Aggregate quality estimation method, quality estimation program, aggregate quality estimation device, ready-mixed concrete manufacturing method, and ready-mixed concrete manufacturing system

Abstract

To improve prediction accuracy when predicting quality by using a prediction model constructed by machine learning.SOLUTION: A quality prediction method of an aggregate includes a first acquisition step of acquiring color captured image data obtained by imaging an aggregate to be evaluated, a pre-processing step of generating input image data by gray-scaling the captured image data, a second acquisition step of acquiring input information including the input image data, a prediction step of acquiring a physical property value corresponding to input information acquired in the second acquisition step using a prediction model constructed in advance so as to output the physical property value of the aggregate according to an input of the input information, and a construction step of constructing the prediction model by machine learning based on training data. IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx], where IA [lx] is the illuminance among the environmental conditions when imaging the aggregate for training, and IB [lx] is the illuminance among the environmental conditions when imaging the aggregate to be evaluated.SELECTED DRAWING: Figure 12

Description

[Technical Field]

[0001] The present disclosure relates to an aggregate quality prediction method, a quality prediction program, an aggregate quality prediction device, a ready-mix concrete manufacturing method, and a ready-mix concrete manufacturing system. [Background technology]

[0002] Patent Document 1 discloses an aggregate quality determination system for determining the quality of aggregate, and an aggregate quality determination method. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-137651 Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure provides an aggregate quality prediction method, a quality prediction program, an aggregate quality prediction device, a ready-mix concrete manufacturing method, and a ready-mix concrete manufacturing system that are useful for improving prediction accuracy when predicting quality using a prediction model constructed by machine learning. [Means for solving the problem]

[0005] [1] A method for predicting the quality of aggregate, comprising: a first acquisition step of acquiring color image data obtained by imaging an aggregate to be evaluated; a preprocessing step of generating input image data by grayscaling the image data; a second acquisition step of acquiring input information including the input image data; a prediction step of acquiring physical property values ​​corresponding to the input information acquired in the second acquisition step using a prediction model previously constructed to output physical property values ​​of the aggregate in response to input of the input information; and a construction step of constructing the prediction model by machine learning based on training data including training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information, wherein when IA [lx] is an illuminance among the environmental conditions when imaging the training aggregate to prepare the training input information, and IB [lx] is an illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the image data, IB [lx] is not less than 0.2 times IA [lx] and not more than 5 times IA [lx].

[0006] [2] The method for predicting the quality of aggregate described in [1] above, wherein the absolute value of the difference between IA [lx], which represents the illuminance among the environmental conditions when the training aggregate is imaged to prepare the training input information, and IB [lx], which represents the illuminance among the environmental conditions when the aggregate to be evaluated is imaged to obtain the image data, is 2500 lx or less.

[0007] [3] A method for predicting the quality of aggregate described in [1] or [2] above, wherein IB [lx], which represents the illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the image data, is 2000 lx or more.

[0008] [4] In the pre-processing step, when the pixel values ​​of the red, green, and blue color components of each pixel in the color image data are represented as R, G, and B, and the pixel value of each pixel in the grayscale image data is represented as Y, grayscale conversion is performed using the following equation (1): Y=0.299×R+0.587×G+0.114×B (1) The method for predicting aggregate quality according to any one of the above [1] to [3].

[0009] [5] A method for predicting the quality of aggregates described in any one of [1] to [4] above, further comprising a preparation step of preparing the training data, wherein the preparation step includes generating extended image data by performing a conversion process and a grayscale process in any order on training image data corresponding to the captured image data, wherein the conversion process includes randomly changing the brightness and contrast of the image, and wherein the input image data in the training input information includes the extended image data.

[0010] [6] The method for predicting quality of aggregates described in [5] above, wherein the transformation process further includes at least one of flipping the image and rotating the image.

[0011] [7] The method for predicting aggregate quality described in any one of [1] to [6] above, wherein the construction process includes preparing, separately from the training data, model evaluation data including input information for model evaluation corresponding to the input information and correct information for the physical property values ​​associated with the input information for model evaluation; continuously updating the intermediate model using the training data so as to reduce the error with the correct information for the physical property values; and determining, using the model evaluation data, one of the intermediate models generated during the update as the prediction model.

[0012] [8] A quality prediction program for causing a computer to execute the aggregate quality prediction method according to any one of [1] to [7] above.

[0013] [9] An aggregate quality prediction device comprising: a first acquisition unit that acquires color image data obtained by imaging an aggregate to be evaluated; a pre-processing unit that generates input image data by grayscaling the image data; a second acquisition unit that acquires input information including the input image data; a prediction unit that acquires physical property values ​​corresponding to the input information acquired by the second acquisition unit using a prediction model previously constructed to output physical property values ​​of the aggregate in response to input of the input information; and a construction unit that constructs the prediction model by machine learning based on training data including training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information, wherein when IA [lx] is an illuminance among the environmental conditions when imaging a training aggregate to prepare the training input information, and IB [lx] is an illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the image data, IB [lx] is not less than 0.2 times IA [lx] and not more than 5 times IA [lx].

[0014]

[10] A method for producing ready-mixed concrete by mixing materials including aggregate, and a quality prediction process for predicting the quality of at least a portion of the aggregate used in the production process, wherein the quality prediction process includes a first acquisition process for acquiring color image data obtained by imaging the aggregate to be evaluated, a pre-processing process for generating input image data by grayscaling the image data, a second acquisition process for acquiring input information including the input image data, and a prediction model for outputting physical property values ​​of the aggregate in response to input of the input information, based on the input information acquired in the second acquisition process. and a construction process for constructing the prediction model by machine learning based on training data including training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information, wherein when illuminance among the environmental conditions when imaging a training aggregate to prepare the training input information is IA [lx] and illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the image data is IB [lx], IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx].

[0015]

[11] A manufacturing apparatus for producing ready-mixed concrete by mixing materials including aggregate, and a quality prediction device for predicting the quality of at least a portion of the aggregate used by the manufacturing apparatus, wherein the quality prediction device includes a first acquisition unit that acquires color image data obtained by imaging the aggregate to be evaluated, a pre-processing unit that generates input image data by grayscaling the image data, a second acquisition unit that acquires input information including the input image data, and a prediction model that is constructed in advance to output physical property values ​​of the aggregate in response to input of the input information, and a quality prediction device that predicts the quality of at least a portion of the aggregate used by the manufacturing apparatus. a prediction unit that acquires the physical property values; and a construction unit that constructs the prediction model by machine learning based on training data including training input information corresponding to the input information and correct answer information for the physical property values ​​associated with the training input information, wherein when the illuminance among the environmental conditions when imaging training aggregate to prepare the training input information is IA [lx] and the illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the captured image data is IB [lx], IB [lx] is not less than 0.2 times IA [lx] and not more than 5 times IA [lx]. [Effects of the Invention]

[0016] According to the present disclosure, there are provided an aggregate quality prediction method, a quality prediction program, an aggregate quality prediction device, a ready-mix concrete manufacturing method, and a ready-mix concrete manufacturing system, which are useful for improving prediction accuracy when predicting quality using a prediction model constructed by machine learning. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a schematic diagram showing an example of a ready-mix concrete manufacturing system. [Figure 2] FIG. 2 is a block diagram showing an example of the hardware configuration of a calculation device included in the quality prediction device. [Figure 3] FIG. 3 is a block diagram illustrating an example of the functional configuration of the arithmetic device. [Figure 4]FIG. 4 is a diagram for explaining problems that arise when the imaging environment is different. [Figure 5] FIG. 5 is a diagram illustrating an example of a calculation process when predicting quality using a prediction model. [Figure 6] FIG. 6 is a diagram showing an example of captured image data for training. [Figure 7] FIG. 7 is a flowchart illustrating a series of processes in the training phase. [Figure 8] FIG. 8 is a diagram illustrating how extended image data is created. [Figure 9] FIG. 9 is a flowchart illustrating a series of processes executed during model construction. [Figure 10] FIG. 10 is a diagram illustrating the calculation process when constructing a model. [Figure 11] FIG. 11 is a flowchart illustrating a series of processes in the evaluation phase. [Figure 12] 12(a) and 12(b) are graphs illustrating the results of comparing predicted values ​​with correct values ​​in a test dataset. DETAILED DESCRIPTION OF THE INVENTION

[0018] An embodiment will be described below with reference to the drawings. In the description, the same elements or elements having the same functions are designated by the same reference numerals, and redundant description will be omitted.

[0019] [Ready-mix concrete manufacturing system] Fig. 1 schematically shows a ready-mixed concrete manufacturing system according to one embodiment. The manufacturing system 1 shown in Fig. 1 is a system for manufacturing ready-mixed concrete. The manufacturing system 1 is installed, for example, in a factory for manufacturing ready-mixed concrete. The manufacturing system 1 manufactures (produces) ready-mixed concrete by mixing concrete materials.

[0020] The concrete materials used in the manufacturing system 1 include cement, admixtures, coarse aggregate, fine aggregate, water, and admixtures. Examples of coarse aggregate include gravel, crushed stone, slag coarse aggregate, lightweight coarse aggregate, recycled coarse aggregate, recovered aggregate, and coarse aggregates made from a mixture of these. Examples of gravel include mountain gravel, land gravel, river gravel, and sea gravel. Examples of slag coarse aggregate include blast furnace slag aggregate, ferronickel slag aggregate, electric arc furnace oxidizing slag aggregate, and coal gasification slag aggregate. Examples of lightweight coarse aggregate include natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate. Examples of coarse aggregate include crushed rock or crushed limestone.

[0021] Examples of fine aggregates include sand, crushed sand, slag fine aggregate, lightweight fine aggregate, recycled fine aggregate, recovered aggregate, and fine aggregates made from mixtures of these. Sand includes mountain sand, land sand, river sand, and sea sand. Slag fine aggregates include blast furnace slag aggregate, ferronickel slag aggregate, copper slag aggregate, electric furnace oxidizing slag aggregate, and coal gasification slag aggregate. Lightweight fine aggregates include natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate.

[0022] Examples of rock types for crushed stone and crushed sand include igneous rocks, sedimentary rocks, metamorphic rocks, quartzite, limestone, domalite, and peridotite. Igneous rocks include granite, diorite, gabbro, porphyrite, diabase, rhyolite, andesite, basalt, and serpentinite. Sedimentary rocks include conglomerate, sandstone, shale, slate, and tuff. Metamorphic rocks include gneiss and schist.

[0023] In addition to the function of producing ready-mixed concrete, the manufacturing system 1 also has the function of predicting the quality of aggregate contained in concrete materials. In this disclosure, coarse aggregate and fine aggregate may be collectively referred to as "aggregate." In this case, "aggregate" refers to coarse aggregate, fine aggregate, or both coarse aggregate and fine aggregate.

[0024] The manufacturing system 1 loads the manufactured ready-mixed concrete onto a transport vehicle C. After the ready-mixed concrete has been loaded onto the transport vehicle C, the transport vehicle C transports the ready-mixed concrete to the site where the ready-mixed concrete will be used (for example, a construction site). Examples of the transport vehicle C include an agitator vehicle (mixer vehicle) or a dump truck. The manufacturing system 1 may manufacture ready-mixed concrete from concrete materials so as to satisfy a target quality (required quality) set for each site. For example, an operator of the manufacturing system 1 determines the mix of concrete materials so as to satisfy the control quality at the time of shipping based on the target quality at the site, and inputs operating instructions to the manufacturing system 1.

[0025] 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.

[0026] The material storage yard 2 is a place where concrete materials are stored. The material storage yard 2 includes a plurality of silos 4. The plurality of silos 4 are containers that store at least a portion of the concrete materials by material type. The plurality of silos 4 include a silo 4 that stores coarse aggregate, a silo 4 that stores fine aggregate, and a silo 4 that stores cement.

[0027] The transporting device 8 is a device that transports the concrete materials stored in the multiple silos 4 to the manufacturing device 10. The transporting device 8 includes, for example, a belt conveyor that transports the concrete materials. The transporting device 8 may transport the concrete materials by type at different times. In one example, based on an operation instruction from a control device included in the manufacturing system 1, a specific material from among the various concrete materials is transferred to the transporting device 8 and transported to the manufacturing device 10.

[0028] The manufacturing apparatus 10 is an apparatus that mixes concrete materials including aggregate to manufacture 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 bottle 12, a measuring bottle 14, a collecting hopper 16, a mixer 20, and a loading hopper 30.

[0029] Storage bottles 12 temporarily store various types of concrete materials. Various types of concrete materials are transported (conveyed) to storage bottles 12 from material storage area 2 by transport device 8. Storage bottles 12 are configured to store various types of concrete materials individually. Hereinafter, "concrete materials" may be simply referred to as "materials." Various materials stored in storage bottles 12 are supplied to measuring bottles 14 as needed.

[0030] The measuring bottle 14 is disposed below the storage bottle 12. The measuring bottle 14 operates based on operational instructions from the control device of the manufacturing system 1, and individually measures various materials. When the measuring bottle 14 detects the target amount of material instructed by the control device, it supplies the material to the collecting hopper 16. When water is supplied to the measuring bottle 14, an admixture may be mixed into the water. The collecting hopper 16 is disposed below the measuring bottle 14. The collecting hopper 16 collects the various materials discharged from the measuring bottle 14 and supplies the collected materials to the mixer 20.

[0031] The mixer 20 is disposed below the collecting hopper 16. The mixer 20 is a device that mixes concrete materials. The mixer 20 produces ready-mixed concrete by mixing (kneading) aggregate, cement, water, admixtures, etc. The ready-mixed concrete is discharged from the bottom of the mixer 20 into the loading hopper 30. The loading hopper 30 is disposed below the mixer 20 and temporarily stores the ready-mixed concrete. The loading hopper 30 supplies the temporarily stored ready-mixed concrete to the transport vehicle C.

[0032] The manufacturing apparatus 10 described above is an example of a ready-mixed concrete manufacturing apparatus, and the ready-mixed concrete manufacturing apparatus may be configured in any way as long as it is capable of mixing concrete materials and manufacturing ready-mixed concrete. For example, the manufacturing apparatus 10 does not need to be equipped with the collection hopper 16, and various materials measured in the measuring bottles 14 may be supplied from the measuring bottles 14 to the mixer 20.

[0033] (Quality prediction device) The quality prediction device 50 (aggregate quality prediction device) is a device that predicts the quality of aggregate contained in concrete material. The quality prediction device 50 may predict the quality of at least some of the aggregate used by the manufacturing apparatus 10. It is known that the quality of the ready-mixed concrete manufactured by the manufacturing apparatus 10 is affected by the quality of the concrete material. Specifically, the particle size of the aggregate used to manufacture the ready-mixed concrete affects the fluidity (e.g., slump and slump flow) of the ready-mixed concrete. For example, for the same unit water content, the larger the particle size of the aggregate, the smaller the surface area of ​​the aggregate, which reduces the water retention capacity and may increase the fluidity of the ready-mixed concrete. Conversely, the smaller the particle size of the aggregate, the larger the surface area of ​​the aggregate, which increases the water retention capacity and may decrease the fluidity of the ready-mixed concrete. Thus, in order to stabilize the quality, including the fluidity, of the ready-mixed concrete, it is necessary to control the particle size of the aggregate.

[0034] The gradation of aggregate can be expressed by the coarse-grained fraction. In order to stabilize the quality of ready-mixed concrete, it is necessary to control, for example, whether the coarse-grained fraction of aggregate is within a desired range. The quality prediction device 50 may predict the coarse-grained fraction of aggregate as the quality of aggregate. The quality prediction device 50 may be configured to predict the coarse-grained fraction of aggregate from information including image data based on image data obtained by capturing an image of the aggregate, using a prediction model constructed by machine learning.

[0035] At least a part of the quality prediction device 50 is configured by one or more computers. The computer that configures at least a part of the quality prediction device 50 may be a personal computer, a tablet computer (tablet terminal), a smartphone, a wearable device, a workstation, a server computer, or a general-purpose computer.

[0036] When the quality prediction device 50 includes two or more computers, these computers may be communicatively connected to each other. The quality prediction device 50 may be communicatively connected to a control device included in the manufacturing system 1. The quality prediction device 50 may be a part of the function of the control device included in the manufacturing system 1 (i.e., may be included in the control device).

[0037] 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 be a computer main body (a main body portion of a computer) that executes the main functions of the computer that constitutes the quality prediction device 50.

[0038] The input / output device 54 is connected to the arithmetic unit 52. The input / output device 54 has a function of inputting information indicating instructions from a user such as a worker to the arithmetic unit 52, and a 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 an input device and an output device are integrated. The arithmetic unit 52 and the input / output device 54 may be integrated.

[0039] The imaging device 58 is a device (camera) capable of capturing an image of the aggregate. The imaging device 58 generates color image data by, for example, capturing an image of the aggregate. Hereinafter, the color image data generated by the imaging device 58 will be referred to as "captured image data P1." The imaging device 58 is a digital camera or video camera that captures an image within its field of view (imaging range) using visible light. The color captured image data P1 generated by the imaging device 58 is used to predict the quality of the aggregate. When the imaging device 58 generates video data, still image data included in the video may be acquired as the captured image data P1.

[0040] In the quality prediction device 50, the arithmetic device 52, the input / output device 54, and the imaging device 58 may be integrated into one device, such as a smartphone with a built-in (mounted) camera or a tablet computer with a built-in camera. In this case, the quality prediction device 50 may be portable by a worker or the like. In one example, the quality prediction device 50 is configured by installing an application for predicting the quality of aggregate on a smartphone or a tablet computer.

[0041] The imaging device 58 may be arranged in the manufacturing system 1 so as to be able to capture an image of the aggregate at a specific location. For example, the imaging device 58 is arranged so as to be able to capture an image of the aggregate while it is being transported by the transporting device 8. The imaging device 58 may be mounted on a mobile object such as a drone, and the imaging device 58 may capture an image of the aggregate by, for example, operating the mobile object by a user. The imaging device 58 may capture an image of the aggregate in a room where the imaging environment, such as illuminance, is adjusted.

[0042] 2, the arithmetic device 52 includes a circuit 91. The circuit 91 has a processor 92, a memory 93, a storage 94, and an input / output port 95. The storage 94 is configured with one or more non-volatile memory devices such as a flash memory or a hard disk. The storage 94 stores at least a quality prediction program for causing a computer to execute a quality prediction process, which will be described later. The storage 94 stores a quality prediction program for configuring each functional block, which will be described later, of the arithmetic device 52.

[0043] The memory 93 is composed of one or more volatile memory devices such as a random access memory. The memory 93 temporarily stores a quality prediction program loaded from the storage 94. The processor 92 is composed of one or more arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The processor 92 configures each functional block of the arithmetic unit 52 by executing the quality prediction program loaded into the memory 93. The results of arithmetic operations by the processor 92 are temporarily stored in the memory 93. The input / output port 95 inputs and outputs information to and from the input / output device 54, the imaging device 58, etc. in response to a request from the processor 92.

[0044] The circuit 91 is not necessarily limited to one in which each function is configured by a program. For example, the circuit 91 may have at least some of its functions configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates such a dedicated logic circuit. The quality prediction program may be provided by being permanently recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, the quality prediction program may be provided via a communication network as a data signal superimposed on a carrier wave.

[0045] 3 shows an example of functional components (referred to as "functional blocks" in this disclosure) included in the arithmetic device 52. The arithmetic device 52 has, as functional blocks, an imaging data acquisition unit 62 (first acquisition unit), a preprocessing unit 64, an input data acquisition unit 66 (second acquisition unit), a model construction unit 68 (construction unit), an imaging condition holding unit 69, a model holding unit 70, a prediction unit 72, and an output unit 74. The processing executed by these functional blocks corresponds to the processing executed by the arithmetic device 52 (quality prediction device 50).

[0046] The imaging data acquisition unit 62 is configured to acquire captured image data P1, which is color image data obtained by capturing an image of an aggregate whose quality is to be evaluated. Hereinafter, the aggregate whose quality is to be evaluated will be referred to as the "aggregate to be evaluated." The imaging data acquisition unit 62 acquires, for example, captured image data P1 from the imaging device 58, which is obtained by imaging with the imaging device 58.

[0047] In the color captured image data P1, pixel values ​​for the red, green, and blue color components are determined for each pixel. If the pixel value of the red component, the pixel value of the green component, and the pixel value of the blue component of each pixel are written as "R," "G," and "B," respectively, the pixel value of each pixel is written as (R, G, B). The pixel value is also called the brightness value, density value, shading value, tone value, or color value. The pixel value is expressed as a number ranging from 0 to 255, for example.

[0048] The pre-processing unit 64 performs predetermined pre-processing on the captured image data P1 to generate pre-processed image data. The pre-processing unit 64 performs grayscale conversion on the captured image data P1 to generate pre-processed image data (hereinafter referred to as "input image data P2"). The pre-processing unit 64 may perform other processing as part of the pre-processing, at least either before or after the grayscale conversion processing, while maintaining the input image data P2 as grayscale image data. The pre-processing unit 64 may perform only the grayscale conversion processing as pre-processing.

[0049] Here, the significance of performing preprocessing including grayscaling will be explained with reference to Fig. 4. The inventors have confirmed that the following problems arise when predicting aggregate particle size (coarse particle ratio) using a prediction model constructed by machine learning so as to output physical property values ​​related to aggregate particle size in response to input of color captured image data P1. They have confirmed that prediction accuracy decreases and the variance in prediction results increases when imaging conditions (imaging environment) such as illuminance or camera position differ between training and evaluation, and / or when the particle size of the aggregate to be evaluated is a particle size not used during training.

[0050] FIG. 4 shows image data a1 to a7, each representing a different image brightness, as color captured image data. Image data a4 is the original captured image data, and other image data are generated by adjusting the brightness of image data a4. As illustrated in FIG. 4, differences in brightness can result in different image features, even if the captured object is the same aggregate. Therefore, the inventors investigated various image preprocessing methods to prevent a decrease in prediction accuracy even when the imaging environment is different, and came up with the idea of ​​performing preprocessing, including grayscaling, on the captured image data P1 and using the grayscale image data as input data for the prediction model. However, further testing by the inventors revealed that simply grayscaling is not desirable from the perspective of quality prediction accuracy; details of this point will be discussed later. In the following description, "preprocessing" refers to preprocessing, including grayscaling, that generates grayscale image data as processed image data.

[0051] 5 shows a schematic diagram of a calculation process for prediction using a prediction model for predicting quality. The pre-processing unit 64 generates input image data P2 by performing pre-processing (for example, only grayscale processing) on ​​captured image data P1. In the pre-processing, the pre-processing unit 64 converts a 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 represented as "Y," the pre-processing 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)

[0052] 3, the input data acquisition unit 66 acquires input information including the input image data P2. The input data acquisition unit 66 may acquire only the input image data P2 as the input information. The input data acquisition unit 66 may also acquire, as the input information, information that may affect the granularity of the aggregate in addition to the input image data P2.

[0053] The model construction unit 68 constructs a model (hereinafter referred to as "prediction model M") for predicting the particle size of aggregate. The prediction model M is a model constructed by machine learning to output physical property values ​​of aggregate according to input information including input image data P2. The physical property values ​​of aggregate are physical property values ​​related to the particle size of aggregate. Specific examples of physical property values ​​related to the particle size of aggregate include the mass fraction retained between successive sieves (mass fraction for each particle size division), the mass fraction retained on each sieve, the mass fraction passing through each sieve, and the coarse particle fraction of aggregate, as defined in JIS A 1102:2014 "Test Method for Sieving of Aggregates"; the fine particle content of aggregate, as defined in JIS A 1102:2014 "Test Method for Fine Particle Content of Aggregates"; and the oversized particle content of aggregate (the amount of fine aggregate retained on a 5 mm sieve) and the undersized particle content of aggregate (the amount of coarse aggregate passing through a 5 mm sieve). The prediction unit 72 constructs a prediction model M by machine learning based on, for example, input information including input image data P2 and correct answer information of the mass fraction remaining between each successive sieve associated with the input information.

[0054] Machine learning is a technique in which a machine (computer) autonomously finds laws or rules by repeatedly learning based on given information. A predictive model M can be constructed using an algorithm and a data structure. A predictive model M is realized, for example, using a neural network, which is an information processing model that mimics the mechanism of the human brain and nerves. There are no particular restrictions on the specific algorithm of machine learning used when constructing a predictive model M. A neural network has an input layer, one or more intermediate layers, and an output layer. By including one or more intermediate layers, a more complex predictive model M can be constructed, thereby improving prediction accuracy.

[0055] The model construction unit 68 may construct the prediction model M by machine learning using a convolutional neural network (CNN). Using a convolutional neural network makes it possible to more appropriately capture image features. The model construction unit 68 may construct the prediction model M by machine learning based on the input image data P2 and correct answer information on the mass fraction remaining between successive sieves. The prediction model M constructed by the model construction unit 68 may output a value indicating the mass fraction remaining between successive sieves in accordance with the input image data P2. In other words, the prediction model M may output multiple values ​​indicating the mass fraction for multiple particle size classes.

[0056] The model construction unit 68 may autonomously construct a prediction model M by performing machine learning using data provided as input for machine learning and correct answer information of the output of the machine learning (such as correct values ​​of the mass fraction remaining between successive sieves). The input for the machine learning is various data sets of input information including the input image data P2. The output of the machine learning is data (numerical values) indicating physical property values ​​such as the mass fraction remaining between successive sieves. The model construction unit 68 iteratively trains a model that outputs physical property values ​​related to the granularity of the aggregate using multiple combinations of data sets of input information and correct values ​​of physical property values.

[0057] The stage of autonomously constructing the predictive model M corresponds to the training phase. The training phase is also called the learning phase. The same (mutually corresponding) information and aggregates are used in the training phase and the evaluation phase. Hereinafter, the terms "for training" and "for evaluation" may be used to distinguish between the information and aggregates used in the training phase and the information and aggregates used in the evaluation phase.

[0058] The imaging condition holding unit 69 holds information representing the environmental conditions (imaging conditions) when the training bones are imaged in the training phase. The information representing the environmental conditions held (stored) by the imaging condition holding unit 69 includes the illuminance when the training bones are imaged. The model holding unit 70 holds the prediction model M constructed by the model construction unit 68. The prediction model M, which is 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.

[0059] In the evaluation phase, the prediction unit 72 predicts the quality of the aggregate to be evaluated. The prediction unit 72 uses the prediction model M to acquire physical property values ​​corresponding to the input information for evaluation acquired by the input data acquisition unit 66. For example, the prediction unit 72 inputs the input information for evaluation to the prediction model M, and then acquires values ​​indicating the mass fraction remaining between successive sieves that are output from the prediction model M. The prediction unit 72 may calculate a predicted value of the coarse particle fraction of the aggregate to be evaluated by converting the values ​​indicating the mass fraction remaining between successive sieves into a coarse particle fraction. The prediction unit 72 may correct the values ​​indicating the multiple mass fractions output from the prediction model M before converting them into a coarse particle fraction.

[0060] The output unit 74 outputs the quality prediction result by the prediction unit 72 to the monitor of the input / output device 54. The output unit 74 may display the predicted value of the coarse particle ratio calculated by the prediction unit 72 on the monitor. In addition to the predicted value of the coarse particle ratio, the output unit 74 may display the value obtained from the prediction model M regarding the mass fraction remaining between successive sieves on the monitor. The output unit 74 may display a graph of the particle size distribution on the monitor based on the predicted value of the mass fraction remaining between successive sieves.

[0061] <Environmental conditions when capturing image data P1> In quality prediction using the prediction model M by the quality prediction device 50, the relationship between illuminance, which is one of the environmental conditions when imaging aggregates to obtain captured image data P1, is defined between the training phase and the evaluation phase. In the training phase, the illuminance in the vicinity of the training aggregates imaged by the imaging device 58 may be measured. For example, when the imaging device 58 images the training aggregates stacked on a tray, an illuminance meter may be installed on the imaging table on which the tray is placed, and the illuminance may be measured as one of the environmental conditions (imaging environment).

[0062] Here, "illuminance IA" and "illuminance IB" are defined as follows: Each of illuminance IA and illuminance IB represents the illuminance in the vicinity of the bone material to be imaged. · Illuminance IA: Illuminance [lux: lx], which is one of the environmental conditions when imaging the training aggregate to prepare training input information including the training image data P1 during the training phase. Illuminance IB: Illuminance [lux: lx], which is one of the environmental conditions when capturing an image of the aggregate to be evaluated in order to obtain the captured image data P1 for evaluation during the evaluation phase.

[0063] In order to prevent a decrease in the prediction accuracy of the prediction model M, the illuminance IB is equal to or greater than 0.2 times the illuminance IA and equal to or less than 5 times the illuminance IA. That is, the imaging data acquisition unit 62 acquires, as the captured image data P1, image data obtained by capturing an image under environmental conditions (imaging conditions) in which the illuminance IB is equal to or greater than 0.2 times the illuminance IA and equal to or less than 5 times the illuminance IA. In this case, in the evaluation phase, the imaging device 58 captures an image of the aggregate for evaluation under environmental conditions (imaging conditions) in which the illuminance IB is equal to or greater than 0.2 times the illuminance IA and equal to or less than 5 times the illuminance IA.

[0064] The minimum value of the range specifying the illuminance IB may be 0.3 times the illuminance IA, 0.4 times the illuminance IA, or 0.5 times the illuminance IA, from the viewpoint of further suppressing a decrease in the prediction accuracy of the prediction model M. The maximum value of the range specifying the illuminance IB may be 4 times the illuminance IA, 3 times the illuminance IA, 2 times the illuminance IA, or 1.5 times the illuminance IA, from the viewpoint of further suppressing a decrease in the prediction accuracy of the prediction model M. In one example, the illuminance IB is 0.5 times or more and 2 times or less the illuminance IA. When the illuminance IA is 3000 lx to 4000 lx, the illuminance IB may be 0.5 times or more and 2 times or less the illuminance IA.

[0065] From the viewpoint of suppressing a decrease in the prediction accuracy of the prediction model M, the absolute value of the difference between the illuminance IA and the illuminance IB may be 2500 lx or less. From the viewpoint of further suppressing a decrease in the prediction accuracy of the prediction model M, the absolute value of the difference between the illuminance IA and the illuminance IB may be 2400 lx, 2300 lx, 2200 lx, 2100 lx, or 2000 lx or less. In one example, the illuminance IB is equal to or greater than a value obtained by subtracting 2000 lx from the illuminance IA and equal to or less than a value obtained by adding 2000 lx to the illuminance IA (within a range of IA±2000 lx). The illuminance IA may be greater than 2500 lx. The illuminance IA may be greater than 2600 lx, 2700 lx, 2800 lx, or 2900 lx. When the illuminance IA is 3000 lx to 4000 lx, the absolute value of the difference between the illuminance IA and the illuminance IB may be within 2000 lx.

[0066] The illuminance IB may be 2000 lx or more. The illuminance IB may be in the range of 2000 lx to 6000 lx. Generally, on a sunny day, the illuminance at a location 1 m inward from a south-facing window in a room is 3000 lx to 5000 lx. Therefore, by setting the illuminance IB within the range of 2000 lx to 6000 lx, it is easy to adjust the environmental conditions when capturing images indoors. The illuminance IB may be 2200 lx or more, 2400 lx or more, 2500 lx or more, 2600 lx or more, 2700 lx or more, 2800 lx or more, or 2900 lx or more. The illuminance IB may be 6000 lx or less, 5800 lx or less, 5600 lx or less, 5500 lx or less, 5400 lx or less, 5300 lx or less, 5200 lx or less, or 5100 lx or less. In one example, the illuminance IB is in the range of 2500 lx to 5500 lx, or in the range of 2900 lx to 5100 lx. Both the illuminance IA and the illuminance IB may be in the range of 2000 lx to 6000 lx. When the illuminance IA is 3000 lx to 5000 lx, the illuminance IB may be 2000 lx to 6000 lx, or 2500 lx to 5500 lx. When the illuminance IA is 3000 lx to 4000 lx, the illuminance IB may be 2000 lx to 6000 lx, or 2500 lx to 5500 lx.

[0067] To summarize the range of illuminance IB, the range of illuminance IB satisfies at least the following condition 1. The range of illuminance IB may also satisfy both the following conditions 1 and 2, both the following conditions 1 and 3, or all of the following conditions 1, 2, and 3. Condition 1: Illuminance 0.2 times or more than IA and 5 times or less than IA. Condition 2: The absolute value of the difference between illuminance IA and illuminance IB is 2500 lx or less. Condition 3: 2000lx or more. For example, when both Condition 1 and Condition 2 are satisfied, the minimum value of the range of illuminance IB is determined by the larger of the minimum values ​​of Condition 1 and Condition 2, and the maximum value of the range of illuminance IB is determined by the smaller of the maximum values ​​of Condition 1 and Condition 2. When both Condition 1 and Condition 3 are satisfied, the minimum value of the range of illuminance IB is determined by the larger of the minimum values ​​of Condition 1 and Condition 3, and the maximum value of the range of illuminance IB is determined by the maximum value of Condition 1. When all of Conditions 1, 2, and 3 are satisfied, the minimum and maximum values ​​of the range of illuminance IB are determined in the same way as when both Conditions 1 and 2 are satisfied.

[0068] <Prediction model M> Next, the prediction model M used to predict the coarse particle ratio will be further described with reference to Figures 5 and 6. A portion of Figure 5 schematically shows the calculation process of predicting the coarse particle ratio using the prediction model M from input image data P2, which is at least a portion of the input information to the prediction model M. The input to the prediction model M is, for example, only the input image data P2, and the output of the prediction model M is, for example, the mass fraction remaining between each successive sieve.

[0069] In the training phase for constructing the prediction model M, the same type of aggregate as the aggregate to be evaluated is used as the training aggregate. When constructing the prediction model M, training data (hereinafter referred to as "training data TD") for machine learning is prepared. The training data TD includes training input information and correct answer information for physical property values ​​associated with the training input information. The training data TD is composed of multiple data sets, and each of the multiple data sets includes, for example, training input image data P2 and correct answer information for the mass fraction remaining between each successive sieve associated with the input image data P2. The training input image data P2 differs among the multiple data sets.

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

[0071] The imaging device 58 captures an image of an area including the aggregate on the tray, from a position a predetermined distance away from the tray on which the training aggregate is piled (for example, a predetermined position above). The imaging by the imaging device 58 may be performed indoors or outdoors. The amount of aggregate placed on the tray when imaging is performed is not particularly limited, but may be approximately 0.1 kg to 10 kg. At least a portion of the aggregate being transported on the belt conveyor included in the transporting device 8 may be used as the training aggregate. In this case, the imaging device 58 may capture an image of the aggregate being transported on the belt conveyor of the transporting device 8. The imaging device 58 may be installed so that the amount of aggregate being transported included in the imaging field of the imaging device 58 is approximately 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 the prepared samples are imaged to obtain training image data P1. The mass fraction and coarse particle fraction remaining between successive sieves for the various samples prepared under the conditions shown in Table 1 are known, and at least a portion of these known values ​​are used as ground truth information in machine learning.

[0073] [Table 1]

[0074] The fineness fraction (FM) can be calculated, for example, by the following formula (2) specified in JIS A 1102:2014 "Sieve Analysis Test Method for Aggregates." Note that in the "Standard Specifications for Building Construction and Commentary JASS5 Reinforced Concrete Construction," nine sieves, excluding the 80 mm sieve, are used in formula (2). The fineness fraction predicted in this disclosure may be specified by either method.

[0075]

number

[0076] In Table 1, this means that aggregates of 10 levels of gradation (coarseness ratio) are prepared. Each level of aggregate is sieved successively in order of coarseness of the sieve and classified into categories I to IV. Category I represents the particle size category for particles greater than 10 mm and smaller than 20 mm. Particles that make up Category I are those that are retained on a 10 mm sieve but pass through a 20 mm sieve. Category II represents the particle size category for particles greater than 5 mm and smaller than 10 mm. Particles that make up Category II are those that are retained on a 5 mm sieve but pass through a 10 mm sieve. Category III represents the particle size category for particles larger than 2.5 mm and smaller than 5 mm. Particles that make up Category III are those that are retained on a 2.5 mm sieve but pass through a 5 mm sieve. Category IV represents the particle size category for particles greater than 1.2 mm and smaller than 2.5 mm. Particles that make up Category IV are those that are retained on a 1.2 mm sieve but pass through a 2.5 mm sieve. The mass fraction retained on each successive sieve means that, for example, level 1 in Table 1, the particles in category I were 85 wt%, the particles in category II were 15 wt%, the particles in category III were 0 wt%, and the particles in category IV were 0 wt%.

[0077] Image data T1 to T10 shown in FIG. 6 are examples of image data obtained by imaging training aggregates prepared according to the conditions of each of Levels 1 to 10 in Table 1. The image data illustrated in FIG. 6 was obtained by imaging the training aggregates placed on a sheet that simulates the transportation of aggregates on a belt conveyor. Instead of such a sheet, the training aggregates may be placed on a flat tray, and image data related to the training aggregates prepared according to the conditions of each of Levels 1 to 10 in Table 1 may be obtained. Level k (k is an integer greater than or equal to 1) corresponds to image data Tk. For example, image data T1 is obtained by imaging training aggregates prepared according to Level 1. Training aggregates according to each Level k contain particles of multiple gradation classes.

[0078] For each level k, multiple image data Tk may be prepared, each showing a different state (more specifically, pile state) of the training aggregate. In one example, a first image is taken with the training aggregate placed on a tray, thereby acquiring one image data Tk. Then, the training aggregate is temporarily removed from the tray, and the same training aggregate (same sample) is placed on the tray again so that the pile state of the multiple particles is different from that in the first image. After that, a second image is taken. By repeating the above-described image capture, multiple image data Tk may be prepared for each level k.

[0079] In addition to the conditions shown in Table 1, various samples may be prepared as training bone materials under the conditions shown in Tables 2 and 3 below, and training captured image data P1 (image data Tk) may be prepared.

[0080] [Table 2]

[0081] [Table 3]

[0082] In one example, one data set in the training data is composed of one image data Tk and the correct answer information for the mass fraction at level k associated with that image data Tk. Multiple data sets, each consisting of a combination of image data Tk and correct answer information for the mass fraction, may be prepared as training data. Machine learning using such training data may then be performed to construct a prediction model M that outputs predicted values ​​for four mass fractions for four particle size classes. The mass fractions remaining between successive sieves at each level in the training data, the number of levels (the value of k), and the particle size classes may be set arbitrarily, without being limited to the examples shown in Tables 1 to 3.

[0083] An example of the prediction model M will be described using simplified formulas for ease of understanding. The prediction model M constructed by the model construction unit 68 can be simply expressed as, for example, the following formulas (3) and (4).

number

number

[0084] In equation (4), Y represents the output value of the physical property, for example, the output value of the mass fraction retained between successive sieves. In this case, equations (3) and (4) are calculated for each particle size classification. N is an integer of 2 or greater and represents the number of data sets in the input information. xl represents various input values ​​included in the input information, and the input information includes at least input image data P2 based on captured image data P1.

[0085] wl is a weight (coefficient), and b is a bias term (coefficient). f(U) represents the activation function. The activation function can be 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, softsign function, tanh function, tanhShrink function, hardtanh function, tanhexp function, ReLU function, ReLU6 function, Leaky-ReLU function, PReLU function, ELU function, SELU function, CELU function, Swith function, Mish function, or ACON function.

[0086] The model construction unit 68 may use training data to repeatedly evaluate the error (loss) between Y (predicted value) obtained by equation (4) and the correct information for the physical property values, and determine the weight wl and bias term b in equation (3) so as to minimize the error. The model construction unit 68 may use any type of loss function as a function for evaluating the error between Y (output value from an intermediate model at an intermediate stage in constructing the prediction model M) obtained by equation (4) and the correct information for the physical property values. The loss function is a function that calculates a loss value based on the error between the predicted value and the correct value, and the weight wl is updated so as to minimize the loss value. Specific examples of loss functions include the Huber loss function, the mean absolute error, the mean squared error, and the ε-allowable loss function.

[0087] The model construction unit 68 may repeatedly update the weight wl using a gradient method so as to minimize the loss value evaluated by the loss function. The model construction unit 68 may use any type of update formula (weight update formula) when updating the weight wl. Specific examples of weight update formulas include Adam, AdamW, AdamBelief, Adamax, AdaBound, Adagrad, AMSGRAD, AMSBound, Lars, RMSprop, Adadelta, Sgd, SgdW, Momentum, and Nesterov. The weight update formula is also referred to as an optimization algorithm or an optimization method.

[0088] [Ready-mix concrete manufacturing method] Next, an example of a method for producing ready-mixed concrete executed in the production system 1 will be described. The method for producing ready-mixed concrete includes a production process and a quality prediction process. The production process is a process for kneading materials including aggregate to produce ready-mixed concrete. The quality prediction process is a process for predicting the quality of at least a portion of the aggregate used in the production process. The quality prediction process may be executed during a period that overlaps with at least a portion of the period during which the production process is executed.

[0089] (manufacturing process) The manufacturing process includes, for example, a transporting process, a measuring process, a feeding process, a mixing process, a discharging process, and a loading process. In the transporting process, various concrete materials are transported by a transporting device 8 to storage bottles 12, and the various materials are individually supplied to storage bottles 12. In the measuring process, the various materials are individually supplied from storage bottles 12 to measuring bottles 14, and the various materials are measured in measuring bottles 14. In the measuring process, when the measured amount of each material reaches a predetermined set amount, the material is discharged into a collecting hopper 16.

[0090] In the charging process, after all types of materials have been collected in the collecting hopper 16, the materials in the collecting hopper 16 are charged (supplied) into the mixer 20. In the mixing process, multiple types of concrete materials are mixed in the mixer 20. In the discharging process, after mixing of the concrete materials is completed in the mixer 20, 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.

[0091] (Quality prediction process) The quality prediction process (aggregate quality prediction method) includes a model construction process in a training phase and a quality evaluation process in an evaluation phase. In the quality prediction process, the model construction process is executed before the quality evaluation process.

[0092] The model construction process includes, for example, a preparation process and a construction process. The preparation process and the construction process may be executed by the model construction unit 68 of the calculation device 52. The preparation process is a process of preparing training data TD including training input information corresponding to the evaluation input information and correct answer information for physical property values ​​associated with the training input information. The construction process is a process of constructing a prediction model M by machine learning based on the training data TD prepared in the preparation process.

[0093] The preparation process may include generating extended image data by performing a conversion process and a grayscale process, in any order, on the training image data P1 corresponding to the evaluation image data P1. The term "extension" in the extended image data refers to extending the dataset initially prepared for use in the training phase (increasing the number of datasets). The conversion process in the preparation process may include randomly changing the brightness and contrast of the image. In addition to changing the brightness and contrast, the conversion process in the preparation process may also include at least one of flipping the image and rotating the image. That is, the conversion process may include either flipping the image or rotating the image, or both flipping and rotating the image. The input image data P2 in the training input information may include the extended image data, which is a grayscale image. Details of the conversion process for extending the training data TD will be described later.

[0094] The construction process may include preparing model evaluation data MED separately from the training data TD. The model evaluation data MED includes input information for model evaluation and correct information for physical properties associated with the input information for model evaluation. The model evaluation data MED is composed of multiple datasets of the same type as the multiple datasets included in the training data TD. The "model evaluation" in the model evaluation data MED is different from evaluating quality in the evaluation phase and refers to evaluating an intermediate model generated during the construction stage of a predictive model when building the predictive model in the training phase. The construction process may include continuously updating the intermediate model using the training data TD to reduce the error from the correct information for physical properties, and determining one of the intermediate models generated during the update as the predictive model M using the model evaluation data MED. A specific example of a method for determining the predictive model M using the model evaluation data MED will be described later.

[0095] The quality evaluation process includes, for example, a first acquisition process, a preprocessing process, a second acquisition process, a prediction process, and an output process. The first acquisition process is a process of acquiring captured image data P1, which is color image data obtained by capturing an image of the aggregate to be evaluated. The captured image data acquisition unit 62 may execute the first acquisition process. The preprocessing process is a process of grayscaling the captured image data P1 acquired in the first acquisition process to generate input image data P2. The preprocessing unit 64 may execute the preprocessing process.

[0096] The second acquisition step is a step of acquiring input information including the input image data P2 generated in the preprocessing step. The input data acquisition unit 66 may execute the second acquisition step. The prediction step is a step of acquiring physical property values ​​(physical property values ​​related to the particle size of the aggregate) corresponding to the input information acquired in the second acquisition step, using a prediction model M constructed in advance by machine learning so as to output physical property values ​​of the aggregate in response to input of the input information. In the prediction step, the quality of the aggregate, such as the coarseness ratio, is predicted based on the predicted values ​​of the physical property values ​​output from the prediction model M. The prediction unit 72 may execute the prediction step. The output step is a step of outputting the quality of the aggregate predicted in the prediction step to the input / output device 54. The output unit 74 may execute the output step.

[0097] An example of the model construction process and an example of the quality evaluation process will be described below with reference to Figures 7 to 11. Note that the description will be given using an example in which the aggregate coarseness fraction is predicted as the quality of the aggregate, and the input information to the prediction model M includes only input image data P2. The description will also be given using an example in which the prediction model M outputs a predicted value of the mass fraction for each of the four particle size classifications.

[0098] <Model building process> 7 is a flowchart showing an example of a series of processes executed in the model construction step. In the model construction step, a portion of the aggregate (e.g., coarse aggregate) used in the production of ready-mixed concrete in the production apparatus 10 may be extracted, and the extracted aggregate may be used as training aggregate.

[0099] In the model building process, step S11 is first executed. In step S11, for example, a worker or the like acquires information necessary for preparing training data TD for machine learning and model evaluation data MED used in building a prediction model M. In one example, the worker or the like prepares various types of aggregates whose mass fractions remaining between successive sieves are known. The worker or the like may prepare various types of aggregates having gradations of levels 1 to 30 in Tables 1 to 3 described above. The worker or the like may prepare M pieces of training image data P1, each differing from the other in at least one of gradation level and deposition state. M may be 10 to 100,000, or may be 100 to 10,000. The known mass fractions for each gradation level are associated with each of the M pieces of training image data P1 as correct answer information for physical properties.

[0100] In the imaging to obtain the M pieces of training image data P1, the training bone may be imaged by the imaging device 58 under substantially the same illuminance conditions. The illuminance IA, which is one of the environmental conditions when imaging the training bone to obtain the training image data P1, may be within the range exemplified in the above-mentioned "environmental conditions when obtaining the image data P1."

[0101] Next, step S12 is executed. In step S12, for example, the environmental conditions when an operator or the like images the training aggregate to obtain the training image data P1 are stored in the imaging condition storage unit 69. The operator or the like may input the illuminance IA, which is one of the environmental conditions, to the calculation device 52 by operating an input device of the input / output device 54.

[0102] Next, step S13 is executed. In step S13, for example, the model construction unit 68 creates (generates) extended image data based on the M pieces of training captured image data P1 prepared in step S11 in accordance with a predetermined calculation procedure in order to extend the data used in the training phase. Fig. 8 schematically shows the process of creating extended image data from the training captured image data P1. In Fig. 8, the training captured image data P1 is indicated by "TP".

[0103] For example, the model construction unit 68 generates Q pieces of image data TP1 to TPQ by randomly changing the brightness and contrast of each of the M pieces of image data TP (for each piece of image data TP). Q is an integer equal to or greater than 2, and FIG. 8 illustrates an example where Q is 7. As a result of randomly changing the brightness and contrast, only the brightness or only the contrast may be changed in some of the image data TP1 to TPQ. As a result of randomly changing the brightness and contrast, the image data TP1 to TPQ may include images for which neither the brightness nor the contrast has been changed. The model construction unit 68 may generate new image data based on the image data TP so that one of the image data TP1 to TPQ includes the image data TP as is without any change processing. Any one of the image data TP1 to TPQ will be referred to as "image data TPq" (q is an integer from 1 to Q).

[0104] The model construction unit 68 generates R pieces of image data TPq1 to TPqR by, for example, at least one of image flipping and image rotation for each of the image data TP1 to TPQ (for each piece of image data TPq). R is an integer equal to or greater than 2, and FIG. 8 illustrates an example where R is 4. While FIG. 8 illustrates an example where new image data is generated based on the image data TP3, new image data is similarly generated for each of the image data TP1, TP2, TP4 to TP7. The model construction unit 68 may generate new image data based on the image data TPq so that the image data TPq is included as is in one of the image data TPq1 to TPqR.

[0105] In one example, the model construction unit 68 generates four pieces of image data from the image data TPq. In Fig. 8, image data TP31 is the same as the image data TP3, and image data TP32 is data generated by flipping the image data TP3 horizontally. Image data TP33 is data generated by flipping the image data TP3 vertically, and image data TP34 is data generated by flipping the image data TP3 both vertically and horizontally.

[0106] In the inversion for generating new image data based on one image data, the inversion may be performed about any axis, for example, horizontal inversion, vertical inversion, or both horizontal and vertical inversion. In the rotation for generating new image data based on one image data, the rotation may be performed at any angle in the range of 0° to 360°, and the rotation direction may be either clockwise or counterclockwise. At least one of the following rotations may be performed: rotating the image clockwise by 90°, rotating the image counterclockwise by 90°, and rotating the image 180°. After rotating at an angle other than 90° or 180°, correction may be performed so that the outer edges of the image are aligned vertically and horizontally while maintaining the rotated state.

[0107] By generating new image data as described above, the model construction unit 68 generates (M×Q×R) pieces of image data based on the M pieces of image data TP. The model construction unit 68 maintains the state in which the correct answer information for physical property values ​​is associated with the newly generated image data. The correct answer information associated between the newly generated image data and the image data on which the newly generated image data is based is the same. As described above, M data sets may be expanded to (M×Q×R) data sets.

[0108] Next, step S14 is executed. In step S14, for example, the model construction unit 68 performs grayscale conversion processing on each of the (M×Q×R) pieces of image data prepared in step S13 in accordance with a predetermined calculation procedure. The model construction unit 68 may perform grayscale conversion using the above-mentioned equation (1) on each piece of image data prepared in step S13. As a result, (M×Q×R) pieces of extended image data, which are grayscaled image data, are obtained, and this grayscaled extended image data is used as the input image data P2.

[0109] By executing steps S11, S13, and S14, (M×Q×R) pieces of training input image data P2 are prepared. The (M×Q×R) pieces of training input image data P2 have been subjected to conversion processing to expand the dataset used in the training phase, but have also been generated by grayscaling the training captured image data P1. Each of the (M×Q×R) pieces of training input image data P2 is associated with a known mass fraction for each particle size classification as ground truth information for the physical property value.

[0110] As described above, the model construction unit 68 may prepare (M×Q×R) data sets as data to be used in the training phase. Each of the (M×Q×R) data sets (each data set) includes training input image data P2 and correct answer information on the mass fraction remaining between successive sieves, which is associated with the input image data P2. A portion of the (M×Q×R) data sets may be used as training data TD, and the remaining portion may be used as model evaluation data MED.

[0111] Next, steps S15 and S16 are executed. In step S15, for example, the model construction unit 68 constructs a prediction model M by machine learning based on the training data TD prepared up to the execution of step S14. In the process of constructing the prediction model M, the model construction unit 68 may use the model evaluation data MED to determine the prediction model M to be used in the quality evaluation step. The model construction unit 68 may construct a prediction model M that outputs multiple mass fraction values ​​for multiple particle size classes in response to input of input image data P2. In step S16, for example, the model holding unit 70 stores the prediction model M constructed in step S15. This completes the series of processes executed in the model construction step.

[0112] FIG. 9 is a flowchart showing an example of a series of processes executed when constructing a prediction model M in step S15. FIG. 10 is a diagram for explaining the series of processes shown in FIG. 9. In machine learning for constructing a prediction model M, the number of epochs (total number of learning times) is represented as "K", and the number of epochs is represented as i (i is an integer equal to or greater than 2). The number of divisions of the training data TD in one epoch (each epoch) is represented as "L", and each piece of data obtained by dividing the training data TD into L pieces is referred to as "batch data". Each batch of data in the training data TD is represented as "batch data Bj" (j is an integer equal to or greater than 2).

[0113] To facilitate understanding of the explanation, it is assumed that the training data TD consists of 3,000 data sets, and L, which represents the number of divisions of the training data per epoch, is 10. In this case, each batch of data consists of 300 data sets, and the training data TD is divided into batches B1 to B10. It is also assumed that K, which is the number of epochs indicating the number of times (total number) of learning, is a multiple of 10. K may be 30 to 100,000, 100 to 10,000, or 1,000 to 5,000. L and K may be arbitrarily set by an operator or the like.

[0114] Steps S51 and S52 are executed with L and K already set. In step S51, for example, the model construction unit 68 sets each of the variables i and j to 1. In step S52, for example, the model construction unit 68 randomly selects 300 pieces of data (data sets) from the training data TD to prepare batch data B1 to B10.

[0115] Next, steps S53 and S54 are executed. In step S53, for example, the model construction unit 68 uses the batch data Bj to update the weight wl as exemplified by the above equation (2). The model construction unit 68 uses the batch data Bj to determine the weight wl so that the above-mentioned loss value is minimized. In step S54, for example, the model construction unit 68 increments j.

[0116] Next, step S55 is executed. In step S55, for example, the model construction unit 68 determines whether j is greater than L (=10). If it is determined that j is equal to or less than L (step S55: NO), the process returns to step S53. In this case, the model construction unit 68 executes the series of processes from steps S53 to S55 again. The model construction unit 68 repeats the series of processes from steps S53 to S55 until j exceeds L. A part of FIG. 10 illustrates how ten weight updates are performed in the first epoch (i=1). Batch data B1 is used in the first weight update, batch data B2 is used in the second weight update, and batch data B10 is used in the tenth weight update. Although omitted, corresponding batch data Bj are also used in the third to ninth weight updates.

[0117] If it is determined in step S55 that j is greater than L (step S55: YES), the process proceeds to step S56. In step S56, for example, the model construction unit 68 stores the model Mi defined by the updated weight wl as an intermediate model of the prediction model M. The model Mi refers to the intermediate model constructed at the time when the i-th epoch (i epochs in total) has been executed.

[0118] Next, steps S57 and S58 are executed. In step S57, for example, the model construction unit 68 increments i and sets j to 1. In step S58, for example, the model construction unit 68 determines whether i is greater than K. If it is determined that i is equal to or less than K (step S58: NO), the process returns to step S52. In step S52, which is executed again, the model construction unit 68 randomly reselects 300 pieces of data (data sets) from the training data TD to prepare new batch data B1 to B10. As a result, different batch data B1 to B10 are prepared between the ith epoch and the (i-1)th epoch.

[0119] The model construction unit 68 repeats the series of processes in steps S52 to S58 until i exceeds K. As a result, updating of the weight wl using randomly selected batch data B1 to B10 is repeated for each epoch until training in the Kth epoch is completed, as shown in Fig. 10. Note that Fig. 10 only illustrates the case where the weight wl is updated using batch data B10 in the second and subsequent epochs, and omits the process of updating the weight wl using each of the batch data B1 to B9.

[0120] If it is determined in step S58 that i is greater than K (step S58: YES), the process proceeds to step S59. In step S59, for example, the model construction unit 68 calculates a loss value for the latest model Mi obtained at that time point using the model evaluation data MED for every arbitrarily set number of epochs (for example, every 1 to 10 epochs). FIG. 10 illustrates an example in which the loss value for model Mi is evaluated using the model evaluation data MED every 10 epochs. In this case, the model Mi obtained at the time when training (weight update) is completed at the 10th, 20th, . . . , Kth epochs is evaluated for its loss value using the model evaluation data MED.

[0121] In the evaluation of the loss value using the model evaluation data MED, the loss value may be calculated for each data set included in the model evaluation data MED, and statistical information (e.g., arithmetic mean) of the loss values ​​may be used as the evaluation index. The model construction unit 68 may determine (select) the model Mi having the smallest loss value, which is the evaluation index, as the prediction model M from among the multiple models Mi that have been evaluated using the model evaluation data MED.

[0122] <Quality evaluation process> 11 is a flowchart showing an example of a series of processes executed in the quality evaluation step. This quality evaluation step is performed, for example, during a period that overlaps with at least a part of the period during which the above-described production step is executed by the production apparatus 10. In the quality evaluation step, a part of the aggregate used in the production of ready-mixed concrete by the production apparatus 10 may be extracted, and the quality of the aggregate may be evaluated as the evaluation target.

[0123] In the quality evaluation process, step S21 is executed first. In step S21, the imaging data acquisition unit 62 may display on the monitor of the input / output device 54 the range of illuminance IB when the captured image data P1 is obtained in the quality evaluation process, based on the illuminance IA stored in the imaging condition storage unit 69. The range of illuminance IB is the same as the range exemplified in the above-mentioned "environmental conditions when acquiring the captured image data P1." Note that if it is guaranteed that the illuminance IB when the captured image data P1 is obtained in the quality evaluation process is within the range, step S21 may be omitted.

[0124] Next, step S22 is executed. In step S22, for example, the imaging data acquisition unit 62 acquires evaluation image data P1 obtained by capturing an image of the aggregate for evaluation. The imaging data acquisition unit 62 may acquire the image data P1 from the imaging device 58. The imaging device 58 may capture an image of the aggregate to be evaluated while a worker or the like is performing some of the work. In this case, the worker or the like may check whether the illuminance IB is within the range displayed on the monitor in step S21. Step S22 may be executed at a predetermined evaluation timing. The evaluation timing may be set to a certain time of day or may be a timing instructed by a worker or the like. The conditions other than the illuminance when capturing an image by the imaging device 58 may be set to approximately match the conditions in the model construction process, within a settable range.

[0125] Next, step S23 is executed. In step S23, for example, the pre-processing unit 64 performs a grayscale conversion process on the evaluation captured image data P1 acquired in step S22, thereby generating input image data P2. The pre-processing unit 64 performs grayscale conversion using the same calculation procedure (same image processing procedure) as the grayscale conversion process in the model construction step. The pre-processing unit 64 may perform grayscale conversion on the evaluation captured image data P1 using the above-mentioned equation (1).

[0126] 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 into the prediction model M held by the model holding unit 70, and then obtains a predicted value of the mass fraction remaining between successive sieves output from the prediction model M. In step S25, for example, the prediction unit 72 converts the predicted value of the mass fraction remaining between successive sieves obtained in step S24 into a coarse particle fraction, thereby calculating a predicted value of the coarse particle fraction.

[0127] Next, step S26 is executed. In step S26, for example, the output unit 74 outputs the predicted value of the coarse-grained fraction calculated in step S25 to the monitor of the input / output device 54. In addition to the predicted value of the coarse-grained fraction, the output unit 74 may also output the predicted value of the mass fraction remaining between each successive sieve obtained in step S24 to the monitor of the input / output device 54. By executing step S26, a worker or the like can check the predicted result of the coarse-grained fraction. If the coarse-grained fraction is outside the control range, the worker or the like who has checked the predicted result of the coarse-grained fraction may take measures to maintain the quality of the ready-mixed concrete (for example, adjusting the gradation of the aggregate, correcting (modifying) the mix, temporarily suspending production, investigating the cause, etc.).

[0128] This completes the series of processes for predicting the coarse particle ratio once. After step S26 is executed, the series of processes from steps S21 to S26 may be repeated. In the above example, some aggregates are extracted and quality evaluation (sampling inspection) is performed, but all aggregates used in the production of ready-mix concrete may also be inspected.

[0129] [Verification of prediction accuracy] Next, with reference to Figures 12(a) and 12(b), we will explain the results of verifying the prediction accuracy of prediction model M, which was constructed by machine learning using a convolutional neural network. In this verification, we used a dataset in which the correct values ​​were known, and compared the prediction results by prediction model M with the correct values.

[0130] (Verification example 1A: air-dry condition, illuminance IB=3220lx) First, 10 levels of training aggregate with different coarseness ratios were prepared according to the conditions in Table 1 above. For each level, 4 kg of training aggregate was placed on a flat tray (a rimmed tray). Ten images were captured indoors, each with a different state of the aggregate on the tray. Air-dried aggregate, which represents a dry state in a natural environment, was used as the training aggregate. The illuminance IA, one of the environmental conditions during imaging, was set to 3420 lx. The 10 images were obtained according to the conditions in Table 1, resulting in 100 data sets in which the image data was associated with the correct value for the mass fraction remaining between each successive sieve.

[0131] Next, 10 levels of aggregate (air-dried state) with different coarseness ratios were prepared according to the conditions in Table 2 above. Using a method similar to the preparation according to the conditions in Table 1, 100 image data were obtained according to the conditions in Table 2, and 100 data sets were obtained in which the image data was associated with the correct value for the mass fraction remaining between each successive sieve. Furthermore, 10 levels of aggregate (air-dried state) with different coarseness ratios were prepared according to the conditions in Table 3 above. Using a method similar to the preparation according to the conditions in Table 1, 100 image data were obtained according to the conditions in Table 3, and 100 data sets were obtained in which the image data was associated with the correct value for the mass fraction remaining between each successive sieve.

[0132] After obtaining 300 datasets according to the conditions in Tables 1, 2, and 3, new image data were created to expand the dataset used in the training phase by performing the process in step S13 described above. Specifically, when randomly changing the brightness and contrast, 11 patterns of new image data were generated from one original image data. When changing the image by flipping or rotating it, four patterns of new image data were generated from one original image data. As a result, the 300 image data were expanded to (300 x 11 x 4 = 13,200) image data. When randomly changing the brightness and contrast, random numbers were generated by computer processing and the change conditions were varied. When changing the image by flipping or rotating it, new image data were generated using four conversion patterns: no flip, vertical flip, horizontal flip, and vertical and horizontal flip.

[0133] Next, each of the 13,200 pieces of image data was subjected to grayscale conversion in step S14 to generate 13,200 pieces of input image data P2. In the grayscale conversion process, grayscale conversion was performed for each pixel using the above-mentioned formula (1). As a result, 13,200 data sets were obtained in which the input image data P2 was associated with the correct information on mass fraction.

[0134] Of the 13,200 datasets, 9,240 datasets were used as training data TD, and 2,640 datasets were used as model evaluation data MED. The remaining 1,320 datasets were not used in the comparative verification according to the present disclosure. A series of processes from steps S51 to S59 in step S15 was performed to construct a prediction model M from the training data TD and the model evaluation data MED. K, which represents the number of epochs (total number of learning iterations), was set to 10,000, and the model Mi, which is an intermediate model being updated, was evaluated every 10 epochs using the model evaluation data MED.

[0135] In addition to the training dataset, a dataset (hereinafter referred to as "test data") was prepared to compare and verify the predicted values ​​by the prediction model M with the correct values. In preparing the test data, 10 levels of aggregate with different coarseness ratios were prepared according to the conditions shown in Table 4 below.

[0136] [Table 4]

[0137] Under the conditions shown in Table 4, the combination of mass fractions for multiple particle size classes is set to be different from the combinations under the conditions of Tables 1, 2, and 3. After preparing the aggregate in an air-dried state according to the conditions in Table 4, 100 image data (imaged image data P1) according to the conditions in Table 4 were obtained using a method similar to the preparation (imaging) according to the conditions in Table 1. However, the illuminance IB, one of the environmental conditions during imaging, was set to 3220 lx. The illuminances IA and IB were measured using a digital illuminance meter (product name: CANA-0010, manufactured by Tokyo Koden Co., Ltd.) placed on an imaging platform on which a tray containing the aggregate to be imaged was placed.

[0138] Using a method similar to the preparation according to the conditions in Table 1, the captured image data P1 was grayscaled to obtain 100 data sets as test data, each of which corresponded to the correct value of the mass fraction remaining between successive sieves. This test data was based on aggregates with a different gradation level than the aggregate gradation level used to prepare the training data TD. In addition to the correct value of the mass fraction remaining between successive sieves, the 100 data sets included in the test data also corresponded to the image data and the correct value of the coarse particle ratio for comparative verification.

[0139] Using the test data, comparative verification was performed for each dataset included in the test data using the following procedure. First, input image data P2 (input information) included in the dataset was input into prediction model M to obtain predicted values ​​of the mass fraction (mass fraction for each of multiple particle size divisions) retained between successive sieves. Next, the predicted values ​​of the mass fraction retained between successive sieves were converted to coarse particle fractions to obtain predicted values ​​of the coarse particle fraction using prediction model M. After that, the error between the predicted value of the coarse particle fraction by prediction model M and the correct value of the coarse particle fraction for the dataset was calculated.

[0140] In the comparative verification, the error between the predicted value of the coarse grain ratio and the correct value was evaluated using the following two indicators. Coarse-grained accuracy rate (%): The ratio of the number of correct datasets to the total number of datasets, assuming that a dataset with an error of within ±0.10 is defined as correct. ·RMSE: Root Mean Square Error.

[0141] (Verification example 1B: surface dry, illuminance IB=3220lx) While using the same prediction model M as in Verification Example 1A, the test data preparation method was different from that in Verification Example 1A. In Verification Example 1B, instead of air-dried aggregate, test data was prepared using aggregate that simply simulated surface-dried aggregate, which means that the internal voids are filled with water and the surface is dry. Specifically, the surface-dried aggregate was prepared by adding 0.25 wt% (10 ml) of water to 4 kg of air-dried aggregate and stirring the aggregate until the water covered the entire aggregate surface. In the explanation of this comparative verification, for convenience of explanation, the state that simply simulated the surface-dried state will simply be referred to as the "surface-dried state." In other words, surface-dried aggregate refers to aggregate that simply simulates surface-dried aggregate. The test data was prepared using the same method as in Verification Example 1A, except that surface-dried aggregate was used.

[0142] (Verification example 2A: air-dry condition, illuminance IB=3600lx) The same prediction model M as in Verification Example 1A was used, but the method for preparing test data was different from that in Verification Example 1A. In Verification Example 2A, test data was prepared with the illuminance IB set to 3600 lx. The test data was prepared in the same manner as in Verification Example 1A, except that the illuminance IB was set to 3600 lx.

[0143] (Verification example 2B: surface dry, illuminance IB=3600lx) The same prediction model M as in Verification Example 1A was used, but the method for preparing the test data was different from that in Verification Example 2A. In Verification Example 2B, test data was prepared using surface-dried aggregate instead of air-dried aggregate. The test data was prepared in the same manner as in Verification Example 2A, except that surface-dried aggregate was used.

[0144] (Verification example 3A: air-dry condition, illuminance IB=4190lx) The same prediction model M as in Verification Example 1A was used, but the method of preparing the test data was different from that in Verification Example 1A. In Verification Example 3A, the test data was prepared with the illuminance IB set to 4190 lx. The test data was prepared in the same manner as in Verification Example 1A, except that the illuminance IB was set to 4190 lx.

[0145] (Verification example 3B: surface dry, illuminance IB=4190lx) The same prediction model M as in Verification Example 1A was used, but the method for preparing the test data was different from that in Verification Example 3A. In Verification Example 3B, test data was prepared using surface-dried aggregate instead of air-dried aggregate. The test data was prepared in the same manner as in Verification Example 3A, except that surface-dried aggregate was used.

[0146] (Verification example 4A: air-dry condition, illuminance IB=4810lx) The same prediction model M as in Verification Example 1A was used, but the method for preparing test data was different from that in Verification Example 1A. In Verification Example 4A, test data was prepared with the illuminance IB set to 4810 lx. The test data was prepared in the same manner as in Verification Example 1A, except that the illuminance IB was set to 4810 lx.

[0147] (Verification example 4B: surface dry, illuminance IB=4810lx) The same prediction model M as in Verification Example 1A was used, but the method for preparing the test data was different from that in Verification Example 4A. In Verification Example 4B, test data was prepared using surface-dried aggregate instead of air-dried aggregate. The test data was prepared in the same manner as in Verification Example 4A, except that surface-dried aggregate was used.

[0148] (Reference example A: air-dry condition, illuminance IB=610lx) The same prediction model M as in Verification Example 1A was used, but the method of preparing the test data was different from that in Verification Example 1A. In Reference Example A, the test data was prepared with the illuminance IB set to 610 lx. The test data was prepared in the same manner as in Verification Example 1A, except that the illuminance IB was set to 610 lx.

[0149] (Reference example B: surface dry, illuminance IB=610lx) The same prediction model M as in Verification Example 1A was used, but the method of preparing the test data was different from that in Reference Example A. In Reference Example B, test data was prepared using surface-dried aggregate instead of air-dried aggregate. The test data was prepared in the same manner as Reference Example A, except that surface-dried aggregate was used.

[0150] (Difference between air-dried and surface-dried) The color of the aggregate in the image was different between the air-dry state and the surface-dry state. Specifically, the aggregate in the image appeared whitish in the air-dry state, whereas the aggregate in the image appeared blackish in the surface-dry state. That is, in Reference Example A and Verification Examples 1A to 4A, the color of the aggregate in the image was the same between the training phase and the evaluation phase, whereas in Reference Example B and Verification Examples 1B to 4B, the color of the aggregate in the image was different between the training phase and the evaluation phase.

[0151] (Verification results) Tables 5 and 6 below show the verification results regarding errors in each verification example and each reference example.

[0152] [Table 5]

[0153] [Table 6]

[0154] In Fig. 12(a), the evaluation results for accuracy rate are plotted on a graph, and in Fig. 12(b), the evaluation results for RMSE are plotted on a graph. In the graphs of Fig. 12(a) and Fig. 12(b), the circles and solid lines correspond to the results shown in Table 5, and the triangles and dashed lines correspond to the results shown in Table 6.

[0155] The results shown in Tables 5 and 6 and FIG. 12(a) indicate that the accuracy rates are higher in Verification Examples 1A to 4A and 1B to 4B, in which the illuminance Ia was 3220 lx, 3600 lx, 4190 lx, and 4810 lx, than in Reference Examples A and B, in which the illuminance Ib was 610 lx while the illuminance Ia was 3420 lx. Furthermore, the results shown in Tables 5 and 6 and FIG. 12(b) indicate that the RMSE is smaller in Verification Examples 1A to 4A and 1B to 4B, in which the illuminance Ia was 3220 lx, 3600 lx, 4190 lx, and 4810 lx, than in Reference Examples A and B, in which the illuminance Ib was 610 lx while the illuminance Ia was 3420 lx. Furthermore, it can be seen that the error index for the surface-dry condition can be significantly reduced in Verification Examples 1B to 4B compared to Reference Example B.

[0156] [Variations] The series of processes shown in each of Figures 7, 9, and 11 are examples and can be modified as appropriate. In the series of processes, one step and the next step may be executed in parallel, or several steps may be executed in an order different from that of the above-mentioned examples. Steps different from those of the above-mentioned examples may be executed instead of at least some of the steps of the series of processes, or in addition to the series of processes. In step S15, the model evaluation data MED may not be used, and the model at the time when training (weight update) for the set number of epochs is completed may be determined as the prediction model M.

[0157] In the example shown in FIG. 7 , after step S11 is performed, the grayscaling process of step S14 may be performed, followed by the generation of extended image data of step S13. Training input image data P2 may be generated by subjecting training captured image data P1 to a grayscaling process and a conversion process for extending the training data, in this order. The conversion process for extending the training data does not need to involve conversion including at least one of image inversion and rotation. Extension of training data involving the generation of new image data may be performed by a method different from the extension method exemplified in step S13. In preparing training data, extension of training data by the method of step S13 or another method does not need to be performed.

[0158] In step S15, the prediction model M may be configured to output a coarse particle ratio in response to input of input image data P2. In this case, in the evaluation phase, instead of steps S24 and S25, the coarse particle ratio output when the evaluation input image data P2 is input to the prediction model M is acquired as a prediction result. The prediction model M may be configured to output two or more physical property values ​​related to the aggregate particle size. For example, the prediction model M may be configured to output the mass fraction and coarse particle ratio remaining between successive sieves in response to input information including the input image data P2. In this case, the prediction unit 72 may calculate the coarse particle ratio based on the predicted value of the mass fraction from the prediction model M and the predicted value of the coarse particle ratio output from the prediction model M as quality prediction results. The input information, which is data input to the prediction model M, may include any information that may be correlated with aggregate quality in addition to the input image data P2.

[0159] In addition to the quality prediction device 50 predicting the quality of the aggregate, the manufacturing system 1 may periodically (e.g., several times a day) measure the physical properties of the aggregate. In this case, the prediction model M may be updated based on the actual measured values ​​of the physical properties of the aggregate and input information (e.g., input image data P2) when the actual measured values ​​were obtained. In the prediction step, the physical properties of the aggregate may be predicted using the updated prediction model M. Note that even when the updated prediction model M is used, the step of predicting the physical properties of the aggregate is still performed based on the prediction model M and the input information acquired in the second acquisition step.

[0160] The quality prediction device 50 may be used in a location other than the manufacturing system 1 (other than a process inspection in the manufacturing process of ready-mixed concrete). The quality prediction device 50 may be used in an acceptance inspection when a factory that manufactures ready-mixed concrete receives raw materials. In one example, the quality prediction device 50 is used for quality prediction (acceptance inspection) performed on a transport truck or a transport ship.

[0161] The imaging of the aggregate for evaluation by the imaging device 58 may be performed autonomously by the device including the quality prediction device 50 itself while checking the illuminance IB. In this case, the device including the quality prediction device 50 itself may determine, based on the illuminance IB, whether or not it is appropriate to perform prediction using the prediction model M, and may perform prediction using the prediction model M if the illuminance IB is appropriate. The imaging device 58 may obtain color image data P1 by imaging, and then perform a grayscale process on the image data P1 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 device 52. In one example of the various examples described above, at least some of the features described in the other examples may be combined.

[0162] Summary of this disclosure The aggregate quality prediction method described above includes a first acquisition step of acquiring color image data (P1) obtained by imaging the aggregate to be evaluated, a pre-processing step of generating input image data (P2) by grayscaling the image data (P1), a second acquisition step of acquiring input information including the input image data (P2), a prediction step of acquiring physical property values ​​corresponding to the input information acquired in the second acquisition step using a prediction model (M) previously constructed to output physical property values ​​of the aggregate in response to input of the input information, and a construction step of constructing the prediction model (M) by machine learning based on training data (TD) including training input information corresponding to the above input information and ground truth information for the physical property values ​​associated with the training input information. When the illuminance among the environmental conditions when imaging the training bone material to prepare the training input information is defined as IA [lx], and the illuminance among the environmental conditions when imaging the bone material to be evaluated to obtain the captured image data (P1) is defined as IB [lx], IB [lx] is 0.2 times or more than IA [lx] and 5 times or less than IA [lx].

[0163] A technology for predicting aggregate quality from aggregate image data using a prediction model constructed through machine learning has been investigated. Differences in imaging conditions or particle size distribution, which represent the size of aggregate particles, between the training and evaluation phases can lead to reduced prediction accuracy. After extensive investigation, the inventors came up with the idea of ​​preprocessing the image data to be input into the prediction model by grayscaling it. Further testing revealed a new problem: even with grayscaling preprocessing, prediction accuracy can be reduced if illuminance increases between the training and evaluation phases. In response to this problem, the quality prediction method uses an illuminance IB in the evaluation phase that is at least 0.2 times but no greater than 5 times the illuminance IA in the training phase. Therefore, the illuminance IB used to obtain image data with unknown physical properties in the evaluation phase is close to the illuminance IA used to prepare training data in the training phase. Therefore, assuming grayscaling is performed as a preprocessing step, the prediction accuracy of the prediction model M can be prevented from deteriorating due to differences in illuminance. Therefore, the above quality prediction method is useful for improving the prediction accuracy when predicting quality using a prediction model constructed by machine learning.

[0164] In the aggregate quality prediction method described above, the absolute value of the difference between IA [lx], which represents the illuminance among the environmental conditions when an image of a training aggregate is taken to prepare input information for training, and IB [lx], which represents the illuminance among the environmental conditions when an image of an aggregate to be evaluated is taken to obtain the captured image data (P1), may be 2500 lx or less. It has been confirmed that a decrease in prediction accuracy is suppressed when the difference between illuminance IB and illuminance IA is 2500 lx or less, compared to when the difference between illuminance IB and illuminance IA is greater than 2500. Therefore, the above quality prediction method is useful for improving the accuracy of predictions using a prediction model for predicting aggregate quality from image data.

[0165] In the aggregate quality prediction method described above, IB [lx], which represents the illuminance among the environmental conditions when capturing an image of the aggregate to be evaluated to obtain the captured image data (P1), may be 2000 lx or more. In this case, the illuminance IB is close to the range of general illuminance indoors on a sunny day. Therefore, it is easy to adjust the environmental conditions during the image capturing to a desired range in the evaluation phase.

[0166] In the aggregate quality prediction method described above, in the preprocessing step, when the pixel values ​​of the red, green, and blue color components of each pixel in the color image data are represented as R, G, and B, respectively, and the pixel value of each pixel in the grayscale image data is represented as Y, grayscale conversion may be performed using the above formula (1). In this case, the grayscale conversion is performed taking into account the sensitivity of the human eye. This increases the likelihood that the prediction model can capture and predict differences perceived by humans as features from the grayscaled image data.

[0167] The aggregate quality prediction method described above may further include a preparation step of preparing training data (TD). The preparation step may include generating extended image data by performing a conversion process and a grayscale process, in any order, on training captured image data (P1) corresponding to the captured image data (P1). The conversion process may include randomly changing the brightness and contrast of the image. The input image data (P2) in the training input information may include extended image data. The brightness and contrast of the captured image data may change depending on the surrounding environment at the time of capture. In this case, even if the correct values ​​of the physical properties are the same or even if the difference in illuminance between the training phase and the evaluation phase is reduced, the trained prediction model may recognize the image data as having different characteristics, resulting in reduced prediction accuracy. Therefore, it is possible to perform machine learning by preparing various training image data after adjusting the capturing environment, but this would complicate the preparation of the training data. In contrast, in the quality prediction method described above, the brightness and contrast of the image are randomly changed to generate new image data and extend the training data. Therefore, it is useful for improving prediction accuracy while avoiding the complication of preparing training data for machine learning.

[0168] In the aggregate quality prediction method described above, the transformation process may further include at least one of inverting and rotating the image. Even if the correct values ​​of the physical properties are the same and the difference in illuminance between the training phase and the evaluation phase is reduced, differences in the deposition state of the aggregate being imaged can cause the trained prediction model to recognize the image data as having different characteristics, resulting in reduced prediction accuracy. Therefore, while it is possible to perform machine learning by preparing multiple training image data with different deposition states, this complicates the preparation of the training data. In contrast, the quality prediction method described above involves at least one of inverting and rotating the image to generate new image data and expand the training data. Therefore, this method is even more useful for improving prediction accuracy while avoiding the complication of preparing training data for machine learning.

[0169] In the above-described aggregate quality prediction method, the construction step may include preparing model evaluation data (MED) separately from the training data (TD), the model evaluation data (MED) including input information for model evaluation corresponding to the input information and correct information on physical properties associated with the input information for model evaluation; continuously updating the intermediate models (Mi) using the training data (TD) so as to reduce the error between the training data and the correct information on physical properties; and using the model evaluation data (MED) to determine one of the intermediate models (Mi) generated during the update as the prediction model (M). During training in machine learning, parameters such as weights are continuously adjusted, and the intermediate models are continuously updated. In such machine learning, it is possible to use a model generated when a set number of learning cycles (number of epochs) has been completed as the prediction model to be used in the evaluation phase. However, the predictive accuracy of a model at the end of training is not necessarily higher than the predictive accuracy of a model that has been updated before. In contrast, in the above-described quality prediction method, intermediate models (Mi) generated during the update are evaluated using model evaluation data (MED) separate from the training data (TD), and one of the intermediate models (Mi) is then determined as the prediction model (M). As a result, even if an intermediate model generated during the update is preferable in terms of prediction accuracy, that intermediate model can be used as the prediction model (M) for prediction in the evaluation phase. This is therefore useful for improving the prediction accuracy of a prediction model for predicting aggregate quality.

[0170] The aggregate quality prediction program described above is a program for causing a computer to execute the quality prediction methods described in [1] to [7] above. Similar to the aggregate quality prediction methods, this quality prediction program is useful for improving the prediction accuracy when predicting quality using a prediction model constructed by machine learning.

[0171] The quality prediction device (50) described above includes a first acquisition unit (62) that acquires color image data (P1) obtained by capturing an image of the aggregate to be evaluated, a pre-processing unit (64) that generates input image data (P2) by grayscaling the captured image data (P1), a second acquisition unit (66) that acquires input information including the input image data (P2), a prediction unit (72) that acquires physical property values ​​corresponding to the input information acquired by the second acquisition unit (66) using a prediction model (M) that has been constructed in advance to output physical property values ​​of the aggregate in response to input of the input information, and a construction unit (68) that constructs the prediction model (M) by machine learning based on training data (TD) that includes training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information. When the illuminance among the environmental conditions when an image of a training aggregate is taken to prepare input information for training is defined as IA [lx], and the illuminance among the environmental conditions when an image of an aggregate to be evaluated is taken to obtain captured image data (P1) is defined as IB [lx], IB [lx] is 0.2 times or more than IA [lx] and 5 times or less than IA [lx]. This quality prediction device (50), like the above-described aggregate quality prediction method, is useful for improving prediction accuracy when predicting quality using a prediction model constructed by machine learning.

[0172] The above-described method for producing ready-mixed concrete includes a production step of kneading materials including 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 production step. The quality prediction step includes a first acquisition step of acquiring color image data (P1) obtained by imaging the aggregate to be evaluated, a preprocessing step of grayscaling the image data (P1) to generate input image data (P2), a second acquisition step of acquiring input information including the input image data (P2), a prediction step of acquiring physical property values ​​corresponding to the input information acquired in the second acquisition step using a prediction model (M) previously constructed to output physical property values ​​of the aggregate in response to input of the input information, and a construction step of constructing the prediction model (M) by machine learning based on training data (TD) including training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information. When the illuminance among the environmental conditions when imaging a training aggregate to prepare input information for training is defined as IA [lx], and the illuminance among the environmental conditions when imaging an aggregate to be evaluated to obtain image data (P1) is defined as IB [lx], IB [lx] is at least 0.2 times IA [lx] and at most 5 times IA [lx]. This method for producing ready-mixed concrete is useful for improving the accuracy of predictions when predicting quality using a prediction model constructed by machine learning, similar to the above-mentioned aggregate quality prediction method.

[0173] The manufacturing system (1) described above includes a manufacturing apparatus (10) that mixes materials including aggregate to manufacture ready-mixed concrete, and a quality prediction device (50) that predicts the quality of at least a portion of the aggregate used by the manufacturing apparatus (10). The quality prediction device (50) includes a first acquisition unit (62) that acquires color image data (P1) obtained by capturing an image of an aggregate to be evaluated, a pre-processing unit (64) that generates input image data (P2) by grayscaling the captured image data (P1), a second acquisition unit (66) that acquires input information including the input image data (P2), a prediction unit (72) that acquires physical property values ​​corresponding to the input information acquired by the second acquisition unit (66) using a prediction model (M) that has been previously constructed to output physical property values ​​of the aggregate in response to input of the input information, and a construction unit (68) that constructs the prediction model (M) by machine learning based on training data (TD) that includes training input information corresponding to the input information and ground truth information for the physical property values ​​associated with the training input information. When the illuminance among the environmental conditions when imaging a training aggregate to prepare input information for training is defined as IA [lx], and the illuminance among the environmental conditions when imaging an aggregate to be evaluated to obtain image data (P1) is defined as IB [lx], IB [lx] is 0.2 times or more than IA [lx] and 5 times or less than IA [lx]. This manufacturing system (1), like the above-mentioned aggregate quality prediction method, is useful for improving the prediction accuracy when predicting quality using a prediction model constructed by machine learning. [Explanation of symbols]

[0174] 1...manufacturing system, 10...manufacturing equipment, 50...quality prediction device, 62...imaging data acquisition unit, 64...preprocessing unit, 66...input data acquisition unit, 68...model construction unit, 72...prediction unit, P1...imaging image data, P2...input image data, TD...training data, MED...model evaluation data, M...prediction model, Mi...model (intermediate model).

Claims

1. a first acquisition step of acquiring color image data obtained by imaging the aggregate to be evaluated; a pre-processing step of grayscaling the captured image data to generate input image data; a second acquisition step of acquiring input information including the input image data; a prediction step of acquiring physical property values ​​corresponding to the input information acquired in the second acquisition step by using a prediction model previously constructed to output physical property values ​​of the aggregate in response to input of the input information; a construction step of constructing the prediction model by machine learning based on training data including training input information corresponding to the input information and correct answer information of the physical property value associated with the training input information, When the illuminance among the environmental conditions when the training bone material is imaged to prepare the training input information is defined as IA [lx], and the illuminance among the environmental conditions when the evaluation target bone material is imaged to obtain the image data is defined as IB [lx], IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx]. Methods for predicting aggregate quality.

2. The absolute value of the difference between IA [lx], which represents the illuminance among the environmental conditions when the training bone is imaged to prepare the training input information, and IB [lx], which represents the illuminance among the environmental conditions when the evaluation target bone is imaged to obtain the image data, is 2500 lx or less. The method for predicting aggregate quality according to claim 1.

3. IB [lx], which represents the illuminance among the environmental conditions when imaging the aggregate to be evaluated to obtain the image data, is 2000 lx or more. The method for predicting aggregate quality according to claim 1 or 2.

4. In the pre-processing step, when the pixel values ​​of the red, green, and blue color components of each pixel in the color image data are represented as R, G, and B, and the pixel value of each pixel in the grayscale image data is represented as Y, grayscale conversion is performed using the following equation (1): Y=0.299×R+0.587×G+0.114×B...(1) The method for predicting aggregate quality according to claim 1 or 2.

5. further comprising the step of preparing the training data; The preparation step includes generating extended image data by performing a conversion process and a grayscale process in any order on training captured image data corresponding to the captured image data, the transformation process includes randomly changing the brightness and contrast of the image; The input image data in the training input information includes the augmented image data. The method for predicting aggregate quality according to claim 1 or 2.

6. The transformation process further includes at least one of flipping the image and rotating the image. The method for predicting aggregate quality according to claim 5.

7. The construction step includes: preparing model evaluation data including input information for model evaluation corresponding to the input information and correct answer information of the physical property value associated with the input information for model evaluation, separately from the training data; continuously updating the intermediate model using the training data so that an error between the intermediate model and the correct information of the physical property value becomes small; and determining, using the model evaluation data, one of the intermediate models generated during the update as the predictive model. The method for predicting aggregate quality according to claim 1 or 2.

8. A quality prediction program for causing a computer to execute the aggregate quality prediction method according to claim 1 or 2.

9. a first acquisition unit that acquires color image data obtained by capturing an image of the aggregate to be evaluated; a pre-processing unit that generates input image data by grayscaling the captured image data; a second acquisition unit that acquires input information including the input image data; a prediction unit that acquires the physical property values ​​according to the input information acquired by the second acquisition unit using a prediction model that is pre-constructed to output the physical property values ​​of the aggregate in response to the input of the input information; a construction unit that constructs the prediction model by machine learning based on training data including training input information corresponding to the input information and correct answer information for the physical property value associated with the training input information; When the illuminance among the environmental conditions when the training bone material is imaged to prepare the training input information is defined as IA [lx], and the illuminance among the environmental conditions when the evaluation target bone material is imaged to obtain the image data is defined as IB [lx], IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx]. Aggregate quality prediction device.

10. A manufacturing process in which materials including aggregate are mixed to produce ready-mix concrete; a quality prediction step of predicting the quality of at least a portion of the aggregate used in the manufacturing step; Including, The quality prediction step includes: a first acquisition step of acquiring color image data obtained by imaging the aggregate to be evaluated; a pre-processing step of grayscaling the captured image data to generate input image data; a second acquisition step of acquiring input information including the input image data; a prediction step of acquiring physical property values ​​corresponding to the input information acquired in the second acquisition step by using a prediction model previously constructed to output physical property values ​​of the aggregate in response to input of the input information; a construction step of constructing the prediction model by machine learning based on training data including training input information corresponding to the input information and correct answer information of the physical property value associated with the training input information, When the illuminance among the environmental conditions when the training bone material is imaged to prepare the training input information is defined as IA [lx], and the illuminance among the environmental conditions when the evaluation target bone material is imaged to obtain the image data is defined as IB [lx], IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx]. Method for manufacturing ready-mix concrete.

11. A manufacturing device that mixes materials including aggregate to manufacture ready-mix concrete; a quality prediction device that predicts the quality of at least a portion of the aggregate used by the manufacturing device; The quality prediction device includes: a first acquisition unit that acquires color image data obtained by capturing an image of the aggregate to be evaluated; a pre-processing unit that generates input image data by grayscaling the captured image data; a second acquisition unit that acquires input information including the input image data; a prediction unit that acquires the physical property values ​​according to the input information acquired by the second acquisition unit using a prediction model that is pre-constructed to output the physical property values ​​of the aggregate in response to the input of the input information; a construction unit that constructs the prediction model by machine learning based on training data including training input information corresponding to the input information and correct answer information for the physical property value associated with the training input information; When the illuminance among the environmental conditions when the training bone material is imaged to prepare the training input information is defined as IA [lx], and the illuminance among the environmental conditions when the evaluation target bone material is imaged to obtain the image data is defined as IB [lx], IB [lx] is 0.2 times or more of IA [lx] and 5 times or less of IA [lx]. Ready-mix concrete manufacturing system.

Images