JP2024079337A - Aggregate quality prediction method, ready-mixed concrete manufacturing method, and ready-mixed concrete manufacturing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To stabilize quality of ready-mixed concrete.

SOLUTION: An aggregate quality prediction method includes: a preparation step of, through machine learning based on input information including image data obtained by picking up an image of aggregate and a correct answer value of a fineness modulus of the aggregate associated with the input information, building a prediction model that outputs a value indicating the fineness modulus of the aggregate according to input of the input information; an acquisition step of acquiring evaluation information including image data obtained by picking up an image of aggregate to be evaluated and corresponding to the input information; and a prediction step of inputting the evaluation information to the prediction model to acquire a first prediction value to be output from the prediction model in relation to the fineness modulus of the aggregate.

SELECTED DRAWING: Figure 3

COPYRIGHT: (C)2024,JPO&INPIT

Description

This disclosure relates to a method for predicting aggregate quality, a method for producing ready-mix concrete, and a system for producing ready-mix concrete.

Patent Document 1 discloses a method for estimating the particle size distribution of soil. Patent Document 2 discloses a method for estimating the quality of aggregate.

JP 2021-117625 A JP 2021-135199 A

This disclosure provides a method for predicting aggregate quality, a method for producing ready-mix concrete, and a system for producing ready-mix concrete that are useful for stabilizing the quality of ready-mix concrete.

[1] A method for predicting quality of aggregate, comprising: a preparation step of constructing a prediction model that outputs a value indicating the coarse aggregate ratio in response to input of input information, the input information including image data obtained by photographing the aggregate, and machine learning based on a ground truth value of the coarse aggregate ratio corresponding to the input information; an acquisition step of acquiring evaluation information that includes image data obtained by photographing the aggregate to be evaluated and corresponds to the input information; and a prediction step of inputting the evaluation information into the prediction model and acquiring a first predicted value output from the prediction model regarding the coarse aggregate ratio.
[2] The quality prediction method for aggregate described in [1] above, wherein in the preparation step, the prediction model is constructed to output only a value indicating the coarse aggregate ratio in accordance with the input information.
[3] The quality prediction method of aggregate described in [1] above, wherein the preparation step includes constructing the prediction model by machine learning further based on correct values of the mass ratio of each of a plurality of granularity classifications of the aggregate corresponding to the input information, and in the preparation step, the prediction model is constructed to further output a value indicating the mass ratio of each classification included in the plurality of granularity classifications in accordance with the input information, and the prediction step includes inputting the evaluation information into the prediction model to further obtain a second predicted value output from the prediction model regarding the mass ratio of each classification, and converting the second predicted value into a coarse aggregate ratio.
[4] The quality prediction method for aggregate described in [3] above, wherein the preparation process includes updating an intermediate model generated in the process of constructing the predictive model based on a first error related to the coarse aggregate ratio and a second error related to the mass ratio for each category, and in the preparation process, the intermediate model is updated using an error corrected so as to reduce the difference between the first error and the second error.
[5] The method for predicting quality of aggregate described in any one of [1] to [4] above, wherein the preparation step includes constructing the prediction model by machine learning using a convolutional neural network.
[6] A method for predicting quality of aggregate described in any one of the above [1] to [5], wherein the acquisition step includes imaging an area including the aggregate to be evaluated from a position away from the aggregate to be evaluated while the aggregate to be evaluated is piled up, thereby acquiring the evaluation information.
[7] A method for predicting quality of aggregate described in any one of [1] to [6] above, wherein the preparation step includes preparing training aggregate so that it contains grains of two or more grain size divisions, and the input information includes image data obtained by photographing the training aggregate.
[8] A method for producing ready-mixed concrete, comprising: a production process for mixing materials including aggregate to produce ready-mix concrete; and a quality prediction process for predicting the coarse aggregate ratio of at least a portion of the aggregate used in the production process, wherein the quality prediction process comprises: a preparation process for constructing a prediction model that outputs a value indicating the coarse aggregate ratio in response to input of input information by machine learning based on input information including image data obtained by photographing the aggregate and a correct value of the coarse aggregate ratio corresponding to the input information; an acquisition process for acquiring evaluation information corresponding to the input information, including image data obtained by photographing the aggregate to be evaluated; and a prediction process for inputting the evaluation information into the prediction model and acquiring a first predicted value output from the prediction model regarding the coarse aggregate ratio.
[9] A ready-mix concrete manufacturing system comprising: a manufacturing apparatus for mixing materials including aggregate to manufacture ready-mix concrete; and a quality prediction device for predicting the coarse aggregate ratio of at least a portion of the aggregate used by the manufacturing apparatus, wherein the quality prediction device is configured to execute the following steps: a preparation step for constructing a prediction model that outputs a value indicating the coarse aggregate ratio in response to input of input information by machine learning based on input information including image data obtained by photographing the aggregate and a ground truth value of the coarse aggregate ratio corresponding to the input information; an acquisition step for acquiring evaluation information corresponding to the input information, the evaluation information including image data obtained by photographing the aggregate to be evaluated; and a prediction step for inputting the evaluation information into the prediction model and acquiring a first predicted value output from the prediction model regarding the coarse aggregate ratio.

The present disclosure provides an aggregate quality prediction method, a ready-mix concrete manufacturing method, and a ready-mix concrete manufacturing system that are useful for stabilizing the quality of ready-mix concrete.

FIG. 1 is a schematic diagram showing an example of a ready-mix concrete manufacturing system. FIG. 2 is a schematic diagram illustrating an example of a hardware configuration of the control device. 3A and 3B are flow charts showing an example of a process executed in the learning phase and an evaluation phase, respectively. Fig. 4A is a schematic diagram showing an example of a predictive calculation using a predictive model, and Fig. 4B is a schematic cross-sectional view showing an example of a state in which image data for learning is acquired. FIG. 5 is a diagram showing an example of image data for learning. 6A and 6B are graphs showing an example of a comparison result between a predicted value and a correct value. 7A and 7B are graphs showing an example of a comparison result between a predicted value and a correct value. 8A and 8B are graphs showing an example of a comparison result between a predicted value and a correct value. 9A and 9B are graphs showing an example of the comparison result between the predicted value and the correct value. 10A and 10B are graphs showing an example of a comparison result between a predicted value and a correct value. 11A and 11B are graphs showing an example of the comparison result between the predicted value and the correct value. FIG. 12 is a graph showing an example of a comparison result between a predicted value and a correct value. FIG. 13 is a graph showing an example of a comparison result between a predicted value and a correct value. FIG. 14 is a graph showing an example of a comparison result between a predicted value and a correct value.

One embodiment will be described below with reference to the drawings. In the description, identical elements or elements having the same functions are given the same reference numerals, and duplicate descriptions will be omitted.

Figure 1 shows a schematic diagram of a ready-mix concrete manufacturing system according to one embodiment. The manufacturing system 1 shown in Figure 1 is a system for manufacturing ready-mix concrete. The manufacturing system 1 mixes concrete materials to manufacture (produce) ready-mix concrete. In addition to the function of manufacturing ready-mix concrete, the manufacturing system 1 has a function of predicting the quality of the concrete materials.

The concrete materials used by the manufacturing system 1 include cement, admixtures, coarse aggregates, fine aggregates, water, and admixtures. Examples of coarse aggregates include gravel such as mountain gravel, land gravel, river gravel, and sea gravel; crushed stone; slag coarse aggregates such as blast furnace slag aggregate, ferronickel slag aggregate, electric furnace oxidized slag aggregate, and coal gasification slag aggregate; lightweight coarse aggregates such as natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate; recycled coarse aggregate; recovered aggregate, or coarse aggregates made by mixing these. The coarse aggregate may include, for example, crushed rock, or limestone crushed stone. Examples of fine aggregates include sand such as mountain sand, land sand, river sand, and sea sand; crushed sand; slag fine aggregates such as blast furnace slag aggregate, ferronickel slag aggregate, copper slag aggregate, electric furnace oxidized slag aggregate, and coal gasification slag aggregate; lightweight fine aggregates such as natural lightweight aggregate, by-product lightweight aggregate, and artificial lightweight aggregate; recycled fine aggregate; and fine aggregates made of recycled aggregates or mixtures thereof. Rock types of crushed stone and crushed sand include igneous rocks such as granite, diorite, gabbro, porphyrite, diabase, rhyolite, andesite, basalt, and serpentine; sedimentary rocks such as conglomerate, sandstone, shale, slate, and tuff; metamorphic rocks such as gneiss and crystalline schist; and others, silica stone, limestone, domalite, and peridotite. In the present disclosure, coarse aggregates and fine aggregates may be collectively referred to as "aggregates". In this case, "aggregate" means coarse aggregate, fine aggregate, or both coarse and fine aggregate.

The manufacturing system 1 loads the manufactured ready-mix concrete onto a transport vehicle C. After the ready-mix concrete is loaded onto the transport vehicle C, the transport vehicle C transports the ready-mix concrete to a site (e.g., a construction site) where the ready-mix concrete is to be used. Examples of the transport vehicle C include an agitator vehicle (mixer vehicle) and a dump truck. The manufacturing system 1 may manufacture ready-mix concrete from concrete materials so as to satisfy a target quality (required quality) set for each site. In one example, the target quality of the ready-mix concrete includes a target value for slump or slump flow. The manufacturing system 1 includes, for example, a material storage area 2, a transport device 8, a manufacturing device 10, an imaging device 50, and a control device 60.

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 contains cement.

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 width of the belt conveyor may be approximately 30 cm to 160 cm. The transporting device 8 may transport the concrete materials by material type at different times. In one example, specific materials from among the various concrete materials are transferred to the transporting device 8 based on operational instructions from the control device 60, and transported to the manufacturing device 10.

The manufacturing device 10 is a device that mixes concrete materials including aggregate to manufacture ready-mix concrete. The manufacturing device 10 operates based on operation instructions from the control device 60. The manufacturing device 10 includes, for example, a storage bottle 12, a measuring bottle 14, a collection hopper 16, a mixer 20, and a loading hopper 30.

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

The measuring bottle 14 is disposed below the storage bottle 12. The measuring bottle 14 operates based on operational instructions from the control device 60, and weighs each material individually. When the measuring bottle 14 detects the target amount of material instructed by the control device 60, 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 various materials to the mixer 20. When the collecting hopper 16 is not provided, the various materials are supplied to the mixer 20 from the measuring bottle 14.

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.

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

The imaging device 50 is a device (camera) capable of capturing an image of aggregate, which is one of the concrete materials. The imaging device 50 generates image data by capturing an image of the aggregate. The imaging device 50 is, for example, a digital camera or video camera that captures an image of the imaging range based on visible light. The image data generated by the imaging device 50 is used to predict the quality of the aggregate. The prediction of the quality of the aggregate will be described later. Note that the imaging device 50 may generate a video of the subject to be captured, and still images included in the video may be acquired as image data.

The imaging device 50 is arranged to image an area including the aggregate to be imaged from a position away from the aggregate to be imaged. The imaging device 50 may image an area including the aggregate to be imaged from above. The imaging device 50 may image an area including the aggregate in a piled state. The piled state refers to aggregate in a state where some grains included in the aggregate are overlapped with other grains when viewed vertically from above. The imaging device 50 may be arranged so as to be able to image the aggregate being transported by the transport device 8. The imaging device 50 may image the aggregate being transported while the belt conveyor included in the transport device 8 is stopped or moving. The imaging device 50 may image the aggregate loaded on a transport truck or transport ship when the aggregate is received. The imaging device 50 may be arranged in an imaging room where a floor or a rubber sheet is prepared, and may image the aggregate placed on the floor or the like.

The control device 60 is a device that controls the manufacturing device 10. The control device 60 is composed of one or more computers. When the control device 60 is composed of multiple computers, these computers are connected so that they can communicate with each other. The control device 60 controls the manufacturing device 10 according to set operating conditions. At least a part of the operating conditions may be determined by instructions from an operator such as a worker.

An input/output device 62 may be connected to the control device 60 (see FIG. 2). The input/output device 62 is a device for inputting information indicating instructions from a worker or the like to the control device 60 and outputting information from the control device 60 to the worker or the like. The input/output device 62 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 62 may be a touch panel in which the input device and the output device are integrated. The control device 60 and the input/output device 62 may be integrated.

It is known that the quality of the ready-mix concrete produced by the production apparatus 10 is affected by the quality of the concrete materials. Specifically, the particle size of the aggregate used in producing the ready-mix concrete affects the fluidity (e.g., slump and slump flow) of the ready-mix 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 amount of water retention, and the greater the fluidity of the ready-mix concrete. Conversely, the smaller the particle size of the aggregate, the larger the surface area of the aggregate, which increases the amount of water retention, and the less fluidity of the ready-mix concrete. In this way, in order to stabilize the quality of the ready-mix concrete, including the fluidity, it is necessary to manage the particle size of the aggregate. The particle size of the aggregate can be expressed by the coarse particle ratio, and in order to stabilize the quality of the ready-mix concrete, it is necessary to manage, for example, whether the coarse particle ratio of the aggregate is within a desired range.

In addition to controlling the manufacturing apparatus 10, the control device 60 may be configured to predict the quality of the aggregate used in the manufacturing of ready-mix concrete by the manufacturing apparatus 10. In this case, the control device 60 constitutes a quality prediction device that predicts the quality of the aggregate. The control device 60 functioning as a quality prediction device predicts the coarse aggregate ratio of at least a portion of the aggregate used by the manufacturing apparatus 10.

The control device 60 is configured to execute a preparation step of constructing a prediction model that outputs a value indicating the aggregate coarse aggregate ratio in response to input information by machine learning based on at least input information including image data obtained by imaging the aggregate and a ground truth value of the aggregate coarse aggregate ratio associated with the input information. The control device 60 is also configured to execute an acquisition step of acquiring evaluation information that includes image data obtained by imaging the aggregate to be evaluated and corresponds to the input information, and a prediction step of inputting the evaluation information into the prediction model and acquiring a value output from the prediction model regarding the aggregate coarse aggregate ratio.

The control device 60 has, as its functional configuration (hereinafter referred to as "functional modules"), for example, an operation control unit 68, a model construction unit 72, a model holding unit 74, a quality prediction unit 76, and an output unit 78. The processes executed by these functional modules correspond to the processes executed by the control device 60. The operation control unit 68 is a functional module that controls the manufacturing device 10 to manufacture ready-mix concrete according to the above operating conditions. Workers, etc. may adjust (change) the operating conditions according to the target quality of the ready-mix concrete.

The model construction unit 72 is a functional module that constructs a model for predicting the coarse aggregate ratio (hereinafter referred to as the "prediction model M"). The prediction model M is a model that outputs a value indicating at least the coarse aggregate ratio in response to input information including image data obtained by photographing the aggregate. The model construction unit 72 constructs the prediction model M by machine learning based on the input information and the correct value of the coarse aggregate ratio associated with the input information.

Machine learning is a technique in which a machine (computer) autonomously finds laws or rules by repeatedly learning based on given information. The predictive model M can be constructed using an algorithm and a data structure. The 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. There are no particular limitations on the specific algorithm of the machine learning used when constructing the predictive model M. The 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, improving prediction accuracy.

The model construction unit 72 may construct the prediction model M by machine learning using a convolutional neural network (CNN). The model construction unit 72 may construct the prediction model M by machine learning based on the correct value of the coarse aggregate ratio and the correct value of the mass ratio of each of the multiple granularity classifications of the aggregate. The prediction model M constructed by the model construction unit 72 may output a value indicating the coarse aggregate ratio and a value indicating the mass ratio of each classification included in the multiple granularity classifications according to the above input information. In other words, the prediction model M may output one value indicating the coarse aggregate ratio and multiple values indicating the mass ratio for each of the multiple granularity classifications.

The model construction unit 72 may autonomously construct a prediction model M for predicting the coarse aggregate ratio by performing machine learning using, for example, data provided as input for machine learning and correct answer data (correct answer values such as the coarse aggregate ratio) of the machine learning output. The machine learning input is various data sets of input information including images of aggregate. At least the image data included in the input information differs from one another in the various data sets of input information. The machine learning output is data (numerical values) indicating the coarse aggregate ratio of the aggregate and the mass ratio for each gradation class. The model construction unit 72 iteratively learns a model that outputs predicted values of the coarse aggregate ratio of the aggregate and the mass ratio for each gradation class using multiple combinations of the data sets of input information and correct answer values such as the coarse aggregate ratio.

The stage of autonomously constructing the predictive model M corresponds to a learning phase. The learning phase may be performed before the production phase in which ready-mix concrete is manufactured, or may be performed at an early stage of the production phase. The predictive model M, which is a trained model, may be portable between computers. Therefore, the predictive model M constructed in the control device 60 may be used in another manufacturing system different from the manufacturing system 1.

The model holding unit 74 is a functional module that holds (stores) the prediction model M. The quality prediction unit 76 is a functional module that predicts the coarse aggregate ratio. Hereinafter, the target aggregate for which the coarse aggregate ratio is predicted is referred to as the "aggregate to be evaluated." The quality prediction unit 76 acquires information that includes image data obtained by photographing the aggregate to be evaluated and that corresponds to the input information (hereinafter referred to as "evaluation information"). While the input information corresponds to the correct value of the coarse aggregate ratio, the evaluation information is information in which the coarse aggregate ratio is unknown. The stage in which the quality prediction unit 76 predicts the coarse aggregate ratio using the prediction model M corresponds to the evaluation phase.

The quality prediction unit 76 inputs the evaluation information into the prediction model M and obtains a predicted value (first predicted value) output from the prediction model M regarding the coarse aggregate ratio. The quality prediction unit 76 also inputs the above evaluation information into the prediction model M and further obtains a predicted value (second predicted value) output from the prediction model M regarding the mass ratio for each grading class. That is, the quality prediction unit 76 obtains from the prediction model M a plurality of predicted values (a plurality of second predicted values) each indicating a mass ratio value for a plurality of grading classes. The quality prediction unit 76 may convert the predicted value (a plurality of predicted values) regarding the mass ratio for each grading class into a coarse aggregate ratio. By predicting the coarse aggregate ratio using the prediction model M using two different methods, the reliability of the predicted value of the coarse aggregate ratio can be evaluated.

The output unit 78 is a functional module that outputs at least the predicted value of the coarse aggregate ratio output from the prediction model M as the predicted result of the coarse aggregate ratio. The output unit 78 may further output a value obtained by converting the predicted value of the mass ratio for each gradation class output from the prediction model M into a coarse aggregate ratio as the predicted result of the coarse aggregate ratio. The output unit 78 may notify a worker or the like of the predicted result of the coarse aggregate ratio by outputting the predicted result of the coarse aggregate ratio to an output device (e.g., a monitor) of the input/output device 62.

As shown in FIG. 2, the control device 60 has a circuit 91. The circuit 91 includes one or more processors 92, a memory 93, a storage 94, an input/output port 95, and a timer 96. The storage 94 has a computer-readable storage medium, such as a non-volatile semiconductor memory. The storage 94 stores a program for causing the control device 60 to control the manufacturing device 10 using a preset control procedure, and a program for causing the control device 60 to predict the coarse aggregate fraction using a preset prediction procedure. For example, the storage 94 stores a program for configuring each of the functional modules described above.

The memory 93 temporarily stores the programs loaded from the storage medium of the storage 94 and the results of calculations by the processor 92. The processor 92 executes the programs in cooperation with the memory 93 to configure each functional module of the control device 60. The input/output port 95 inputs and outputs electrical signals between the manufacturing device 10, the imaging device 50, the input/output device 62, etc., in accordance with instructions from the processor 92. The timer 96 measures the elapsed time, for example, by counting a reference pulse at a constant period. Note that the circuit 91 is not necessarily limited to one that configures each function by a program. For example, at least some of the functions of the circuit 91 may be configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates the dedicated logic circuit.

[Method of manufacturing ready-mix concrete]
Next, a description will be given of an example of a method for producing ready-mixed concrete executed in the production system 1. The method for producing ready-mixed concrete includes a step of producing ready-mixed concrete and a step of predicting the quality of aggregate. The step of predicting the quality of aggregate may be executed during a period overlapping at least a part of the period during which the step of producing ready-mixed concrete is executed.

The process of producing ready-mix concrete (hereinafter referred to as the "production process") includes, for example, a transport process, a weighing process, a pouring process, a mixing process, a discharging process, and a loading process. In the transport process, various concrete materials are transported by a transport device 8 to a storage bottle 12, and the various materials are individually supplied to the storage bottle 12. In the weighing process, the various materials are individually supplied from the storage bottle 12 to a weighing bottle 14, and the various materials are weighed in the weighing bottle 14. In the weighing process, when the measured amount for each material reaches a predetermined set amount, the material is discharged into a collection hopper 16.

In the charging process, after all types of materials are collected in the collecting hopper 16, the materials in the collecting hopper 16 are charged (supplied) to 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, the ready-mixed concrete is discharged from the mixer 20 to 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.

The process of predicting aggregate quality (hereinafter referred to as the "quality prediction process") includes a model construction process in the learning phase and a quality evaluation process in the evaluation phase. In the quality prediction process (quality prediction method), the model construction process is executed before the quality evaluation process. An example of the model construction process and an example of the quality evaluation process are described below. Note that an example will be described in which the coarse aggregate fraction is predicted as the aggregate quality, and the input information includes only image data of the aggregate. Also, a specific example will be described in which the prediction model M outputs a predicted value of the coarse aggregate fraction and a predicted value of the mass ratio for each grading class.

(Model building process)
3A is a flowchart showing an example of a series of processes executed in the model construction step. This model construction step may be executed before the above-mentioned manufacturing step is executed in the manufacturing apparatus 10, or at an early stage when the above-mentioned manufacturing step is started. In this model construction step, a part of the coarse aggregate used in manufacturing ready-mixed concrete in the manufacturing apparatus 10 may be extracted, and the part of the coarse aggregate may be used as a "learning coarse aggregate".

In the model construction process, step S11 is executed first. In step S11, for example, a worker or the like prepares learning data for machine learning. The learning data includes various image data (input information) obtained by imaging the learning coarse aggregate, and correct answer data corresponding to each of the various image data. Figure 4(a) shows a schematic diagram of a calculation method for predicting the coarse aggregate ratio using a prediction model M from an image Im, which is the input information.

In this disclosure, as shown in FIG. 4(a), the prediction model M outputting two types of information, the mass ratio for each particle size division and the coarse particle ratio, is referred to as "multitasking." In addition, the prediction model M outputting either the mass ratio for each particle size division or the coarse particle ratio is referred to as "single tasking." FIG. 4(a) illustrates an example of a calculation method in which the coarse particle ratio is predicted by two types of methods using the prediction model M that performs multitasking.

In preparing learning images, for example, various types of coarse aggregates with known mass ratios and coarse aggregate ratios for each of multiple gradation classes are prepared by workers, etc. Various types of coarse aggregates do not mean different varieties of coarse aggregates (e.g., crushed stone), but rather coarse aggregates with different gradations. Then, as shown in FIG. 4(b), for example, the various types of learning coarse aggregates are individually imaged by camera 52 while stacked on measurement sheet 8a. This allows learning image data to be obtained for each type of coarse aggregate. Camera 52 may be a digital camera, and is set up to image an area including the coarse aggregates from a predetermined position away from the coarse aggregates on measurement sheet 8a (e.g., a predetermined position above).

The sheet 8a may be a sheet formed to resemble the belt conveyor of the transport device 8, for example a rubber sheet. The width of the sheet 8a may be approximately half the width of the belt conveyor of the transport device 8. The amount of aggregate placed on the sheet 8a is not particularly limited, but may be approximately 1 kg to 10 kg. Figure 4(b) shows a schematic example of aggregate piled up on the sheet 8a, with piled up states shown when the amount of aggregate is 2 kg, 5 kg, and 8 kg.

In one example, various types of coarse aggregate are prepared under the conditions shown in Table 1 below, and images are obtained by photographing the coarse aggregate. The mass ratios and coarse aggregate ratios for each grain size category of the various types of coarse aggregate prepared under the conditions shown in Table 1 are known, and these known values are the correct answer data for machine learning.

The coarse-grained ratio (F.M.) is calculated, for example, by the following formula (1) specified in the Standard Method for Concrete. In the measurement specified in the "Standard Specifications for Building Construction and Commentary JASS5 Reinforced Concrete Construction", nine sieves excluding the 80 mm sieve in formula (1) are used. The coarse-grained ratio predicted in the present disclosure may be specified by any method.

In Table 1, this means that 10 levels of grain size (coarse grain ratio) of coarse aggregate are prepared. Category I represents a grain size category of grains smaller than 2.5 mm. Grains that pass through a 2.5 mm sieve are classified into category I. Category II represents a grain size category of grains 2.5 mm or more and smaller than 5 mm. Grains that are retained on a 2.5 mm sieve but pass through a 5 mm sieve are classified into category II. Category III represents a grain size category of grains 5 mm or more and smaller than 10 mm. Grains that are retained on a 5 mm sieve but pass through a 10 mm sieve are classified into category III. Category IV represents a grain size category of grains 10 mm or more and smaller than 20 mm. Grains that are retained on a 10 mm sieve but pass through a 20 mm sieve are classified into category IV.

Regarding level 1, the coarse aggregate ratio can be calculated from the mass ratios of each gradation class (mass ratios of each of classes I to IV) as follows. For level 1 coarse aggregate, the percentages of the mass of the sample that are retained on the 80 mm, 40 mm, and 20 mm sieves are each 0, the percentages of the mass of the sample that are retained on the 10 mm sieve are 85, and the percentages of the mass of the sample that are retained on the 5 mm sieve are 100 (=15+85). In addition, since all the particles are retained on the 5 mm sieve, the percentages of the mass of the sample that are retained on the 2.5 mm, 1.2 mm, 600 μm, 300 μm, and 150 μm sieves are each 100. Therefore, the coarse aggregate ratio is calculated as shown in the following formula (2).

For levels 2 to 10, the coarse aggregate ratio is calculated in the same manner as for level 1. The conditions for the learning aggregate shown in Table 1 are merely examples, and the number of levels, the coarse aggregate ratio and mass ratio for each size category at each level, and the method for setting the size categories may be determined arbitrarily by an operator or the like. As with the multiple size categories shown in Table 1, the sizes in multiple categories including the smallest category (<150 μm) may be combined into one size category. The size above a certain category (e.g., 20 to 40 mm) may not be included in the multiple size categories. Also, although different from the setting method in Table 1, the size below a certain category (e.g., 1.2 to 2.5 μm) may not be included in the multiple size categories.

Figure 5 shows images T1 to T10 obtained by imaging training aggregate prepared according to the conditions of each of levels 1 to 10 in Table 1. Level n (n is an integer from 1 to 10) corresponds to image Tn. For example, image T1 is obtained by imaging training aggregate prepared according to level 1. Training aggregate according to each level n contains grains of multiple grain size divisions. Thus, in step S11, training aggregate may be prepared so as to contain grains of two or more grain size divisions, and input information may be prepared that includes image data obtained by imaging the training aggregate.

For each level n, multiple images Tn may be prepared in which the state (more specifically, the pile state) of the learning aggregate is different from one another. In one example, a first image is taken with the learning aggregate placed on the measurement sheet 8a, and one image Tn is obtained. Then, the learning aggregate is temporarily removed from the sheet 8a, and the same learning aggregate is placed again on the sheet 8a so that the pile state of the multiple particles is different from that at the time of the first image. Then, a second image is taken. Multiple images Tn may be prepared for each level n by repeating the above-mentioned image taking by a worker or the like.

The worker or the like may input the prepared image data and correct answer data to the control device 60 via the input/output device 62. One data set is composed of one image Tn and the correct answer data of the mass ratio and coarse grain ratio at level n associated with that image Tn. The worker or the like prepares multiple data sets as learning data.

Returning to FIG. 3(a), after step S11 is executed, step S12 is executed. In step S12, for example, the model construction unit 72 of the control device 60 constructs a prediction model M by performing machine learning using the image data prepared in step S11 and the ground truth data linked to the image data. The prediction model M may be constructed to output four mass ratio values related to the four grain size classifications and one value related to the coarse grain ratio (see also FIG. 4(a)). The four mass ratio values output from the prediction model M are converted as shown in the above formula (2) to obtain a predicted value of the coarse grain ratio.

In this way, the prediction model M may be a model that performs multitasking. When the prediction model M is a model that performs multitasking, two predicted values of the coarse grain ratio are obtained by different methods. In this case, for example, the reliability of the predicted value of the coarse grain ratio can be evaluated. In one example, when the difference between the two predicted values for the coarse grain ratio is small, the reliability of the prediction result can be evaluated as high.

An example of the prediction model M will be described below using simplified formulas for ease of understanding. The prediction model M constructed by the model construction unit 72 can be simply expressed as, for example, the following formulas (3) and (4).

In equation (4), Y represents one of y1 to y5, and in the prediction model M, equation (4) is calculated for each of y1 to y5. y1 to y5 are output data representing the mass ratio and coarse grain ratio for each granularity classification. If y1, y2, y3, and y4 are the output values of the mass ratios of each of the four granularity classifications, y5 is the output value of the coarse grain ratio. N is an integer equal to or greater than 2 and represents the number of input data. The input data includes at least the values of each pixel in the image of the aggregate. wi is a weight (coefficient), and b is a bias term (coefficient). f(U) represents the activation function. Activation functions are linear functions (identity functions) or nonlinear functions such as polynomials, absolute values, step functions, sigmoid functions, hardsigmoid functions, logsigmoid functions, softmax functions, logsoftmax functions, softmin functions, softplus functions, softsign functions, tanh functions, tanhShrink functions, hardtanh functions, tanhexp functions, ReLU functions, ReLU6 functions, Leaky-ReLU functions, PReLU functions, ELU functions, SELU functions, CELU functions, Swith functions, Mish functions, and ACON functions.

The model construction unit 72 uses the learning data to construct a prediction model M that outputs values of y1 to y5 by determining the weight wi and bias term b in formula (4) for each of y1 to y5. The model construction unit 72 determines the weight wi for each of y1 to y5 by repeating the calculations shown in the following formulas (5) and (6).

In formula (5), yj (j is an integer between 1 and 5) is the output of the model defined by the weight wi being updated, and tj is the correct value of the mass ratio or the coarse grain ratio. E and Ej are errors representing the difference between the output yj and the correct value tj. In formula (6), w' represents the weight w after updating. Repeating the calculations of formulas (5) and (6) means calculating the sum of the squared errors between the value of the output yj and the correct value tj when a certain value is input to the weight w, and updating the value of the weight w so that the error is minimized by the gradient method. The number of times the weight w is updated may be determined by an operator or the like. As the gradient method, for example, the steepest descent method, the stochastic gradient descent method (SGD), the mini-batch gradient descent method, Adam, the Nesterov accelerated algorithm (NAG), AdaGrad, RMSprop, Adadelta, Adam, Adamax, AMSGRAD, and Momentum can be used.

The explanation of the prediction model M using formulas (3) to (6) is a general explanation of a deep neural network. Examples of deep neural networks include a convolutional neural network, a regression neural network, and a recurrent neural network. In the aggregate quality prediction method according to the present embodiment, it is preferable to use, for example, a convolutional neural network. This is because, when input information including image data is used, a convolutional neural network is advantageous over a normal deep neural network in updating weights. When a normal deep neural network contains a large amount of information such as image data, the number of places where weights are calculated becomes very large. On the other hand, a convolutional neural network can share weights using a filter, simplifying the weight update. In this way, the use of a convolutional neural network is preferable in that it makes it possible to more appropriately capture the features of an image.

After step S12 is executed, step S13 is executed. In step S13, for example, the model holding unit 74 stores the prediction model M constructed in step S12. This completes the model construction process.

(Quality evaluation process)
3B is a flowchart showing an example of a series of processes executed in the quality evaluation process. This quality evaluation process is executed, for example, during the period when the manufacturing process is being executed in the manufacturing apparatus 10.

In this quality evaluation process, first, the control device 60 executes step S21. In step S21, for example, the quality prediction unit 76 waits until the measurement timing arrives. The measurement timing may be set in advance to a certain time during the day, or may be a timing when an instruction is received from an operator such as a worker. When the measurement timing arrives, the control device 60 may execute control so that the coarse aggregate is loaded onto the transport device 8 from the material storage area 2. The measurement timing may be set to coincide with the timing when the coarse aggregate is being transported to the transport device 8.

Next, the control device 60 executes step S22. In step S22, for example, the quality prediction unit 76 causes the imaging device 50 to capture an image of the aggregate to be evaluated, and acquires image data from the imaging device 50. As a result, the quality prediction unit 76 acquires evaluation information including image data obtained by capturing an image of the aggregate to be evaluated. In one example, the evaluation information includes only image data obtained by capturing an image of the aggregate to be evaluated. In step S22, for example, with the aggregate to be evaluated piled up on the conveyor belt of the transport device 8, the imaging device 50 captures an image of an area including the aggregate to be evaluated from a position above and away from the aggregate to be evaluated.

Next, the control device 60 executes step S23. In step S23, for example, the quality prediction unit 76 predicts the coarse aggregate ratio of the evaluation target based on the image data (evaluation information) acquired in step S22 and the prediction model M held by the model holding unit 74. In one example, the quality prediction unit 76 inputs the image data acquired in step S22 into the prediction model M to obtain an output value related to the coarse aggregate ratio that is directly output from the prediction model M. The quality prediction unit 76 also inputs the image data into the prediction model M to obtain an output value related to the mass ratio for each gradation class that is directly output from the prediction model M. The quality prediction unit 76 then converts the output value related to the mass ratio for each gradation class into a coarse aggregate ratio, thereby indirectly obtaining a predicted value of the coarse aggregate ratio.

Next, the control device 60 executes step S24. In step S24, for example, the output unit 78 outputs the two prediction results regarding the coarse aggregate ratio acquired in step S23. In one example, the output unit 78 displays the two prediction results regarding the coarse aggregate ratio on the monitor of the input/output device 62. This allows workers, etc. to check the prediction results regarding the coarse aggregate ratio. If the coarse aggregate ratio is outside the specified range, the workers, etc. who have checked the prediction results may take measures to maintain the quality of the ready-mix concrete (for example, adjusting the aggregate particle size, modifying the mix, temporarily suspending production, investigating the cause, etc.).

This completes the series of processes for predicting the coarse aggregate ratio once. After executing step S24, the control device 60 may repeat the series of processes from steps S21 to S24. In the above example, the coarse aggregate ratio is predicted at each measurement timing, but images of the aggregate being transported by the transport device 8 may be continuously captured, and all the aggregate used in the production of ready-mix concrete may be inspected.

The above-mentioned quality prediction method can be used as a substitute for a quality test by sampling, in addition to stabilizing the quality by performing a total inspection of the raw materials (materials) of the ready-mixed concrete in the production process (production line) of ready-mixed concrete. Examples of quality tests by sampling include an acceptance inspection, which is an inspection at the time of receiving raw materials, and a process inspection, which is an inspection at each production process of ready-mixed concrete. Quality tests by sampling are performed on a daily basis to supply ready-mixed concrete of stable quality. Conventionally, when measuring the coarse aggregate ratio by a quality test by sampling, for example, the coarse aggregate is sieved and the coarse aggregate ratio is calculated from the mass ratio. If the aggregate quality prediction method according to this embodiment is applied to a daily quality test, the time for the quality prediction test can be significantly shortened, and for example, the frequency of the quality prediction test in one day can be increased. This makes it possible to supply ready-mixed concrete of more stable quality. When a quality test by sampling is performed using the aggregate quality prediction method according to this embodiment, for example, the extracted aggregate is piled on a rubber sheet, and image data can be generated from the sheet by the imaging device 50. When the aggregate quality prediction method according to this embodiment is used for receiving inspection, it is also possible to attach an imaging device 50 to a transport truck or a transport ship, for example, and perform quality prediction during transport.

[Verification of prediction methods using prediction models]
In the following, in order to verify the method of predicting the coarse aggregate ratio using the prediction model M based on machine learning using a convolutional neural network, a comparison was made between the prediction results by the prediction model M and the correct value using a data set in which the correct value was known. For the learning aggregate, 10 levels of aggregate with different coarse aggregate ratios were prepared according to the conditions in Table 1 above. For each level, images were taken 10 times indoors with the aggregate in different states on the measurement sheet 8a, and 10 images were obtained.

In addition, a total of 100 images (10 levels x 10 shots) were prepared according to the standard conditions shown in Table 2 below. As a result, 100 data sets were prepared as learning data, with images and correct answers corresponding to each data set.

The prediction model M was constructed using 80 of the 100 datasets, and 20 datasets were used for validation. Specifically, for each level, eight pieces of data were used for model construction, and two datasets were used for validation. In validation using 20 datasets, for each dataset, two prediction results calculated using the image and the prediction model M were compared with the correct answer value, such as the coarse grain rate, associated with the image. The prediction model M was constructed using a convolutional neural network with six intermediate layers, each consisting of a convolutional layer and a pooling layer, followed by four fully connected layers. The Huber function was used as the loss function. Note that the number of images was increased in each dataset by the ImageAugmentation layer.

Figure 6(a) is a graph showing the relationship between the predicted values of mass ratios (%) in sections III and IV in Table 1 above and the corresponding correct values of mass ratios (%). In the graph shown in Figure 6(a), the horizontal axis is the correct value of mass ratio, and the vertical axis is the predicted value of mass ratio output from prediction model M. Prediction results in section III (5mm to 10mm) are plotted with △ marks, and prediction results in section IV (10mm to 20mm) are plotted with ◯ marks. A solid line is drawn where the correct value and the predicted value match, and the closer the plotted results are to the line, the higher the prediction accuracy. From the results shown in Figure 6(a), the prediction results are plotted near the line, and it can be seen that prediction model M is able to predict mass ratios with high accuracy.

Figure 6 (b) is a graph showing the relationship between the predicted value of the coarse grain ratio output from the prediction model M and the correct value of the coarse grain ratio. In the graph shown in Figure 6 (b), the horizontal axis is the correct value of the coarse grain ratio, and the vertical axis is the predicted value of the coarse grain ratio output from the prediction model M. The prediction results are plotted with squares. The solid line is the line where the correct value and the predicted value match, and a dashed upper limit line shifted vertically from the solid line by +0.10 and a lower limit line shifted vertically by -0.10 are drawn. The prediction results included in the area between the upper limit line and the lower limit line are defined as being within the allowable range. With reference to the error in the coarse grain ratio that is currently allowed in ready-mix concrete plants, a range of ±0.10 was set as the allowable range. It can be seen that all of the prediction results using the prediction model M are included in the allowable range, and the prediction model M can predict the coarse grain ratio with high accuracy.

Below, we verified the influence of various factors when preparing training data on predictions by prediction model M. Specifically, training data was prepared by changing at least one of the factors of the presence or absence of lighting, image size, amount of coarse aggregate, and moisture state of the aggregate surface from the reference conditions shown in Table 2. Then, for each condition, a prediction model M was constructed from the corresponding training data, and the prediction accuracy of prediction model M was evaluated using the corresponding validation data.

The graphs shown in Figure 7(a) and Figure 7(b) differ in the presence or absence of lighting. Figure 7(a) shows the comparison results when images were taken using only indoor lighting (no lighting), while Figure 7(b) shows the comparison results when images were taken using indoor lighting and LED lighting directed at the subject (with lighting). Furthermore, in both cases with and without lighting, verification was performed for the amounts of coarse aggregate of 2 kg, 5 kg, and 8 kg.

The average absolute difference was calculated as an index showing how far the predicted value deviates from the correct value. The average absolute difference is the arithmetic mean of the absolute values of the differences between the correct value and the predicted value in the 20 validation data sets. In the graphs shown in Figures 7(a) and 7(b), "direct prediction" represents the verification result regarding the output value of the coarse grain ratio directly output from the prediction model M. "indirect prediction" represents the verification result regarding the predicted value of the coarse grain ratio obtained indirectly by converting the output value of the mass ratio for each grain size category output from the prediction model M.

The vertical axis represents the average absolute difference, and "Tg" represents the target value. When the average absolute difference is equal to or less than the target value Tg (0.10), the accuracy of the prediction using prediction model M can be evaluated as good. As shown in Figures 7(a) and 7(b), whether the amount of coarse aggregate was 2 kg, 5 kg, or 8 kg, whether direct prediction or indirect prediction was used, and regardless of whether lighting was present or not, the average absolute difference was equal to or less than the target value Tg. Therefore, it is considered that the presence or absence of lighting has little effect on the accuracy of prediction by prediction model M.

The graphs shown in Fig. 8(a) and Fig. 8(b) use different prediction methods. Fig. 8(a) shows the comparison results when direct prediction among predictions by prediction model M is used, and Fig. 8(b) shows the comparison results when indirect prediction among predictions by prediction model M is used. As shown in Fig. 8(a) and Fig. 8(b), the absolute difference average value was below the target value Tg regardless of whether the amount of coarse aggregate was 2 kg, 5 kg, or 8 kg, whether lighting was present or not, and regardless of whether direct prediction or indirect prediction was used. Therefore, it is considered that the selection of direct prediction or indirect prediction has little effect on the accuracy of prediction by prediction model M.

In each of the graphs shown in Figures 9(a) and 9(b), verification was performed by varying the amount of coarse aggregate. Figure 9(a) shows the comparison results when direct prediction was used and the amount of coarse aggregate was varied, and Figure 9(b) shows the comparison results when indirect prediction was used and the amount of coarse aggregate was varied. As shown in Figures 9(a) and 9(b), the average absolute difference was below the target value Tg, regardless of whether direct prediction or indirect prediction was used, whether lighting was present or not, and regardless of the amount of coarse aggregate. Therefore, it is considered that the amount of coarse aggregate has little effect on the accuracy of predictions made by prediction model M.

In each of the graphs shown in FIG. 10(a) and FIG. 10(b), the image size was changed to perform verification. FIG. 10(a) shows the comparison results when direct prediction was used and the image size was changed, and FIG. 10(b) shows the comparison results when indirect prediction was used and the image size was changed. In each graph, the horizontal axis represents the image size ratio, and the image size under the reference conditions in Table 2 is set to 1. The lower limit of the image size is set to a range in which the aggregate can be visually recognized by a human from the image. As shown in FIG. 10(a) and FIG. 10(b), regardless of whether direct prediction or indirect prediction was used, whether the amount of aggregate was 2 kg, 5 kg, or 8 kg, and regardless of the image size ratio, the absolute difference average value was below the target value Tg. Therefore, it is considered that the image size (number of pixels or resolution) has little effect on the accuracy of prediction by the prediction model M.

11(a) and 11(b) show the results of the verification performed with different moisture conditions on the surface of the aggregate. In this verification, the "air-dry state" is a state in which the learning aggregate is left indoors for a specified time to dry (naturally dried). In addition, the "wet state" is a state in which a specified ratio of water (for example, 1% to 5% by mass) is added to the learning aggregate and mixed in a container. From the results shown in FIG. 11(a) and FIG. 11(b), it can be seen that in the air-dry state, the absolute difference average value is equal to or less than the target value Tg, and the prediction model M can predict with high accuracy. In addition, in the image of the aggregate obtained in the air-dry state, when there is no illumination, there is an overall uneven color compared to the air-dry state, and the edges of each particle of the aggregate appear blurred. In addition, in the image of the aggregate obtained in the air-dry state, when there is illumination, there are aggregate particles that appear white due to the reflection of surface water compared to the air-dry state.

[Modification]
As shown in Fig. 11(a), when direct prediction is adopted and the amount of coarse aggregate is 5 kg and 8 kg, the absolute difference average value exceeds the target value Tg. As shown in Fig. 11(b), when indirect prediction is adopted and the amount of coarse aggregate is 8 kg, the absolute difference average value exceeds the target value Tg. As such, it can be seen that in a wet state, the prediction accuracy may decrease depending on the amount of coarse aggregate. In order to avoid a decrease in prediction accuracy in a wet state, a prediction model M may be constructed differently from the above example.

The prediction model M constructed by the model construction unit 72 may output only a value indicating the coarse aggregate ratio in response to input information including image data obtained by photographing the training aggregate. In other words, the prediction model M may be a model that performs a single task. In this case, in the training data for constructing the prediction model M, only the correct value of the coarse aggregate ratio is associated with the input information including image data.

The model constructing unit 72 may construct a prediction model M shown in the following formula (7) instead of the above formula (4).

In the process of updating the weights w for generating the prediction model M, the model construction unit 72 may repeat the calculation shown in the following equation (8) instead of the above equation (5).

Figure 12 shows the results of a verification performed under the same conditions as the verification results shown in Figure 11 (a), except that the prediction model M shown in equation (7) was used as the prediction model. As shown in Figure 12, the results show that the average absolute difference is below the target value Tg in both air-dry and wet conditions. In other words, it can be seen that the coarse grain ratio can be predicted with high accuracy by using the prediction model M that performs single tasking, even in wet conditions.

When the prediction model M performs multitasking, the calculation method in the process of updating the weight w may be different from the example described above. One of the factors that may cause a decrease in prediction accuracy in a wet state is the difference between the order of error related to the mass ratio and the order of error related to the coarse grain ratio in the process of updating the weight w. For example, among E, which is the total error in the above formula (5), there is a difference of about 10 to 100 times between the order of errors E1, E2, E3, and E4 related to the mass ratio (%) and the order of error E5 related to the coarse grain ratio. In one example, the order of error related to the mass ratio is about 1.0 to 10, and the order of error related to the coarse grain ratio is about 0.001 to 0.10.

Repeating the calculations of the above formulas (5) and (6) to update the weight w is equivalent to updating the intermediate model that is generated in the process of constructing a predictive model based on the error in the coarse aggregate fraction (first error) and the error in the mass ratio for each grading class (second error). If the order of the error in the coarse aggregate fraction is smaller than the order of the error in the mass ratio, the contribution of E5 in the above formula (5) becomes small, and there is a possibility that the calculation to update the weight w to reduce E5 will hardly be performed in the process of updating the intermediate model.

The model construction unit 72 may update the intermediate model when constructing the prediction model M, using errors corrected so as to reduce the difference between the error related to the coarse aggregate fraction and the error related to the mass ratio for each gradation class. For example, the model construction unit 72 multiplies each of the two types of errors by a coefficient so as to reduce the difference between the errors. In one example, the model construction unit 72 may repeat the calculation shown in the following formula (9) instead of the above formula (5) in the process of updating the weight w for generating the prediction model M (process of updating the intermediate model).

In formula (9), λ1 and λ2 are coefficients. The coefficients λ1 and λ2 may be determined by an operator or the like based on the order of error in the coarse grain ratio and the order of error in the mass ratio. The coefficient λ2 may be about 5 to 200 times the coefficient λ1. The coefficient λ2 may be 10 times the coefficient λ1, or 100 times the coefficient λ1. In this disclosure, the multitasking performed by the prediction model M constructed by updating the weight w by the calculation of formula (9) is referred to as "coefficient adjustment multitasking." In addition, when simply referring to "multitasking," it means multitasking in which error adjustment is not performed in the process of updating the weight w.

Figure 13 shows the verification results in a wet state for each of multitasking, coefficient adjustment multitasking, and single tasking. The amount of coarse aggregate is 8 kg, and the image size ratio is 0.25. In multitasking and coefficient adjustment multitasking, the coarse aggregate ratio is predicted by direct prediction, one of the two prediction methods. Coefficient λ1 is set to 1.00, and coefficient λ2 is set to 100. From the results shown in Figure 13, it can be seen that in a wet state, the prediction accuracy is higher in coefficient adjustment multitasking and single tasking than in multitasking.

Figure 14 shows the results of the verification when the values of coefficient λ1 and coefficient λ2 were changed under the same conditions as the verification results shown in Figure 13. Specifically, the evaluation was performed by setting the coefficient combinations (λ1, λ2) to (1.0, 10), (0.1, 1.0), (1.0, 100), and (0.01, 1.0). As shown in Figure 14, whether the ratio of coefficient λ2 to coefficient λ1 was 10 or 100, the average absolute difference was below the target value of 0.1. From the results shown in Figure 14, it can be seen that the average absolute difference tends to be smaller when the ratio of coefficient λ2 to coefficient λ1 is 100 compared to when the ratio is 10.

The various prediction models described above are just examples, and the prediction model may be any model that can output at least a predicted value regarding the coarse aggregate from information including image data of the aggregate. In the learning phase, the aggregate on the conveying device 8 may be imaged and learning images (learning data) may be prepared. The input information and evaluation information input to the prediction model may include, in addition to image data of the aggregate, information indicating physical quantities that may affect the coarse aggregate ratio.

The manufacturing system 1 may include a computer that controls the manufacturing device 10 and a computer that predicts the coarse aggregate ratio (quality prediction device). The computer that predicts the coarse aggregate ratio may have a model construction unit 72, a model holding unit 74, a quality prediction unit 76, and an output unit 78. In one example of the various examples described above, at least some of the matters described in the other examples may be applied.

[Summary of the Disclosure]
The above-described aggregate quality prediction method includes a preparation step of constructing a prediction model M that outputs a value indicating the coarse aggregate ratio in response to input of input information, by machine learning based on input information including image data obtained by photographing the aggregate and a ground truth value of the coarse aggregate ratio corresponding to the input information, an acquisition step of acquiring evaluation information that includes image data obtained by photographing the aggregate to be evaluated and corresponds to the input information, and a prediction step of inputting the evaluation information into the prediction model M and acquiring a first predicted value regarding the coarse aggregate ratio output from the prediction model M.

The coarse aggregate ratio of aggregates affects the quality of ready-mix concrete produced by mixing concrete materials containing aggregates. For example, in the ready-mix concrete industry, a great deal of effort is spent on process inspections such as fresh property tests, which are part of quality control work. For this reason, it is desirable to reduce the labor required for quality control in order to improve the efficiency of ready-mix concrete production. In the above quality prediction method, the coarse aggregate ratio of aggregates is predicted using a prediction model M constructed by machine learning, so that the work required to inspect the coarse aggregate ratio of aggregates used in the production of ready-mix concrete can be simplified. As a result, the inspection frequency can be increased, or labor can be spent on quality control work other than the management of the coarse aggregate ratio, which is useful for stabilizing the quality of ready-mix concrete.

In the quality prediction method described above, in the preparation process, the prediction model M may be constructed to output only a value indicating the coarse aggregate ratio in accordance with the input information. By having the prediction model M perform the above single task, as described above, the coarse aggregate ratio can be predicted with high accuracy using the prediction model M even when the surface condition of the aggregate is wet. In factories that manufacture ready-mix concrete, aggregate often contains some moisture, and improving the prediction accuracy in wet conditions is more beneficial for stabilizing the quality of ready-mix concrete.

In the quality prediction method described above, the preparation step may include constructing a prediction model M by machine learning based on the correct value of the mass ratio of each of the multiple granularity classifications of the aggregate associated with the input information. In the preparation step, the prediction model M may be constructed to further output a value indicating the mass ratio of each classification included in the multiple granularity classifications according to the input information. The prediction step may include inputting evaluation information into the prediction model M to further obtain a second predicted value output from the prediction model M regarding the mass ratio of each classification, and converting the second predicted value into the coarse aggregate ratio. In this case, the coarse aggregate ratio can be predicted using one prediction model M by two different methods. Therefore, the reliability of the prediction result of the coarse aggregate ratio can be evaluated while avoiding an increase in the work required for model construction.

In the quality prediction method described above, the preparation process includes updating an intermediate model generated in the process of constructing the prediction model M based on a first error related to the coarse aggregate ratio and a second error related to the mass ratio for each category, and in the preparation process, the intermediate model is updated using an error corrected so as to reduce the difference between the first error and the second error. As described above, by the prediction model M performing the coefficient adjustment multitasking, the coarse aggregate ratio can be accurately predicted using the prediction model M even when the surface condition of the aggregate is wet. In factories that manufacture ready-mixed concrete, aggregate often contains some moisture, and improving the prediction accuracy in a wet state is more beneficial for stabilizing the quality of ready-mixed concrete.

In the quality prediction method described above, the preparation process may include constructing a prediction model M by machine learning using a convolutional neural network. In this case, a more complex model can be constructed by performing machine learning using a convolutional neural network. As a result, the predicted value of the coarse grain ratio can be brought closer to the correct value.

In the quality prediction method described above, the acquisition step may include acquiring evaluation information by imaging an area including the aggregate to be evaluated from a position away from the aggregate to be evaluated while the aggregate to be evaluated is piled up. In a factory that manufactures ready-mix concrete, aggregate is often piled up. For example, at the time of acceptance, the aggregate is transported loaded on a transport truck or a transport ship. Also, at the time of production, the aggregate is transported in a stacked state by a belt conveyor or the like. In the above configuration, the image of the aggregate is taken in a stacked state, so that the work for acquiring evaluation information can be simplified. In addition to the above-mentioned transportation, the above quality prediction method can also be used as a substitute for quality control work such as a quality screening test. For example, at the time of acceptance or production, a part of the aggregate may be taken and laid out on a floor or a sheet and imaged.

In the quality prediction method described above, the preparation step may include preparing training aggregate so that it contains particles of two or more particle size categories. The input information may include image data obtained by imaging the training aggregate. In this case, the training aggregate becomes closer to the actual aggregate used in the factory that produces ready-mix concrete, and the accuracy of prediction by the prediction model M can be improved.

The manufacturing method of ready-mixed concrete described above includes a manufacturing process of mixing materials including aggregate to manufacture ready-mixed concrete, and a quality prediction process of predicting the coarse aggregate ratio of at least a portion of the aggregate used in the manufacturing process. The quality prediction process includes a preparation process of constructing a prediction model M that outputs a value indicating the coarse aggregate ratio in response to input of input information by machine learning based on input information including image data obtained by imaging the aggregate and a correct answer value of the coarse aggregate ratio associated with the input information, an acquisition process of acquiring evaluation information including image data obtained by imaging the aggregate to be evaluated and corresponding to the input information, and a prediction process of inputting the evaluation information into the prediction model M and acquiring a first predicted value output from the prediction model M regarding the coarse aggregate ratio. This manufacturing method is useful for stabilizing the quality of ready-mixed concrete, similar to the above quality prediction method.

The manufacturing system 1 described above includes a manufacturing device 10 that mixes materials including aggregate to manufacture ready-mix concrete, and a control device 60 (quality prediction device) that predicts the coarse aggregate ratio of at least a portion of the aggregate used by the manufacturing device 10. The control device 60 is configured to execute a preparation process of constructing a prediction model M that outputs a value indicating the coarse aggregate ratio according to the input of input information by machine learning based on input information including image data obtained by imaging the aggregate and a correct answer value of the coarse aggregate ratio associated with the input information, an acquisition process of acquiring evaluation information corresponding to the input information including image data obtained by imaging the aggregate to be evaluated, and a prediction process of inputting the evaluation information into the prediction model M and acquiring a first predicted value output from the prediction model M regarding the coarse aggregate ratio. This manufacturing system 1 is useful for stabilizing the quality of ready-mix concrete, similar to the above-mentioned quality prediction method.

1... manufacturing system, 2... material storage area, 4... silo, 8... transport device, 10... manufacturing device, 50... imaging device, 60... control device, 72... model construction unit, 74... model storage unit, 76... quality prediction unit, 78... output unit, M... prediction model.

The graphs shown in FIG. 11(a) and FIG. 11(b) show the results of the verification performed with different moisture conditions on the surface of the aggregate. In this verification, the "air-dry state" is a state in which the learning aggregate is left indoors for a predetermined time and dried (naturally dried). In addition, the "wet state" is a state in which water is added to the learning aggregate at a predetermined ratio (for example, 1% to 5% by mass ratio) relative to the mass of the aggregate, and the aggregate is mixed in a container. From the results shown in FIG. 11(a) and FIG. 11(b), it can be seen that in the air-dry state, the absolute difference average value is equal to or less than the target value Tg, and the prediction model M can predict with high accuracy. In addition, in the image of the aggregate obtained in the wet state, when there is no illumination, there is an overall unevenness in color compared to the air-dry state, and the edges of each particle of the aggregate appear blurred. In addition, in the image of the aggregate obtained in the wet state, when there is illumination, there are aggregate particles that appear white due to the reflection of surface water compared to the air-dry state.

Claims (9)

a preparation step of constructing a prediction model that outputs a value indicating the coarse aggregate ratio of the aggregate in response to input of the input information by machine learning based on input information including image data obtained by photographing the aggregate and a correct answer value of the coarse aggregate ratio of the aggregate associated with the input information;
an acquisition step of acquiring evaluation information corresponding to the input information, the evaluation information including image data obtained by imaging an aggregate to be evaluated;
A prediction step of inputting the evaluation information into the prediction model and obtaining a first predicted value output from the prediction model regarding the coarse aggregate ratio. The method for predicting aggregate quality according to claim 1, wherein in the preparation step, the prediction model is constructed to output only a value indicating the coarse aggregate ratio in response to the input information. The preparation step includes constructing the prediction model by machine learning further based on correct values of mass ratios of each of a plurality of aggregate granularity classes associated with the input information,
In the preparation step, the prediction model is constructed to further output a value indicating a mass ratio for each of the plurality of particle size categories in response to the input information,
The prediction step includes:
inputting the evaluation information into the prediction model to further obtain a second predicted value output from the prediction model regarding the mass ratio for each of the sections;
The method for predicting quality of aggregate according to claim 1 , further comprising converting the second predicted value into a coarse aggregate ratio. The preparation step includes updating an intermediate model generated in the process of constructing the prediction model based on a first error related to a coarse aggregate ratio and a second error related to a mass ratio for each of the sections;
The method for predicting quality of aggregate according to claim 3 , wherein in the preparation step, the intermediate model is updated using an error corrected so as to reduce a difference between the first error and the second error. The method for predicting aggregate quality according to any one of claims 1 to 4, wherein the preparation step includes constructing the prediction model by machine learning using a convolutional neural network. The method for predicting aggregate quality according to any one of claims 1 to 4, wherein the acquisition step includes capturing an image of an area including the aggregate to be evaluated from a position away from the aggregate to be evaluated while the aggregate to be evaluated is piled up, and acquiring the evaluation information. The preparing step includes preparing a learning aggregate so as to include grains of two or more grain size classes;
5. The method for predicting quality of aggregate according to claim 1, wherein the input information includes image data obtained by imaging the learning aggregate. A manufacturing process in which materials including aggregate are mixed to produce ready-mix concrete;
a quality prediction step of predicting a coarse aggregate ratio of at least a portion of the aggregate used in the manufacturing step;
Including,
The quality prediction step includes:
a preparation step of constructing a prediction model that outputs a value indicating the coarse aggregate ratio of the aggregate in response to input of the input information by machine learning based on input information including image data obtained by photographing the aggregate and a correct answer value of the coarse aggregate ratio of the aggregate associated with the input information;
an acquisition step of acquiring evaluation information corresponding to the input information, the evaluation information including image data obtained by imaging an aggregate to be evaluated;
A prediction process of inputting the evaluation information into the prediction model and obtaining a first predicted value output from the prediction model regarding the coarse aggregate ratio. A manufacturing device for mixing materials including aggregate to manufacture ready-mix concrete;
a quality prediction device for predicting a coarse aggregate ratio of at least a portion of the aggregate used by the manufacturing device;
The quality prediction device includes:
a preparation step of constructing a prediction model that outputs a value indicating the coarse aggregate ratio of the aggregate in response to input of the input information by machine learning based on input information including image data obtained by photographing the aggregate and a correct answer value of the coarse aggregate ratio of the aggregate associated with the input information;
an acquisition step of acquiring evaluation information corresponding to the input information, the evaluation information including image data obtained by imaging an aggregate to be evaluated;
A prediction process of inputting the evaluation information into the prediction model and obtaining a first predicted value output from the prediction model regarding the coarse aggregate ratio.

Images