper limits of the ratio C / A may be arbitrarily selected and combined from the various values ​​described above.

[0096] [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 a process of mixing materials including aggregates to produce ready-mix concrete. The quality prediction process is a process of predicting the quality of at least some of the aggregates used in 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 performed.

[0097] (manufacturing process) The manufacturing process includes, for example, a conveying process, a weighing process, a input process, a mixing process, a discharge process, and a loading process. In the conveying process, various concrete materials are conveyed to the storage bottle 12 by a conveying device 8, and each material is supplied to the storage bottle 12 individually. In the weighing process, each material is supplied individually from the storage bottle 12 to the weighing bottle 14, and each material is weighed in the weighing bottle 14. In the weighing process, when the measured amount of each material reaches a predetermined set amount, that material is discharged to the collection hopper 16.

[0098] In the input process, all types of materials are collected in the collection hopper 16, and then the materials in the collection hopper 16 are fed into the mixer 20. In the mixing process, multiple types of concrete materials are mixed in the mixer 20. In the discharge process, after the mixing of the concrete materials in the mixer 20 is completed, the ready-mixed concrete is discharged from the mixer 20 into the loading hopper 30. In the loading process, the ready-mixed concrete discharged into the loading hopper 30 is loaded onto the transport vehicle C.

[0099] (Quality prediction process) The quality prediction process (method for predicting aggregate quality) 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.

[0100] The model building process includes, for example, a preparation step and a building step. The model building unit 68 of the computing unit 52 may perform the preparation step and the building step. The preparation step is a step of preparing training data TD. The training data TD associates training input information, which includes input image data P2 based on image data P1 obtained by imaging training aggregate, with the physical properties of the training aggregate (e.g., coarseness ratio). The building step is a step of building a predictive model M by machine learning using the training data TD prepared in the preparation step.

[0101] The quality evaluation process includes, for example, an acquisition process, a prediction process, and an output process. The acquisition process is a process of acquiring input information for evaluation (input information including input image data P2) related to the aggregate to be evaluated. The acquisition process may include acquiring imaging image data P1, which is color image data obtained by imaging the aggregate to be evaluated, and generating input image data P2 by converting the imaging image data P1 to grayscale. Thus, the input image data P2 may be image data generated by converting the imaging image data P1 to grayscale. The input data acquisition unit 66 (or imaging data acquisition unit 62, preprocessing unit 64, and input data acquisition unit 66) may perform the acquisition process.

[0102] The prediction process is a process of predicting the quality of the aggregate to be evaluated based on the prediction model M constructed in the construction process and the evaluation input information acquired in the acquisition process. In the prediction process, for example, predicted values ​​of aggregate quality such as the coarseness ratio are calculated from the predicted values ​​of physical properties output from the prediction model M. The prediction unit 72 may perform the prediction process. The output process is a process of outputting the aggregate quality predicted by the prediction process to the input / output device 54. The output unit 74 may perform the output process.

[0103] The following describes an example of the model construction process and an example of the quality evaluation process, referring to Figures 7 and 8. The explanation will use the case where the aggregate quality is predicted based on the aggregate's coarseness ratio. Furthermore, the example will use the case where the prediction model M outputs predicted mass fraction values ​​for each of the four particle size classifications as predicted aggregate physical properties.

[0104] In the model building process, step S11 is performed first. In step S11, for example, a worker prepares training data TD for machine learning and inputs it to the computing device 52 via the input / output device 54. In one example, a worker prepares various aggregates in which the mass fraction remaining between each of the consecutive sieves is known. A worker may prepare various aggregates having particle sizes of levels 1 to 30 as shown in Tables 1 to 3 above. A worker may prepare M images (M is an integer of 2 or more) of training image data P1 in which at least one of the particle size levels and deposition states is different from each other. M may be 10 to 100,000 or 100 to 10,000. The data may be input to the computing device 52 with the known mass fraction for each particle size category associated as correct information for physical properties. The correct information on the mass fractions remaining between each of the consecutive sieves may be obtained by image analysis instead of by actual measurement through sieving tests. Specifically, the correct information in the training data TD may be obtained by converting images of the aggregate to grayscale (or obtaining grayscale images of the aggregate), calculating the area ratio for each particle size, and then multiplying that area ratio by a coefficient to convert it to a weight ratio. As a coefficient to be multiplied by the area ratio, for example, images of the aggregate and the measured correct information on the mass fraction may be used as training data. This simplifies the process of obtaining the correct information on the mass fractions remaining between each of the consecutive sieves.

[0105] In the imaging to obtain M images of training material P1, the training material may be imaged by the imaging device 58 under the same imaging conditions that define the image resolution A. For example, the number of pixels Np in the imaging device 58 may be set to the same value, and the imaging device 58 may be installed at the same position relative to the location where the training material is placed, and the M images of training material P1 may be obtained. The ratio A / B, which represents the ratio of the image resolution A to the minimum dimensions B of the material, may be within the range exemplified in the "ratio of image resolution to minimum dimensions of material" described above.

[0106] Next, step S12 is executed. In step S12, for example, the imaging conditions that define the image resolution A when acquiring training image data P1 by a worker are stored in the imaging condition holding unit 69. The worker may also input the imaging conditions that define the image resolution A (for example, the number of pixels Np and the imaging distance) to the arithmetic unit 52 by operating the input device of the input / output device 54.

[0107] Next, step S13 is executed. In step S13, for example, the model building unit 68 performs a grayscale conversion process on each of the M image data prepared in step S12, according to a predetermined calculation procedure. The model building unit 68 may also perform the grayscale conversion according to equation (1) above for each of the training image data P1 prepared in step S11. This yields M training input image data P2, which are grayscale images.

[0108] Each of the M training input image data P2 is associated with a known mass fraction for each particle size category as correct information for the physical properties. M corresponds to the number C of the multiple training datasets described above. The ratio C / A, which represents the ratio of the number of multiple training datasets to the image resolution A, may be within the range exemplified in the "amount of training data" described above.

[0109] After step S13 is executed, the model building unit 68 may perform a process to expand the training data TD (increase the number of datasets). The model building unit 68 creates (generates) expanded image data based on at least a portion of the M image data. In one example, the model building unit 68 performs a first transformation process to generate two or more first image data by randomly changing the brightness and contrast of each image data included in at least a portion of the M image data. Then, for each first image data generated by the first transformation process, the model building unit 68 generates two or more second image data by performing at least one of image inversion and rotation. The second image data corresponds to the expanded image data. When expanded image data is generated, the input image data P2 becomes that expanded image data. When the training data TD is expanded, the number of training datasets C becomes greater than M.

[0110] Next, steps S14 and S15 are executed. In step S14, for example, the model building unit 68 constructs a predictive model M by machine learning based on the training data TD prepared up to the execution of step S13. The model building unit 68 may also construct a predictive model M that outputs multiple mass fraction values ​​for multiple particle size classifications in response to the input of input image data P2, or the input of data including input image data P2. In step S15, for example, the model storage unit 70 stores the predictive model M constructed in step S14. With this, the series of processes performed in the model building process are completed.

[0111] <Quality Evaluation Process> Figure 8 is a flowchart illustrating an example of a series of processes performed in the quality evaluation process. This quality evaluation process is performed, for example, during a period that overlaps with at least a portion of the time during which the above-mentioned manufacturing process is being carried out in the manufacturing apparatus 10. In the quality evaluation process, a portion of the aggregate used in the production of ready-mixed concrete by the manufacturing apparatus 10 may be extracted, and the quality of that aggregate may be evaluated. The minimum dimension B of the aggregate to be evaluated is the same as the minimum dimension B of the training aggregate used in the model construction process.

[0112] In the quality evaluation process, step S21 is executed first. In step S21, for example, the imaging data acquisition unit 62 displays the image resolution A (image resolution A during the model construction stage) held in the imaging condition holding unit 69, or the number of pixels NP and imaging distance that define the image resolution A, on the monitor of the input / output device 54. In the quality evaluation process (evaluation phase), the ratio A / B, which represents the ratio of the image resolution A to the minimum dimensions B of the aggregate, is the same as the range exemplified in the "ratio of image resolution to the minimum dimensions of the aggregate" described above. The ratio A / B may be the same between the model construction process and the quality evaluation process. In addition, if it is guaranteed in the quality evaluation process that the ratio A / B is within the range described above (for example, that the ratio A / B is the same between the model construction process and the quality evaluation process), step S21 may be omitted.

[0113] Next, step S22 is performed. In step S22, for example, the imaging data acquisition unit 62 acquires evaluation image data P1 obtained by imaging the aggregate to be evaluated. The imaging data acquisition unit 62 may also acquire the image data P1 from the imaging device 58. The imaging of the aggregate to be evaluated by the imaging device 58 may be performed while some of the work is carried out by workers or the like. In this case, the workers or the like may check whether the image resolution A, or the number of pixels Np that defines the image resolution A and the imaging distance are the values ​​displayed on the monitor in step S21. Step S22 may be performed at a predetermined evaluation timing. The evaluation timing may be set to a certain time of day, or it may be at a timing instructed by the workers or the like. Among the conditions for imaging by the imaging device 58, conditions other than the image resolution A may be set to substantially match the conditions in the model construction process, within the settable range.

[0114] Next, step S23 is executed. In step S23, for example, the preprocessing unit 64 generates input image data P2 by performing a grayscale conversion process on the evaluation image data P1 acquired in step S22. The preprocessing unit 64 performs the grayscale conversion using the same calculation procedure (same image processing procedure) as the grayscale conversion process in the model construction step. The preprocessing unit 64 may also perform the grayscale conversion according to equation (1) described above for each pixel of the evaluation image data P1.

[0115] Next, steps S24 and S25 are executed. In step S24, for example, the prediction unit 72 inputs the evaluation input image data P2 generated in step S23, or input data including the evaluation input image data P2, into the prediction model M held by the model holding unit 70, and then obtains the predicted value of the mass fraction remaining between each of the consecutive sieves output from the prediction model M. In step S25, for example, the prediction unit 72 calculates the predicted value of the coarse grain ratio by converting the predicted value of the mass fraction remaining between each of the consecutive sieves obtained in step S24 into a coarse grain ratio.

[0116] Next, step S26 is executed. In step S26, for example, the output unit 74 outputs the predicted value of the coarse particle ratio calculated in step S25 to the monitor of the input / output device 54. In addition to the predicted value of the coarse particle ratio, the output unit 74 may also output the predicted value of the mass fraction remaining between each of the consecutive sieves acquired in step S24 to the monitor of the input / output device 54. By executing step S26, workers can confirm the predicted result of the coarse particle ratio. If the coarse particle ratio falls outside the control range, the workers may take measures to maintain the quality of the ready-mixed concrete (for example, adjusting the aggregate particle size, correcting (modifying) the mix, temporarily suspending production, investigating the cause, etc.).

[0117] With the above steps completed, the series of processes for predicting the coarseness ratio is finished. After step S26 is executed, the series of processes from steps S21 to S26 may be repeated. In the example above, some aggregate is extracted and its quality is evaluated (sampling inspection), but all aggregate used in the production of ready-mixed concrete may be inspected.

[0118] [Verification of prediction accuracy] Next, referring to Figures 9(a) and 9(b), we will explain the results of our verification of the prediction accuracy and training time (feasibility of constructing the prediction model M) of the prediction model M constructed using machine learning with a convolutional neural network. In this verification, we evaluated the prediction accuracy and training time while changing the image resolution A and the number of training datasets C. For the evaluation of prediction accuracy, we used a dataset in which the ground truth values ​​of quality are known, and compared the prediction results of the prediction model M with the ground truth values. For the construction of the prediction model M, we used a personal computer equipped with an NVIDIA RTX A4500 GPU.

[0119] <Verification using coarse aggregate> (Example 11) First, ten levels of training aggregate with different coarseness ratios were prepared according to the conditions in Table 1 above. For the training aggregate, coarse aggregate with a minimum dimension B of 1.2 mm was used. For each level, 4 kg of training aggregate was placed on a flat tray (a tray with a rim) so that the state of the aggregate on the tray was different for each level, and 10 imaging sessions were performed indoors to acquire 10 image data. From the 10 imaging sessions, 100 image data points were obtained according to the conditions in Table 1, and 100 data sets were obtained in which the image data points were correlated with the correct values ​​of the mass fractions remaining between each consecutive sieve. For imaging to obtain the image data, the imaging conditions were set so that the image resolution A was 3.52, and the ratio A / B, which represents the ratio of image resolution A to the minimum dimension B of the aggregate, was 2.93.

[0120] Next, ten levels of aggregate with different particle sizes (coarse aggregate with a minimum dimension B of 1.2 mm) were prepared according to the conditions in Table 2 above. Using the same method as for the preparation according to the conditions in Table 1, 100 image data points were obtained according to the conditions in Table 2, and 100 datasets were obtained in which the image data was associated with the correct mass fraction values ​​remaining between each consecutive sieve. Furthermore, ten levels of aggregate with different particle sizes (coarse aggregate with a minimum dimension B of 1.2 mm) were prepared according to the conditions in Table 3 above. Using the same method as for the preparation according to the conditions in Table 1, 100 image data points were obtained according to the conditions in Table 3, and 100 datasets were obtained in which the image data was associated with the correct mass fraction values ​​remaining between each consecutive sieve. For each of the 300 datasets according to the conditions in Tables 1, 2, and 3, the image data was converted to grayscale. In the grayscale conversion process, grayscale conversion was performed for each pixel using the equation (1) described above.

[0121] After performing the grayscale conversion process, new image data was created to expand the dataset used in the training phase. Specifically, the 300 image data points were expanded to 3600 image data points (input image data point P2) by randomly changing the brightness and contrast of the data points and inverting the images. The expansion process was carried out so that some of the expanded image data points contained the same data as the original image data points. As a result, 3600 datasets were obtained in which the input image data point P2 was associated with the correct mass fraction information, and 2520 of these datasets were used as the training data point TD. In this case, the ratio C / A, which represents the ratio of the number of training datasets C to the image resolution A, is calculated to be 716 (=2520 / 3.52). Machine learning was performed based on the training data point TD to construct a predictive model M.

[0122] In addition to the training dataset, two types of datasets (hereinafter referred to as "first test data" and "second test data") were prepared to compare and verify the predicted values ​​of the prediction model M with the actual values. Of the 3600 datasets mentioned above, 360 were used as the first test data. For the preparation of the second test data, 10 levels of aggregate with different coarseness ratios were prepared according to the conditions shown in Table 4 below.

[0123] [Table 4]

[0124] Under the conditions shown in Table 4, the combinations of mass fractions in multiple particle size categories are set to be different from the combinations under the conditions in Tables 1, 2, and 3, respectively. After preparing aggregate according to the conditions in Table 4, 100 image data (imaging data P1) according to the conditions in Table 4 were obtained using the same method as the preparation (imaging) according to the conditions in Table 1. By applying a grayscale conversion process to the imaging data P1 using the same method as the preparation according to the conditions in Table 1, 100 datasets were obtained as the second test data, each of which was associated with the image data (input image data P2) and the ground truth values ​​of the mass fractions remaining between each consecutive sieve. This second test data is based on aggregate with a different particle size level than the aggregate used to prepare the training data TD. In each dataset included in the first and second test data, in addition to the ground truth values ​​of the mass fractions remaining between each consecutive sieve, the ground truth values ​​of the coarseness percentage were also stored associated with the image data for comparative verification.

[0125] Using the first and second test data sets, we calculated the predicted value of the coarse grain percentage from the prediction results of the prediction model M, and then compared and verified the predicted value with the actual value. In the comparison and verification, the error between the predicted value and the actual value of the coarse grain percentage was evaluated using the following two indicators. • Accuracy rate of coarseness (%): The ratio of the number of correctly identified datasets to the total number of datasets, where a dataset with an error of ±0.10 or less is defined as correct. • Sum of mean and standard deviation: The arithmetic sum of the mean error and the standard deviation of the error. The second indicator, "the sum of the mean and standard deviation," means that the error is evaluated while also taking variability into account. In other words, even if the error itself (mean) is small, if the standard deviation is large, it means that the variability of the error is large, which is undesirable. Also, even if the standard deviation of the error is small, if the mean is large, it means that the error itself is large, which is also undesirable.

[0126] In addition to comparing and verifying the predicted values ​​with the actual values, we measured the training time required to build the predictive model M. Training time refers to the time from when the computer starts machine learning after the preparation of the training data TD is complete until the construction of the predictive model M is finished. Training time can be said to be the time required for the computer to build the predictive model M. In evaluating training time, we set a criterion that if the training time exceeds 10 days, it is "unable to build the predictive model M".

[0127] (Example 12) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 1.76 and the ratio A / B was 1.47 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 1432 (=2520 / 1.76).

[0128] (Example 13) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 1.17 and the ratio A / B was 0.98 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 2154 (=2520 / 1.17).

[0129] (Example 14) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.78 and the ratio A / B was 0.65 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 3231 (=2520 / 0.65).

[0130] (Example 15) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.59 and the ratio A / B was 0.49 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 4271 (=2520 / 0.59).

[0131] (Example 16) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.50 and the ratio A / B was 0.42 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 5040 (=2520 / 0.50).

[0132] (Example 17) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.44 and the ratio A / B was 0.37 when preparing all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 5727 (=2520 / 0.44).

[0133] (Comparative Example 11) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 7.03 and the ratio A / B was 5.86 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 358 (=2520 / 7.03).

[0134] (Comparative Example 12) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.22 and the ratio A / B was 0.18 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 11455 (=2520 / 0.22).

[0135] (Comparative Example 13) The verification process was the same as in Example 11, except that the imaging conditions were set so that the image resolution A was 0.15 and the ratio A / B was 0.13 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 16800 (=2520 / 0.15).

[0136] (Summary 1) The various conditions, measurement results of training time, and evaluation results such as accuracy rates for Examples 11-17 and Comparative Examples 11-13 are summarized in Tables 5 and 6 below. For clarity, the data are listed from top to bottom in descending order of image resolution A.

[0137] [Table 5]

[0138] [Table 6]

[0139] In Table 5, a "-" in the training time column indicates that the training time exceeded 10 days and the evaluation was that "the predictive model M could not be constructed." In Table 5, "d," "h," "m," and "s" represent days, hours, minutes, and seconds, respectively. In Table 6, "test1" means comparative verification using the first test data, and "test2" means comparative verification using the second test data. In Table 6, a "-" indicates that comparative verification could not be performed because the predictive model M could not be constructed.

[0140] From the results shown in Table 5, it can be seen that comparative examples 12 and 13 failed to construct the prediction model M, and that setting the ratio A / B to 0.20 or higher and the ratio A / C to 10000 or lower is useful from the perspective of successfully constructing the prediction model M. Furthermore, from the results shown in Table 6, it can be seen that setting the ratio A / B to 2.93 or lower and the ratio A / C to 600 or higher is useful from the perspective of appropriately predicting quality using the prediction model M. In addition, regardless of whether the first or second test data was used, the evaluation indicators in Examples 12 to 17 were higher than those in Example 11. Therefore, from the perspective of improving prediction accuracy, it can be seen that setting the ratio A / B to 2.5 or lower, 2.0 or lower, or 2.0 or lower is useful, and setting the ratio A / C to 800 or higher, 1000 or higher, 1200 or higher, or 1400 or higher is useful.

[0141] (Example 21) Similar to Example 11, 6000 datasets were prepared, each containing an input image data P2 and corresponding ground truth information for mass fraction. Of these, 4200 datasets were used as training data TD, and 600 datasets were used as the first test data. Also, similar to Example 11, a second test data set consisting of 100 datasets was prepared. However, in preparing all datasets—training data TD, the first test data, and the second test data—the imaging conditions were set so that the image resolution A was 1.76 and the ratio A / B was 1.47. After preparing these data, verification was performed in the same manner as in Example 11. In this case, the ratio C / A was calculated to be 2386 (=4200 / 1.76).

[0142] (Example 22) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 1.17 and the ratio A / B was 0.98 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 3590 (=4200 / 1.17).

[0143] (Example 23) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.78 and the ratio A / B was 0.65 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 5385 (=4200 / 0.78).

[0144] (Example 24) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.59 and the ratio A / B was 0.49 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 7119 (=4200 / 0.59).

[0145] (Example 25) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.50 and the ratio A / B was 0.42 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 8400 (=4200 / 0.50).

[0146] (Example 26) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.44 and the ratio A / B was 0.37 when preparing all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 9545 (=4200 / 0.44).

[0147] (Comparative Example 21) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 7.03 and the ratio A / B was 5.86 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 597 (=4200 / 7.03).

[0148] (Comparative Example 22) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.22 and the ratio A / B was 0.18 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 19091 (=4200 / 0.22).

[0149] (Comparative Example 23) The verification process was the same as in Example 21, except that the imaging conditions were set so that the image resolution A was 0.15 and the ratio A / B was 0.13 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 28000 (=4200 / 0.15).

[0150] (Summary 2) The various conditions, measurement results of learning time, and evaluation results such as accuracy rates for Examples 21-26 and Comparative Examples 21-23 are summarized in Tables 7 and 8 below. Similar to Tables 5 and 6, for clarity, the data are listed from top to bottom in order of increasing image resolution A, and the meanings of "-" and "d" are the same as in Tables 5 and 6.

[0151] [Table 7]

[0152] [Table 8]

[0153] From the results shown in Table 7, it can be seen that comparative examples 22 and 23 failed to construct the prediction model M, and that setting the ratio A / B to 0.20 or higher and the ratio A / C to 10000 or lower is useful from the perspective of successfully constructing the prediction model M. Furthermore, from the results shown in Table 8, it can be seen that setting the ratio A / B to 2.93 or lower and the ratio A / C to 600 or higher is useful from the perspective of appropriately predicting quality using the prediction model M.

[0154] (Example 31) Similar to Example 11, 13,200 datasets were prepared, each containing an input image data P2 associated with the correct mass fraction information. Of these, 9,240 datasets were used as training data TD, and 1,320 datasets were used as the first dataset. Also, similar to Example 11, a second test dataset consisting of 100 datasets was prepared. For all datasets—training data TD, the first test dataset, and the second test dataset—imaging conditions were set so that the image resolution A was 3.52 and the ratio A / B was 2.93. After preparing these datasets, verification was performed in the same manner as in Example 11. In this case, the ratio C / A was calculated to be 2625 (=9240 / 2.93).

[0155] (Example 32) The verification process was the same as in Example 31, except that the imaging conditions were set so that the image resolution A was 1.76 and the ratio A / B was 1.47 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 5250 (=9240 / 1.76).

[0156] (Example 33) The verification process was the same as in Example 31, except that the imaging conditions were set so that the image resolution A was 1.17 and the ratio A / B was 0.98 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 7897 (=9240 / 1.17).

[0157] (Comparative Example 31) The verification process was the same as in Example 31, except that the imaging conditions were set so that the image resolution A was 7.03 and the ratio A / B was 5.86 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 1314 (=9240 / 7.03).

[0158] (Comparative Example 32) The verification process was the same as in Example 31, except that the imaging conditions were set so that the image resolution A was 0.22 and the ratio A / B was 0.18 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 42000 (=9240 / 0.22).

[0159] (Comparative Example 33) The verification process was the same as in Example 31, except that the imaging conditions were set so that the image resolution A was 0.15 and the ratio A / B was 0.13 during the preparation of all datasets, including the training data TD, the first test data, and the second test data. In this case, the ratio C / A is calculated to be 61600 (=9240 / 0.15).

[0160] (Summary 3) Tables 9 and 10 below summarize the various conditions, measurement results of training time, and evaluation results such as accuracy in Examples 31-33 and Comparative Examples 31-33. Similar to Tables 5 and 6, for clarity, the data are listed from top to bottom in order of increasing image resolution A, and the meanings of "-" and "d" are the same as in Tables 5 and 6. Table 9 indicates that the training time for Example 33 could not be measured due to circumstances during the demonstration; however, the prediction model M was constructed in Example 33.

[0161] [Table 9]

[0162] [Table 10]

[0163] From the results shown in Table 9, it can be seen that comparative examples 32 and 33 failed to construct the prediction model M. From the perspective of successfully constructing the prediction model M, it is useful to set the ratio A / B to 0.20 or higher, and to set the ratio A / C to 10000 or lower. Furthermore, from the results shown in Table 10, it can be seen that from the perspective of appropriately predicting quality using the prediction model M, it is useful to set the ratio A / B to 2.93 or lower.

[0164] Figure 9(a) shows a graph of the accuracy evaluation results for each example and comparative example using the second test data. In the graph shown in Figure 9(a), the horizontal axis represents the image resolution A, and the vertical axis represents the accuracy with a tolerance of ±0.10. Figure 9(b) shows an enlarged graph of the portion indicated by "B" in Figure 9(a). From the graphs shown in Figures 9(a) and 9(b), it can be seen that by setting the image resolution A to 3.52 or less, i.e., the ratio A / B to 2.93 or less, a reasonably high accuracy rate (60% or more) can be obtained even when evaluating aggregates with different particle size levels from the training aggregate. Furthermore, it can be seen that by setting the image resolution A to 2.0 or less, i.e., the ratio A / B to 1.68 or less, a high accuracy rate (70% or more) can be obtained even when evaluating aggregates with different particle size levels from the training aggregate.

[0165] <Verification using fine aggregate> Next, we will explain the verification results when coarse aggregate was replaced with fine aggregate. In the verification using fine aggregate, the accuracy rate was evaluated with a tolerance of ±0.10 under the condition that the ratio A / B was between 0.20 and 2.93.

[0166] (Example 41) First, nine levels of training aggregate with different particle size ratios were prepared. Fine aggregate with a minimum dimension B of 0.15 mm was used as the training aggregate. For each level, 500 g of training aggregate was placed on a flat tray (plywood with a rim) so that the state of the aggregate on the tray was different for each level, and 10 imaging sessions were performed indoors to acquire 10 image data. From the 10 imaging sessions, 90 image data points were obtained, and 90 datasets were obtained in which the image data points were correlated with the correct values ​​of the mass fractions remaining between each consecutive sieve. For imaging to obtain the image data, the imaging conditions were set so that the image resolution A was 0.373, and the ratio A / B, which represents the ratio of image resolution A to the minimum aggregate dimension B, was 2.487.

[0167] Each image data in the 90 datasets described above was converted to grayscale in the same manner as in Example 11. The 90 grayscaled datasets were then expanded to 360 datasets in the same manner as in Example 11, and 288 of these datasets were used as training data TD. A predictive model M was constructed using machine learning based on the training data TD. Second test data was prepared using aggregate (fine aggregate) with a different particle size level than the aggregate (fine aggregate) used in the preparation of training data TD. In preparing the second test data, 30 datasets were prepared using three levels of fine aggregate.

[0168] Using the second test data, we calculated the predicted value of the coarseness percentage from the prediction results of the prediction model M, and then compared and verified the predicted value with the actual value. In the comparison and verification, the error between the predicted value and the actual value of the coarseness percentage was evaluated by calculating the accuracy rate.

[0169] (Example 42) The verification process was carried out in the same manner as in Example 41, except that the imaging conditions were set so that the image resolution A was 0.298 and the ratio A / B was 1.987 during the preparation of all datasets, including the training data TD and the second test data.

[0170] (Example 43) The verification process was the same as in Example 41, except that the imaging conditions were set so that the image resolution A was 0.249 and the ratio A / B was 1.660 during the preparation of all datasets, including the training data TD and the second test data.

[0171] (Summary 4) Table 11 below summarizes the various conditions and the evaluation results of the accuracy rate in Examples 41 to 43. From the results shown in Table 11, it can be seen that when predicting the quality of fine aggregate with a different particle size level than that of the training phase, a high accuracy rate (70% or more) can be obtained by setting the ratio A / B to 2.93 or less (or 2.5 or less).

[0172] [Table 11]

[0173] [Differentiation] The processing flows shown in Figures 7 and 8 are examples and can be modified as appropriate. In the above processing flows, one step and the next step may be executed in parallel, and several 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 flows, or in addition to the above processing flows, steps with content different from the example above may be executed. When augmented image data is generated and the training data TD is augmented, processing for augmentation (such as changing the contrast) may be performed before grayscale conversion.

[0174] In step S14, the prediction model M may be configured to output the coarseness ratio in response to input information including input image data P2. In this case, in the evaluation phase, instead of steps S24 and S25, the coarseness ratio output when evaluation input information including evaluation input image data P2 is input to the prediction model M is obtained as the prediction result. The prediction model M may be configured to output two or more physical properties related to the particle size of the aggregate. For example, the prediction model M may be configured to output the mass fraction and coarseness ratio remaining between each of the consecutive sieves in response to input information including input image data P2. In this case, the prediction unit 72 may calculate the coarseness ratio based on the predicted mass fraction value from the prediction model M and the predicted coarseness ratio output from the prediction model M as the quality prediction result. The input information, which is the data input to the prediction model M, may include, in addition to the input image data P2, at least one of any information that may correlate with the quality of the aggregate and any information that may influence the prediction of the quality of the aggregate from the image.

[0175] In addition to predicting aggregate quality using the quality prediction device 50, the manufacturing system 1 may periodically measure the physical properties of the aggregate (for example, several times a day). In this case, the prediction model M may be updated based on the measured physical properties of the aggregate and the input information (for example, input image data P2) at the time the measured values ​​were obtained. In the above prediction process, the physical properties of the aggregate may be predicted using the updated prediction model M. Even when the updated prediction model M is used, the process of predicting the physical properties of the aggregate for evaluation based on the prediction model M and the input information for evaluation remains unchanged.

[0176] The quality prediction device 50 may be used in locations other than the manufacturing system 1 (other than process inspections in the ready-mix concrete manufacturing process). The quality prediction device 50 may also be used in acceptance inspections when a plant that manufactures ready-mix concrete receives raw materials. For example, the quality prediction device 50 may be used in quality prediction (acceptance inspection) performed on a transport truck or transport ship.

[0177] The imaging device 58 may obtain color image data P1 by imaging, and then perform a process to convert the image data P1 to grayscale to generate grayscale image data. In this way, some of the functional blocks described above may be configured by the imaging device 58 instead of the arithmetic unit 52.

[0178] Alternatively, input image data P2 may be generated by applying different preprocessing to the captured image data P1, instead of or in addition to grayscale conversion. For example, input image data P2 may be generated by sequentially applying grayscale conversion and binarization to the captured image data P1. In the binarization process, any method may be used. All pixels may be binarized using a certain threshold, or binarization may be performed using a dynamically set threshold for each pixel, for example, as in the Otsu binarization. In the binarization process, binarization may be performed by detecting edges. After grayscale conversion or after binarization, a blackout process may be performed to darken the entire image.

[0179] It is not necessary to perform any preprocessing, including grayscale conversion, on the color image data P1. In this case, the color image data P1 may be used directly as the input image data P2. That is, the input image data P2 based on the image data P1 may be the color image data P1 itself.

[0180] The imaging device 58 may acquire grayscale image data P1 instead of color image data P1. The imaging device 58 may be a monochrome camera. In this case, the grayscale image data P1 may be used directly as input image data P2 (grayscale image data). If the imaging device 58 itself acquires grayscale image data regardless of whether the grayscale conversion process is performed, steps S13 and S23 are omitted.

[0181] In one of the various examples described above, at least some of the matters described in the other examples may be combined.

[0182] [Summary of this disclosure] The aggregate quality prediction method described above includes a construction step of constructing a prediction model (M) that outputs the physical properties of the aggregate in response to the input of the above input information by machine learning using training data (TD) which associates input information including input image data (P1, P2) based on image data (P1) obtained by imaging the aggregate, and the physical properties of the aggregate; an acquisition step of acquiring the above input information regarding the aggregate to be evaluated; and a prediction step of predicting the quality of the aggregate to be evaluated based on the prediction model (M) constructed in the construction step and the above input information acquired in the acquisition step. When the image resolution representing the size of one pixel in the image data (P1) is A [mm / pixel] and the minimum dimension of the aggregate included in the image data (P1) is B [mm], the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less.

[0183] If the ratio A / B is less than 0.20, it is conceivable that a predictive model (M) cannot be constructed using machine learning due to the large size of the image data, or that it requires an enormous amount of computation on the computer. Conversely, if the ratio A / B is greater than 2.93, it is conceivable that the predictive model (M) constructed using machine learning cannot adequately predict the quality of the aggregate. In the above quality prediction method, by setting the ratio A / B between 0.20 and 2.93, a predictive model (M) capable of predicting quality can be created more reliably and easily, and quality prediction can be performed using the predictive model (M). Therefore, it becomes possible to manage the quality of aggregate using a predictive model constructed using machine learning.

[0184] In the quality prediction method described above, the training data (TD) may consist of multiple training datasets, each of which contains the input information and the physical property values ​​associated with that input information. When the number of training datasets is C, the value obtained by dividing C by A (C / A) may be between 600 and 10000. In this case, a predictive model (M) capable of predicting quality can be created more reliably or easily while maintaining a prediction accuracy above a certain level.

[0185] In the quality prediction method described above, the input image data (P2) may be grayscale image data. In this case, the influence of brightness during imaging is reduced, and since the information contained in a single pixel is smaller than that in color image data, it also contributes to reducing the amount of training data (TD).

[0186] In the quality prediction method described above, B, which represents the minimum size of aggregate included in the image data (P1), may be 5.0 or less. In this case, the aggregate to be predicted includes coarse aggregate and fine aggregate, and this quality prediction method can be widely applied to aggregate quality control.

[0187] The quality prediction program described above is a program that allows a computer to execute the aggregate quality prediction method described above. Similar to the aggregate quality prediction method described above, this quality prediction program makes it possible to manage the quality of aggregates using a prediction model built with machine learning.

[0188] The ready-mix concrete manufacturing method described above includes a manufacturing process of mixing materials containing aggregate to produce ready-mix concrete, and a quality prediction process of predicting the quality of at least a portion of the aggregate used in the manufacturing process using the aggregate quality prediction method described above. Similar to the aggregate quality prediction method described above, this manufacturing method makes it possible to manage the quality of the aggregate using a predictive model constructed by machine learning.

[0189] The aggregate quality prediction device (50) described above comprises: a model construction unit (68) that constructs a prediction model (M) that outputs the physical properties in response to the input of the input information by machine learning using training data (TD) which associates input information including input image data (P1, P2) based on image data (P1) obtained by imaging aggregate, and physical properties of the aggregate; an input data acquisition unit (66) that acquires the input information regarding the aggregate to be evaluated; and a prediction unit (72) that predicts the quality of the aggregate to be evaluated based on the prediction model (M) constructed by the model construction unit (68) and the input information acquired by the input data acquisition unit (66). When the image resolution representing the size of one pixel in the image data (P1) is A [mm / pixel] and the minimum dimension of the aggregate included in the image data (P1) is B [mm], the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less. This quality prediction device (50), similar to the aggregate quality prediction method described above, enables the management of aggregate quality using a prediction model constructed by machine learning.

[0190] The ready-mix concrete manufacturing system (1) described above comprises a manufacturing device (10) that mixes materials including aggregate to produce ready-mix concrete, and the quality prediction device (50) described above. The quality prediction device (50) predicts the quality of at least a portion of the aggregate used in the production of ready-mix concrete by the manufacturing device (10). This manufacturing system enables the management of aggregate quality using a predictive model constructed by machine learning, similar to the aggregate quality prediction method described above. [Explanation of symbols]

[0191] 1...Manufacturing system, 10...Manufacturing equipment, 50...Quality prediction equipment, 62...Image data acquisition unit, 64...Preprocessing unit, 66...Input data acquisition unit, 68...Model construction unit, 72...Prediction unit, P1...Imageed image data, P2...Input image data, M...Prediction model.

Claims

1. A construction step involves constructing a predictive model that outputs the physical properties in response to input information, using machine learning that associates input image data, which includes input image data based on image data obtained by imaging aggregates, with the physical properties of the aggregates. An acquisition step for acquiring the aforementioned input information regarding the aggregate to be evaluated, The process includes a prediction step that predicts the quality of the aggregate to be evaluated based on the prediction model constructed in the construction step and the input information acquired in the acquisition step, When the image resolution representing the size of one pixel in the aforementioned captured image data is A [mm / pixel], and the minimum dimension of the aggregate included in the captured image data is B [mm], the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less. Method for predicting aggregate quality.

2. The aforementioned training data consists of multiple training datasets, each of which includes the input information and the physical property values ​​associated with that input information. When the number of the aforementioned training datasets is C, the value obtained by dividing C by A (C / A) is between 600 and 10000. The method for predicting the quality of aggregate according to claim 1.

3. The input image data is grayscale image data. A method for predicting the quality of aggregate according to claim 1 or 2.

4. B, which represents the minimum dimension of the aggregate included in the aforementioned image data, is 5.0 or less. A method for predicting the quality of aggregate according to claim 1 or 2.

5. A quality prediction program for causing a computer to execute the aggregate quality prediction method described in claim 1 or 2.

6. The manufacturing process involves mixing materials containing aggregates to produce ready-mixed concrete, A quality prediction step includes predicting the quality of at least a portion of the aggregate used in the manufacturing process using the aggregate quality prediction method described in claim 1 or 2. A method for manufacturing ready-mixed concrete.

7. A model building unit constructs a predictive model that outputs the physical properties in response to input information, using machine learning that associates input image data, which includes input image data based on image data obtained by imaging aggregates, with the physical properties of the aggregates. An input data acquisition unit that acquires the aforementioned input information regarding the aggregate to be evaluated, The system comprises a prediction unit that predicts the quality of the aggregate to be evaluated based on the prediction model constructed by the model construction unit and the input information acquired by the input data acquisition unit, When the image resolution representing the size of one pixel in the aforementioned captured image data is A [mm / pixel], and the minimum dimension of the aggregate included in the captured image data is B [mm], the value obtained by dividing A by B (A / B [1 / pixel]) is 0.20 or more and 2.93 or less. Aggregate quality prediction device.

8. A manufacturing apparatus that mixes materials including aggregates to produce ready-mixed concrete, The quality prediction device according to claim 7 comprises, The quality prediction device predicts the quality of at least a portion of the aggregate used in the production of ready-mixed concrete by the manufacturing device. A ready-mix concrete manufacturing system.

Images