JP7748850B2 - Concrete quality prediction method - Google Patents

Description

The present invention relates to a method for predicting the quality of concrete.

At a concrete product factory, ready-mix concrete is first produced by mixing the various materials in a mixer based on a concrete mix adjusted to suit the application, quality, etc. of the concrete product. Next, compaction, surface finishing, curing, etc. are carried out as appropriate according to the properties of the ready-mix concrete to produce the concrete product.
On the other hand, even if the mix proportions and mixing time are the same, the properties of ready-mixed concrete will vary depending on the surface moisture content of the aggregate, the outside temperature, etc., so in order to produce the desired concrete product, it is necessary to adjust the conditions in the finishing process as appropriate.
For example, if ready-mixed concrete with higher fluidity than expected is produced, there is a high possibility that sagging will occur during finishing, so the finishing process is carried out after leaving it to stand for a predetermined period of time. Also, if ready-mixed concrete with lower fluidity than expected is produced, the time required for the finishing process may be longer than expected.

Patent Document 1 describes a method for predicting the quality of ready-mixed concrete using a prediction model, which is capable of predicting the quality of ready-mixed concrete in a short time and with high accuracy without relying on human sense or experience. The prediction model is created by machine learning using multiple pieces of training data, which are a combination of training input data including image data and training output data related to the quality of ready-mixed concrete. The prediction input data including image data is input to the prediction model, and prediction output data related to the quality of ready-mixed concrete is output from the prediction model, and the quality of ready-mixed concrete is predicted using the prediction output data.

Patent Document 2 also describes a method for predicting a suitable time for leveling work before carrying out leveling work on the surface of concrete being constructed. The method includes the steps of: assuming that there is a correlation between the air temperature during the construction period and the temperature of the concrete; applying the predicted air temperature during the construction period to the relationship between the slope information of an approximation equation that represents the relationship between a penetration resistance value obtained by a predetermined penetration resistance test and elapsed time, and the temperature of the concrete, to obtain slope information corresponding to the predicted air temperature; applying the obtained slope information to the relationship between the slope information and the time from the water pouring time when preparing fresh concrete to the time when the concrete's conductivity decreases, which is previously known; estimating the time when the conductivity decreases corresponding to the obtained slope information; estimating the relationship between the penetration resistance value corresponding to the obtained slope information and elapsed time; and predicting the time when the concrete can be leveled from the relationship between the estimated penetration resistance value and elapsed time and the estimated time when the conductivity decreases.

Japanese Patent Application Laid-Open No. 2020-144132 Japanese Patent Application Laid-Open No. 2019-73962

The slump value of concrete is known as one of the indicators for determining finishing processes such as surface finishing and compaction of concrete.
However, when the slump value of concrete is set low from the beginning, as in the case of low-slump concrete commonly used in concrete product factories, the reliability of the slump value as an indicator of the above-mentioned problem is low. For this reason, the adjustment of various conditions in the finishing process largely depends on the sense and experience of skilled workers.
An object of the present invention is to provide a method for predicting the quality of concrete (particularly the finishability of concrete) in a short time with high accuracy without relying on human senses or experience.

As a result of intensive research into solving the above-mentioned problems, the inventors discovered that the above-mentioned object can be achieved by a method of predicting the quality of concrete by inputting input data including image data into a prediction model created by machine learning using multiple combinations of input data including image data and output data related to the finishability of concrete, and using the output data obtained, and thus completed the present invention.
That is, the present invention provides the following [1] to [7].
[1] A method for predicting the quality of concrete using a prediction model, wherein the prediction model is created by machine learning using a plurality of pieces of training data that are a combination of training input data including image data and training output data related to the finishability of concrete, the method comprising: inputting prediction input data including image data into the prediction model; outputting prediction output data related to the finishability of concrete from the prediction model; and predicting the quality of concrete using the prediction output data.
[2] The method for predicting quality of concrete according to [1], wherein the learning output data and the prediction output data include data related to work during finishing of the concrete.

[3] The method for predicting quality of concrete according to [1] or [2], wherein the learning output data and the prediction output data include data on the quality of the concrete after finishing.
[4] The method for predicting quality of concrete according to any one of [1] to [3], wherein at least one type of data among the data used as the learning input data and the data used as the prediction input data is standardized data.
[5] The method for predicting quality of concrete according to [4], wherein the standardized data is image data included in the learning input data and the prediction input data, the image data is a grayscale image, and the pixel values of each pixel constituting the grayscale image are standardized.
[6] A method for predicting the quality of concrete according to any one of [1] to [5], wherein the image data is a plurality of image data continuously captured to show the mixing of concrete materials in a mixer for mixing the materials, and the image data and the prediction output data are displayed on a display means.
[7] The method for predicting quality of concrete according to any one of [1] to [6], wherein at least one type of data among the data included in the learning input data and the prediction input data is obtained from an information management system in a concrete factory.

By using the concrete quality prediction method of the present invention, it is possible to predict concrete quality (especially concrete finishability) in a short amount of time and with high accuracy. This makes it possible to efficiently and stably produce concrete of the desired quality.

In the embodiment, this figure shows a still image taken from above the two-axis mixer, the position of image data extracted from the still image (a), and the position of image data shifted by 5 pixels from the position (b).

The concrete quality prediction method of the present invention is a method for predicting concrete quality using a prediction model, which is created by machine learning using multiple pieces of training data that are a combination of training input data including image data and training output data related to the finishability of concrete.The prediction input data including image data is input to the prediction model, and prediction output data related to the finishability of concrete is output from the prediction model, and the quality of concrete is predicted using the prediction output data.
A detailed explanation is provided below.

The predictive model was created using machine learning.
Examples of learning methods used in machine learning include neural networks, linear regression, decision trees, support vector regression, ensemble methods, support vector machines, discriminant analysis, naive Bayes methods, nearest neighbor methods, etc. These may be used alone or in combination of two or more.
Among these, neural networks are preferred from the viewpoint of being able to predict quality with higher accuracy.
From the viewpoint of being able to predict quality with higher accuracy, a hierarchical neural network having one or more intermediate layers between an input layer and an output layer is preferable.

Examples of neural networks include convolutional neural networks (CNNs) such as a 3D convolutional neural network (3DCNN), recurrent neural networks (RNNs), and long short-term memory (LSTM) neural networks (recurrent neural networks improved using LSTM).
Among these, convolutional neural networks (neural networks with a convolutional layer, a pooling layer, or the like as an intermediate layer) are more suitable because they have excellent performance in the field of image recognition. Convolutional neural networks can detect features from image data and create a prediction model capable of classification or regression using the features.
The number of layers in a convolutional neural network, which is a combination of convolutional layers and pooling layers, is preferably two or more, and more preferably three or more, from the viewpoint of enabling prediction with higher accuracy.
Machine learning can also be performed using, for example, "TensorFlow"("TENSORFLOW" is a registered trademark), a software library developed by Google, or "IBM Watson"("IBMWATSON" is a registered trademark), a system developed by IBM.

The predictive model is created through machine learning using multiple training data, which are combinations of training input data including image data and training output data related to the finishability of concrete.
Examples of image data used as training input data include image data related to concrete production. Specifically, examples include image data photographed showing the mixing of concrete materials in a mixer (hereinafter simply referred to as a "mixer"), image data photographed after mixing the materials in the mixer to obtain ready-mixed concrete and then pouring the ready-mixed concrete from the mixer into a hopper, image data photographed showing the state of ready-mixed concrete being stirred in the drum of a truck agitator (for example, image data photographed by shining a light or the like near the ready-mixed concrete inlet of the truck agitator, and image data photographed from a monitor displaying the history of the mixer's power load value during mixing of concrete materials (visually displaying the change in power load value over time using a graph or the like). These may be used alone or in combination of two or more.
Among these, from the viewpoint of being able to predict quality with higher accuracy, image data photographed while the materials are being mixed in the mixer is preferred. The image data may be image data taken at any time from immediately after mixing the materials to the end of mixing, but from the viewpoint of being able to predict quality at an early stage, it is preferred to take the image data when the rate of decrease in the power load value becomes gradual.

The number of image data used as learning input data is preferably 100 or more, more preferably 1,000 or more, even more preferably 10,000 or more, even more preferably 50,000 or more, and particularly preferably 100,000 or more, from the viewpoint of being able to predict quality with higher accuracy.
The image data may be either moving image data or still image data, and the image may be either a two-dimensional image or a three-dimensional image.
The image data is captured by a camera appropriately installed inside the mixer or around the mixer, etc. For example, a camera is installed at the top of the mixer so that the materials being mixed can be clearly seen.
In addition, in order to make the unevenness of each mixed material more visible, light may be applied from the side or diagonally above each mixed material using a light or the like to create a darker shadow.
In addition, from the viewpoint of being able to predict quality with higher accuracy, it is preferable to capture image data when the stirring blade (mixer blade) rotating in the mixer is at a specific rotation position that has been arbitrarily determined. The position may be one position or two or more positions.

The image data may be an image of a specific range cut out from an image taken inside the mixer (for example, an image taken so that the entire inside of the mixer is visible).
From the viewpoint of enabling predictions with higher accuracy, it is preferable that the above-mentioned specific range be a range in which parts of the concrete that are likely to show the behavior of the materials are likely to be captured when continuous images are taken of the process of mixing concrete materials inside the mixer.
Specifically, the specific range is the vicinity of all of the following parts (1) to (3) of the mixing member consisting of the mixer's rotating shaft and the mixing blades fixed to the rotating shaft, and is a range in which concrete materials may be reflected.
(1) At least a portion of the rotating shaft; (2) the tip portion of the agitator blade; and (3) the portion of the rotating shaft where the agitator blade is fixed. If there are multiple agitator blades, the specific range should be such that the tip portion and the fixed portion of at least one agitator blade are reflected in the image.
Furthermore, if the mixer is a twin-shaft mixer, it is preferable that the range be such that all of the above-mentioned parts (1) to (3) are reflected on both one rotating shaft and the other rotating shaft, and that the concrete material is likely to be reflected.
It should be noted that multiple image data may be extracted from one image taken of the inside of the mixer.

Difference data obtained from two image data arbitrarily selected from two or more captured image data may be used as image data to be used as learning input data. For example, difference data between two image data captured when the mixer blades are positioned at two arbitrarily determined positions within the mixer may be used as image data to be used as learning input data.
In this specification, the differential data refers to image data obtained by comparing two pieces of image data and extracting only the different portions.
Alternatively, image data may be obtained by superimposing multiple still images in succession over time. Such image data can be obtained, for example, by using commercially available image software to adjust the opacity (or transparency) of each of the multiple still images to about 30 to 70% and then superimposing them. From the standpoint of work efficiency, the number of still images to be superimposed is preferably 2 to 20.
Image data is captured for 10 seconds, for example, at 10 frames per second, thereby obtaining data for 100 still images.

From the viewpoint of being able to predict quality with higher accuracy, data on the power load value of the mixer during mixing (hereinafter simply referred to as "power load value") may also be used as learning input data.
Examples of data related to power load values include power load values and their absolute values at specific times from immediately after the start of mixing each material to the end of mixing, maximum or minimum power load values during mixing, amount of change in power load value per unit time, change pattern of power load value, etc. These may be used alone or in combination of two or more.
Furthermore, the number of data relating to the power load value may be one, but from the viewpoint of being able to predict quality with higher accuracy, it is preferably two or more, and more preferably five or more.

Furthermore, image data to be used as learning input data (image data captured by photographing the mixing of the materials in the mixer) may be determined based on data related to the power load values. For example, from the perspective of being able to predict quality with higher accuracy and at an earlier stage, it is preferable to use image data captured after mixing of water and other concrete materials has started and the mixer's power load value has stabilized (the point at which the rate of change in the power load value per unit time has become small).
When mixing concrete ingredients, the time when the mixer's power load value stabilizes varies depending on the water-cement ratio of the concrete, etc., but if the water-cement ratio is around 50 to 70%, it will be around 30 seconds after the start of mixing water and concrete ingredients other than water (for example, after the concrete ingredients other than water are poured into the mixer and dry mixed, water is poured into the mixer and mixing begins).
Furthermore, if the water-cement ratio is set lower than the above-mentioned range (50 to 70%) from the viewpoint of strength development, the time when the power load value stabilizes will be delayed, and in the case of high-strength concrete, it may take 5 to 10 minutes from the start of mixing the water and concrete materials other than water.

Furthermore, two or more image data may be captured based on data related to the power load value, and the difference data obtained from two image data arbitrarily selected from the captured two or more image data may be used as image data to be used as learning input data. For example, once the power load value has stabilized, image capture may be performed for an arbitrary period of time (for example, capturing 30 frames per second), and the difference data obtained from two image data arbitrarily selected from the multiple image data obtained may be used as image data to be used as learning input data.

Moreover, from the viewpoint of predicting quality with higher accuracy, other data may also be used as learning input data.
Other data include data on concrete mix conditions, data on concrete quality, data on cement, data on concrete materials other than cement, data on mixing means and methods, data on the environment during mixing, data on concrete transportation, etc. These may be used alone or in combination of two or more.
These various data include numerical data and classification data.
In this specification, "numerical data" refers to data that can be expressed as specific numerical values. Numerical data also includes data obtained by dividing the quality of concrete into multiple stages and evaluating it based on the worker's visual inspection or sense when working manually (for example, when hitting concrete with a trowel, lifting concrete with a hand shovel, etc.), and replacing each stage with a numerical value.
Data relating to classification means data classified according to criteria such as a specific mix design, a specific type, a specific property, or a specific numerical range.

Examples of data related to the mix conditions of concrete include the mix ratios (for example, the amount (mass%) of admixture relative to 100% by mass of cement) of cement, fine aggregate, coarse aggregate, water, various admixtures (AE agents, water-reducing agents, AE water-reducing agents, high-performance water-reducing agents, high-performance AE water-reducing agents, superplasticizers, setting retarders, etc.), and various admixtures (ground granulated blast furnace slag, silica fume, fly ash, etc.) mixed into the concrete, as well as items in the design mix table, such as the water-cement ratio, air content, fine aggregate rate, unit water content (amount of water per 1 ^m3 of concrete), unit cement content, unit fine aggregate content, unit coarse aggregate content, unit admixture content, and unit admixture content. These may be used alone or in combination of two or more.

Examples of data relating to the quality of concrete include target concrete design data such as strength (nominal strength, compressive strength, flexural strength, etc.), slump, slump flow, air content, chloride content, crack resistance, dynamic modulus of elasticity, dynamic shear modulus of elasticity, dynamic Poisson's ratio, hardened body void volume and void size distribution, durability, color tone, etc. These may be used alone or in combination of two or more.
Of these, slump and slump flow are preferred because they are highly important in terms of concrete quality.
The strength, slump, slump flow, air content, and chloride content of concrete can be measured, for example, by the test methods described in "JIS A 5308:2014 (Ready-mixed concrete)."

Examples of data related to cement include data related to cement as a whole, data related to raw materials for cement clinker, data related to burning conditions for cement clinker, data related to grinding conditions for cement, and data related to cement clinker.
Examples of data related to cement as a whole include: (i) data on cement used as a material for concrete, such as type, chemical composition, mineral composition, wet f. CaO (free lime content), loss on ignition, Blaine specific surface area, particle size distribution, sieve residue, and color; (ii) mineralogical properties and crystallographic properties of each mineral contained in cement; and (iii) the hemihydration rate of gypsum contained in cement.
Examples of data on raw materials for cement clinker include: (i) data on the blended raw materials for cement clinker, such as chemical composition, hydraulic ratio calculated from the chemical composition, sieve residue, Blaine specific surface area (fineness), ignition loss, supply amount, supply amount of auxiliary raw materials (special raw materials such as waste), amount stored in the blending silo (remaining amount), amount stored in the storage silo (remaining amount), (ii) current value of the cyclone located between the raw material mill and the blending silo for blended raw materials (representing the rotation speed of the cyclone, which is correlated with the speed of the raw materials passing through the cyclone), (iii) time from when the raw materials are fed into the kiln, (iv) Data on the raw materials for cement clinker, such as the chemical composition, hydraulic hardness calculated from the chemical composition, of the raw materials for cement clinker (mixed raw materials for cement clinker from which fine particles and the like have been removed by a countercurrent airflow during transportation; hereinafter referred to as the raw materials for cement clinker to be kilned) at a point in time a predetermined time ago (for example, one point in time 5 hours ago, or multiple points in time such as four points in time 3 hours, 4 hours, 5 hours, and 6 hours ago), (v) data on the raw materials obtained by mixing the raw materials for cement clinker to be kilned and auxiliary raw materials, such as the chemical composition, hydraulic hardness calculated from the chemical composition, Blaine specific surface area, sieve residue amount, decarbonation rate, and moisture content.

Examples of data related to cement burning conditions include: (i) data during cement clinker burning, such as the amount of cement clinker raw material inserted into the kiln, kiln rotation speed, outlet temperature, burning zone temperature, cement clinker temperature, kiln average torque, _O2 concentration, _NOx concentration; (ii) clinker cooler temperature; and (iii) preheater gas flow rate (which is correlated with the preheater temperature).
Examples of data related to cement milling conditions include milling temperature, amount of water sprayed inside the finishing mill, separator air volume, type of gypsum, amount of gypsum added, amount of cement clinker added, number of revolutions of the finishing mill, temperature of powder discharged from the finishing mill, amount of powder discharged from the finishing mill, amount of powder not discharged from the finishing mill, etc.
Examples of data related to cement clinker include (i) mineral composition, chemical composition, wet f. CaO (free lime), and volume weight, (ii) crystallographic properties (lattice constant, crystallite size, etc.) of each mineral contained in the cement clinker, and (iii) ratios of two or more minerals contained in the cement clinker.
These may be used alone or in combination of two or more.

Examples of data on concrete materials other than cement include (i) aggregate (fine aggregate and coarse aggregate) data such as type, density, water absorption rate, water content, surface water content, particle size distribution, maximum size, and particle shape, (ii) type of admixture, and (iii) type of admixture. These may be used alone or in combination of two or more.
Examples of data relating to the mixing means and method include the type, model, and capacity of the mixer, the amount of material to be mixed, the mixing time, etc. These may be used alone or in combination of two or more.
Examples of data relating to the environment during mixing include temperature (outside air temperature, temperature inside the mixer, temperature of concrete), temperature of mixing water, humidity, production date, production time, etc. These may be used alone or in combination of two or more.
Examples of data related to concrete transportation include the power load value inside the drum of a truck agitator, the volume, mass, and temperature of the concrete, the outside air temperature during transportation, transportation time (the time from the end of mixing to the end of transportation (unloading)), transportation distance, transportation date, transportation time, etc. These may be used alone or in combination of two or more.

The various learning input data described above may be obtained from an existing information management system in a concrete factory.
An information management system for a concrete factory is a system that centrally manages data necessary to control the production of concrete and data obtained during the production of concrete using computers, networks, etc.
Examples of concrete factories include ready-mix concrete factories, concrete product factories, etc. The various learning input data may be obtained from either an existing information management system in the ready-mix concrete factory or an existing information management system in the concrete product factory, or may be obtained from both information management systems.
Examples of information management systems include systems for managing the composition information of materials used in ready-mixed concrete at ready-mixed concrete factories (information that is disclosed as information indicating the quality of the concrete when the ready-mixed concrete is shipped), systems for setting the measurement values of stored materials (the amount of each material added per batch) and for controlling ready-mixed concrete manufacturing equipment such as mixers, and systems for managing information such as the transportation conditions when transporting manufactured ready-mixed concrete using agitator trucks, etc. (for example, ``Sky One II'' (manufactured by Pacific Systems Co., Ltd.), a ready-mixed concrete distribution system that utilizes GPS).
The above-described information management system is an example, and the method of acquiring various data is not particularly limited.
In this specification, "ready-mixed concrete" refers to concrete that has fluidity before hardening. The term "concrete" includes both concrete that has fluidity before hardening (ready-mixed concrete) and concrete after hardening.

Examples of learning output data regarding the finishability of concrete include data regarding the work performed during finishing of concrete when the concrete is actually produced, data regarding the quality of the concrete after finishing, and data regarding the fluidity of the concrete before finishing (slump value, slump flow value, etc.).
Finishing concrete refers to finishing the surface of the concrete using a trowel or similar tool after the concrete has been poured.
Examples of data related to work during concrete finishing include qualitative evaluation of finishability, finishing work time, cumulative value of finishing work time and number of people working on finishing, and compaction performance (time spent using a vibrator).

The qualitative evaluation of finishability is an evaluation based on the visual inspection and manual sensation of the concrete finishability (ease of finishing the concrete surface using a trowel, etc.) by the worker who actually finished the concrete surface. For example, the finishability may be evaluated on a five-point scale: very hard, hard, good, soft, and very soft.
Furthermore, each of the five-level evaluations may be replaced with a number. For example, very hard may be replaced with "-2," hard may be replaced with "-1," good may be replaced with "0," soft may be replaced with "1," and very soft may be replaced with "2."

Examples of data related to the quality of concrete after finishing include visual evaluation of the finished condition of the concrete, dimensions (actual measured values of the concrete product after hardening), rebound degree measured using a test hammer (for example, rebound degree measured in accordance with "JIS A 1155:2012 (Method for measuring rebound degree of concrete)"), and air permeability coefficient according to the Trent method.
The visual evaluation of the condition of the concrete finish is a visual evaluation of the appearance of the concrete after it has actually been finished. For example, a good appearance is given an "A," junk is given a "B," cracks are given a "C," surface bubbles are given a "D," laitance is given a "E," and cross-sectional defects are given an "F." The appearance of the concrete may be evaluated as, for example, "A,""B and C," etc.

The machine learning in the present invention is performed according to a conventional machine learning method known in the art, using multiple pieces of training data that are combinations of training input data including image data and training output data related to the finishability of concrete.
The learning input data and learning output data used as learning data are data such as images and actual measurements obtained when concrete is actually produced as samples for learning data.
The number of samples for training data varies depending on the types of training input data and training output data required, but from the viewpoint of being able to predict quality with higher accuracy, it is preferably 4 or more, more preferably 6 or more, and particularly preferably 8 or more.

From the viewpoint of being able to predict quality with higher accuracy, the number of learning times is preferably 1,000 or more, more preferably 8,000 or more, and particularly preferably 10,000 or more.
In one learning session, it is not necessary to use all of the data obtained from the sample, and only a portion of the data may be used.
The selection of learning data to be used in one learning is not particularly limited, and the learning data may be selected in the order in which they are arranged from top to bottom according to a specific condition (for example, the order in which they were photographed), or may be selected randomly. Furthermore, when there are multiple samples for the learning data, it is preferable to select the data so that at least one piece of data from each sample is included in the learning data.
Furthermore, the number of image data obtained from one sample is preferably 1,000 or more, more preferably 2,000 or more, and particularly preferably 2,500 or more, from the viewpoint of being able to predict quality with higher accuracy.
In addition, in this specification, "machine learning" refers to learning by machines (particularly computers) alone, without the intervention of human thought.

Prediction input data including image data is input into a prediction model created by machine learning, and prediction output data regarding the finishability of concrete is output from the prediction model.The prediction output data can then be used to predict the quality of the concrete.
From the viewpoint of being able to predict quality with higher accuracy, the prediction input data may further include data on the power load value of the mixer during mixing, data on the concrete mix conditions, data on the desired concrete quality, data on cement, data on concrete materials other than cement, data on mixing means and methods, data on the environment during mixing, and data on concrete transportation. These data may be used alone or in combination of two or more. These data are data obtained in real time during concrete production.

The details of the image data and the like used as the input data for prediction are the same as those of the image data and the like used as the input data for learning described above.
The details of the prediction output data relating to the finishability of concrete are the same as those of the learning output data relating to the finishability of concrete described above.
Furthermore, from the viewpoint of being able to centrally manage and control concrete quality predictions in real time and to periodically retrain the prediction model, it is preferable that at least one type of data included in the above-mentioned learning input data and prediction input data is obtained from an information management system at the concrete factory (at least one of a ready-mix concrete factory and a concrete product factory).

From the viewpoint of predicting the finishability of concrete with higher accuracy, standardized data may be used as at least one of the data used as the learning input data and prediction input data described above.
In addition, when a specific type of data in the training input data is standardized, the same type of prediction input data as the standardized data must also be standardized. For example, when a design slump value used as one type of training input data is standardized, the design slump value used as prediction input data must also be standardized before being input into the prediction model.
Furthermore, data that can be standardized is numerical data (data that can be expressed as a numerical value).
Data normalization can be performed using the following procedure.
Let n be the number of training data, and x be the data to be used as training input data and to be standardized. Then, the mean value μ and standard deviation σ of x _i (where i is an integer from 1 to n) are calculated using the following formulas (1) and (2).
The data to be standardized is data arbitrarily selected from the learning input data, and may be one type or two or more types.

Next, x _i is standardized using the mean value μ, the standard deviation σ, and the following formula (3) to obtain standardized data Z.
Standardized data is a numerical value with a mean of 0 and a variance of 1. By standardizing each of multiple types of data with different units or digits (for example, design slump value (cm), design air content (%), maximum coarse aggregate dimension (mm), unit coarse aggregate amount (kg/m ³ )), it is possible to align data with different units or digits. This makes it easier to compare the impact of each data on the prediction output data, and facilitates the selection of learning data to use when creating a prediction model.
In addition, when predicting the quality of concrete using a prediction model, it is necessary to standardize the prediction input data of the same type as the standardized data (data arbitrarily selected from the learning input data).
In this case, the mean value μ and standard deviation σ used for standardizing the prediction input data are calculated using the learning input data.

From the viewpoint of predicting the finishability of concrete with higher accuracy, it is preferable to use standardized image data as the image data included in the training input data and the prediction input data. By using standardized image data, it is possible to reduce the influence of changes in the brightness value of the entire image (for example, changes in the brightness of the entire screen due to differences in the shooting time or shooting conditions of the image data) on the creation of the prediction model.
Examples of methods for standardizing image data include the following.
The image data to be standardized is grayscale image data. If the captured image data is a color image, the color image must be converted into a grayscale image. The method for converting a color image into a grayscale image is not particularly limited, and a common method such as the NTSC weighted average method can be used using the functions of commercially available image software.

Next, the pixel values (luminance values) of the pixels that make up the grayscale image are standardized.
Specifically, if the image data included in the learning input data is image data consisting of 30 images of 64 x 64 pixels, the average value μ and standard deviation σ are calculated using the pixel values of all pixels that make up the image data (64 x 64 x 30 = 122,880 pixel values) and the above formulas (1) and (2) (in formula (1), n is 122,880, and x is the pixel value (brightness value)).
Next, the x _i can be standardized using the mean value μ, the standard deviation σ, and the equation (3) to obtain the standardized data Z.

According to the concrete quality prediction method of the present invention, in concrete production, by inputting image data of the concrete being produced into a prediction model created in advance, it is possible to quickly predict the finish quality of the concrete to be obtained with high accuracy without relying on human judgment.
In particular, in at least one of the following cases, when the design slump value of the target concrete is set small from the beginning (for example, when the slump value is set to 15 cm or less, preferably 10 cm or less, more preferably 5 cm or less), and when the measured slump value of the produced concrete is small (for example, when the slump value is 15 cm or less, preferably 10 cm or less, more preferably 5 cm or less), the fluidity of the concrete decreases, making finishing difficult, and the reliability of the slump value as an indicator for judging finishing processes such as surface finishing and compaction of the concrete decreases.
According to the concrete quality prediction method of the present invention, when using learning data (a combination of learning input data and learning output data) obtained from concrete whose design slump value is set small from the beginning (particularly set to a value within the range of 0 to 5 cm) and prediction input data, it is possible to predict with higher accuracy the finishability of concrete with a small stamp value, which has traditionally been difficult to predict accurately, and it is easy to make adjustments in the finishing process that are carried out in response to variations in the finishability of concrete.
In particular, since concrete product factories often produce low-slump concrete, the concrete quality prediction method of the present invention is suitable as a method for predicting the quality of concrete in concrete product factories.
Furthermore, since the prediction of concrete finishability can be performed for each of the multiple image data acquired during production, multiple output data (e.g., qualitative evaluation of finishability) can be obtained from the multiple image data. Among the multiple output data obtained, data that is clearly different from the others can be eliminated, or an average value calculated from the multiple output data obtained can be used as the output data for prediction.

If it is predicted that the finishability of the resulting concrete will not satisfy the desired finishability of the concrete, appropriate measures such as changing the concrete production conditions can be taken to produce concrete efficiently and stably.
For example, if the qualitative evaluation of finishability obtained from the predictive model differs significantly from the target qualitative evaluation of finishability, it is possible that there is an abnormality in the quality of the cement, aggregate, admixtures, etc. used as concrete materials, or that the measurement values of these materials were incorrect, and the manufacturing process must be reviewed immediately.It is also possible that the set value of the surface water of the aggregate differs from the actual value, and a correction value can be calculated before the aggregate is discharged from the mixer, and an appropriate amount of additional water can be injected into the mixer based on the correction value.
Furthermore, image data continuously captured of the mixing of concrete ingredients in a mixer may be input to the prediction model as input data for prediction, and the resulting output data for prediction may be displayed on a display device (e.g., a display) together with the image data. By displaying the data in real time, for example, it is possible to more quickly respond by changing the concrete production conditions while checking the output data for prediction and the state of the concrete in the mixer.
In addition, if a qualitative evaluation of finishability is predicted periodically and the predicted value of the qualitative evaluation of finishability deviates slightly from the target value, this may be due to a fluctuation in the surface water content of the aggregate, and appropriate measures such as reviewing the set value of the surface water content can be taken.
Furthermore, the accuracy of predictions may be improved by accumulating the prediction input data and data on the finishability of actual concrete used to predict the quality of concrete, and periodically re-learning the prediction model using the accumulated data.

Furthermore, by connecting a computer that controls concrete production with a computer used to implement the concrete quality prediction method of the present invention, the control system can be automated.
Furthermore, various data relating to concrete production at multiple factories may be transmitted via the Internet and centrally managed and controlled in real time at one location using the concrete quality prediction method of the present invention.

[Preparation of Concrete 1 to 20]
Cement, fine aggregate (mountain sand), coarse aggregate A (crushed stone No. 5), and coarse aggregate B (crushed stone No. 6) were put into a twin-shaft mixer and mixed dry, after which water was added and mixed to produce concrete (ready-mixed concrete) 1 to 20. The amount of each material was adjusted to the unit amount shown in Table 1.
The concrete had a target slump value of 3.0 cm, and the mix design of each material was carried out so as to achieve this slump value.
In addition, when preparing the concrete, a video camera was installed above the twin-shaft mixer to capture video images of the materials being mixed.
For each concrete produced, the actual slump value was measured in accordance with "JIS A 1101:2014 (Concrete slump test method)".
In addition, a qualitative evaluation of the finishability of each concrete (evaluation based on the worker's sense of visual inspection when finishing the surface with a trowel) was conducted with very hard being "-2", hard being "-1", good being "0", soft being "1", and very soft being "2".
The results are shown in Table 1.

[Example 1]
Concrete samples 1, 3 to 7, 11, and 13 to 15 were selected as samples for training data.
For each sample, 450 (15 × 30) still images capturing the entire interior of the twin-shaft mixer were obtained from a 15-second video sequence, starting one minute after water was poured into the mixer and mixing began, at a frame rate of 30 frames per second. From each still image, image data (256 × 256 pixels) was extracted from the approximate center of the twin-shaft mixer, between one and the other rotating shafts, within which concrete material likely appeared when the images were taken during mixing. The images were extracted from the following areas: (1) portions of both rotating shafts; (2) at least one of the tips of the blades attached to the rotating shafts (either or both of the rotating shafts); and (3) at least one of the blades attached to the rotating shafts (either or both of the rotating shafts) (see the area enclosed by the line in Figure 1(a)).
The power load value of the twin-shaft mixer was already stable one minute after the start of mixing, and remained stable until mixing was completed.

1(b), a total of 24 256 x 256 pixel image data pieces (each piece of image data cut at a different position) were cut out from positions shifted by 5 pixels vertically and horizontally (up and down and left and right) based on the position where the image data was cut out. That is, a total of 25 image data pieces were cut out from one still image, and 11,250 (30 x 15 x 25) image data pieces were obtained per sample, for a total of 112,500 (11,250 x 10) image data pieces.
The 25 pieces of image data were image data for the range in which the ready-mixed concrete material located nearby could be reflected for all of the above sections (1) to (3).
Next, each image data was reduced to an image size of 64 x 64 pixels, and 90,000 image data, or 80% of the 112,500 images used as learning input data, were used, and the remaining 20%, or 22,500 images, were used as learning test data.

The obtained 64×64 pixel image data was converted into a grayscale image by converting each pixel constituting the image data into 256 gradations.
The converted image data (grayscale image) was subjected to a standardization process, and the image data after standardization was used as input data for learning. The actual measured slump value of the sample from which the image data used as input data for learning was obtained, and the qualitative evaluation of the concrete finishability (evaluation based on the worker's visual inspection or sense during manual work, with very hard being ``-2'', hard being ``-1'', good being ``0'', soft being ``1'', and very soft being ``2''; data related to the concrete finishability) were used as output data for learning.
Using the learning data consisting of a combination of the learning input data and the learning output data, machine learning of the prediction model was performed to obtain a trained prediction model.
For machine learning, we used "TensorFlow" and conducted training using a six-layer convolutional neural network. Note that the combination of a convolutional layer and a pooling layer is counted as one layer.
The number of learning times was 500,000, and the number of image data (learning input data) input in each learning was 50 (randomly selected). In addition, the least squares method was used as the error function in the learning.

As samples for validation data, concrete samples 2, 8 to 10, 12, and 16 to 20 that were not used for learning were selected.
For each sample, 450 still images showing the entire interior of the twin-shaft mixer were obtained from a 15-second video sequence starting one minute after water was poured into the mixer and mixing began, with a frame rate of 30 frames per second. From each still image, image data (256 x 256 pixels) was extracted from the range in which the ready-mixed concrete materials located nearby could be reflected for all of the above sections (1) to (3). Next, based on the position where the image data was extracted, a total of eight 256 x 256 pixel image data (image data with different extraction positions) were extracted from positions shifted by 10 pixels vertically and horizontally (up and down and left and right). In other words, a total of nine image data were extracted from one still image, resulting in 11,250 images per sample, for a total of 112,500 images.

Each of the obtained image data was reduced to an image size of 64×64 pixels, and then each pixel constituting the image data was converted into a grayscale image with 256 gradations.
The converted image data (grayscale image) was subjected to a standardization process, and the standardized image data was input as prediction input data into the trained prediction model.The prediction model then output, as prediction output data regarding the finishability of concrete, a predicted slump value for the sample from which the image data used as learning input data was obtained, and a qualitative evaluation of the finishability of concrete (prediction output data regarding the finishability of concrete).
For each of the verification data samples (a total of 10 samples: concrete 2, 8-10, 12, and 16-20), the prediction input data (11,250 image data) obtained from the samples was input into the prediction model to calculate the average value of the predicted slump (prediction output data). Next, the actual measured slump value of the sample was compared with the average value, and the average value was considered to be correct (within the acceptable range) if it was within ±0.25 cm of the actual measured value. The accuracy rate (the percentage of samples in which the average value was within ±0.25 cm of the actual measured value of slump) obtained from the verification data samples (10 samples) was 80%.
In addition, for each of the verification data samples (a total of 10 samples: concrete samples 2, 8-10, 12, and 16-20), the prediction input data (11,250 image data) obtained from the samples was input into the prediction model, and the most frequent values (10) of the qualitative evaluation of the concrete finishability were obtained as prediction output data.
The accuracy rate of the obtained data on finishability (the percentage of samples for which the prediction was correct (i.e., the predicted rating (-2, -1, 0, 1, or 2) was the same as the actual rating (-2, -1, 0, 1, or 2)) out of all samples)) was 90%.
From Example 1, it can be seen that according to the present invention, the finishability of concrete can be predicted with high accuracy using image data.

[Example 2]
In the actual production of concrete at a concrete product factory, the finishability of concrete was predicted using the concrete quality prediction method described below.
The concrete materials used were cement (ordinary Portland cement, moderate-heat Portland cement, or low-heat Portland cement), fine aggregate (a mixture of crushed sand and pit sand), coarse aggregate (crushed stone 2005), and admixtures (at least one selected from air-entraining agents, air-entraining water-reducing agents, and high-performance air-entraining water-reducing agents).
The type and amount of each material was determined appropriately so that the water-cement ratio of the concrete was within the range of 30 to 40%, and the slump value of the resulting concrete was within the range of 2 to 4 cm. The materials were mixed using a twin-shaft mixer (capacity 5 m ³ ), with the mixing time within the range of 60 to 120 seconds, and the total amount of each material within the range of 1.00 to 3.25 m ³ .

Seventy batches of concrete with various blending ratios were produced under the above conditions, and training data was obtained from the batches.
Specifically, a video camera was installed on top of the twin-shaft mixer to capture video images of the materials being mixed during the concrete production process. The captured video images were displayed on a monitor in the monitoring room via a video signal distributor and were also saved as video image data on an SD card in the recording device.
The video was taken for 15 seconds, starting 15 seconds before the end of mixing (just before the concrete was discharged from the twin-shaft mixer) and continuing for 15 seconds until the end of mixing. The number of frames per second of the video was 30, and 450 (30 x 15) still images showing the entire interior of the twin-shaft mixer were obtained. From each still image, image data (512 x 512 pixels) was extracted from an area that could potentially capture concrete material located near the mixing blades fixed to the rotating shaft inside the twin-shaft mixer.
The above range was determined by checking multiple consecutive still images over time and determining that the concrete surface reflected in the images had significant movement, and this range was the area near the stirring blades.
The power load value of the twin-shaft mixer was already in a stable state when the video recording started, and remained stable until the mixing was completed.

Furthermore, using the position where the image data was extracted as a reference, 24 pieces of 512 x 512 pixel image data (image data extracted at different positions) were extracted from positions shifted by 5 pixels in the vertical and horizontal directions (up and down and left and right directions). That is, 25 pieces of image data were extracted in total from one still image, and 11,250 (30 x 15 x 25) pieces of image data per batch were obtained, for a total of 787,500 (11,250 x 70) pieces of image data (70 batches of image data).
In addition, for all of the above-mentioned parts, the 25 pieces of image data were image data from a range in which concrete material located near the mixing blades fixed to the rotating shaft inside the twin-shaft mixer could be reflected.
Next, each image data was reduced to an image size of 64 x 64 pixels, and the resulting 64 x 64 pixel image data was converted into a grayscale image by converting each pixel (picture element) constituting the image data into 256 gradations.
The converted image data (grayscale images) were standardized, and 630,000 images, equivalent to 80% of the standardized image data (787,500 images), were used as input data for training. The remaining 20%, or 157,500 images, were used as test data to confirm the reliability of the prediction model obtained after machine learning.

The concrete after production was discharged from the mixer, and the discharged concrete was sampled. The actual slump values of 70 batches were measured in accordance with "JIS A 1101:2014 (Concrete slump test method)," and the actual measured values were used as learning output data or test data.
In addition, qualitative evaluation of the finishability of 70 batches of concrete (evaluation based on the worker's visual inspection or sense during manual work, with very hard being ``-2'', hard being ``-1'', good being ``0'', soft being ``1'', and very soft being ``2''; data regarding the finishability of concrete) was used as learning output data or test data.
The training data was a combination of training input data (image data) and training output data (measured slump value and qualitative evaluation of concrete finishability) of the batch from which the image data used as the training input data was obtained.
The video data in the SD card was saved to another computer, and then the computer was used to perform machine learning of the prediction model using the training data, thereby obtaining a trained prediction model.
For machine learning, TensorFlow was used, and training was performed using a seven-layer convolutional neural network. In training, the least squares method was used as the error function.
The number of learning times was set to 500,000, and the number of image data (learning input data) input in each learning was set to 50 (randomly selected).

For verification data, 26 new batches of concrete were produced under the same conditions as the 70 batches described above, and video images were taken in the same manner as the 70 batches of concrete described above. From the video images, 450 (30 x 15) still images were obtained per batch.
From each still image, one image data was extracted in the same manner as for the 70 batches described above, and then, using the position where the image data was extracted as a reference, eight 512 x 512 pixel image data pieces (each image data piece extracted at a different position) were extracted from positions shifted by 10 pixels in the vertical and horizontal directions (up and down and left and right directions). In other words, a total of nine image data pieces were extracted from one still image, resulting in 4,050 (30 x 15 x 9) pieces of image data per batch, for a total of 105,300 (4,050 x 26) pieces of image data (26 batches of image data).
The image data was image data for all of the above sections (1) to (3) in the range in which concrete material located nearby could be reflected.
Next, each image data was reduced to an image size of 64×64 pixels, and each pixel constituting the image data was converted into a grayscale image with 256 gradations.
The converted image data (grayscale image) was subjected to standardization processing, and the standardized image data was used as verification data.
Furthermore, the actual slump values of the ready-mixed concrete after production and the finishability of the concrete were qualitatively evaluated in the same manner as for the 70 batches of ready-mixed concrete described above.

The image data of the 26 batches from which the verification data was obtained were input as prediction input data into the trained prediction model, and the prediction value of the concrete slump and the qualitative evaluation of the concrete finishability were output from the prediction model as prediction output data related to the concrete finishability. Next, the average value of the predicted value of the concrete slump and the mode value of the qualitative evaluation of the concrete finishability were obtained for each batch.
The actual measured slump value was compared with the average predicted slump value (the average predicted slump value of the concrete output as prediction output data), and batches whose average predicted slump value was within ±0.25 cm of the actual measured value were considered correct (within the acceptable range). The accuracy rate was 88%. Furthermore, the accuracy rate of the most common value for the qualitative evaluation of concrete finishability (the percentage of cases where the prediction was correct (i.e., the predicted classification (-2, -1, 0, 1, or 2) was the same as the actual sample classification (-2, -1, 0, 1, or 2))) was 92%.
These results show that the present invention makes it possible to predict the finishability of concrete with high accuracy using image data.

Claims (7)

1. A method for predicting concrete quality using a predictive model, comprising:
The prediction model is created by machine learning using a plurality of pieces of training data, which are combinations of training input data including image data and training output data related to the finishability of concrete,
inputting prediction input data including image data into the prediction model, outputting prediction output data relating to the finishability of concrete from the prediction model, and predicting the quality of concrete using the prediction output data ;
the learning data is obtained from concrete having a design slump value within a range of 0 to 5 cm, and the design slump value of the concrete to be predicted is within a range of 0 to 5 cm,
A method for predicting the quality of concrete , wherein the learning output data and the prediction output data are data relating to work during concrete finishing . 2. The method for predicting quality of concrete according to claim 1, wherein the learning output data and the prediction output data are qualitative evaluations of the finishability of concrete. The concrete quality prediction method according to claim 1 or 2, wherein the learning output data and the prediction output data include data relating to the quality of the finished concrete. The concrete quality prediction method according to any one of claims 1 to 3, wherein at least one type of data used as the learning input data and the prediction input data is standardized data. the standardized data is image data included in the learning input data and the prediction input data,
5. The method for predicting concrete quality according to claim 4, wherein the image data is a grayscale image, and the pixel values of the pixels constituting the grayscale image are standardized. The image data is a plurality of image data continuously captured to capture a state in which concrete materials are mixed in a mixer for mixing the materials,
The method for predicting quality of concrete according to any one of claims 1 to 5, wherein the image data and the prediction output data are displayed on a display means. The concrete quality prediction method according to any one of claims 1 to 6, wherein at least one type of data included in the learning input data and the prediction input data is obtained from an information management system at a concrete factory.