JP6680936B1 - Method of predicting quality of ready-mixed concrete - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method capable of predicting the quality of ready-mixed concrete in a short time and with high accuracy without depending on human sense or experience. A method for predicting the quality of ready-mixed concrete using a prediction model, wherein the prediction model is a combination of learning input data including image data and learning output data concerning the quality of ready-mixed concrete. It is created by machine learning using multiple learning data.The prediction input data including image data is input to the prediction model, and the prediction output data related to the quality of fresh concrete is output from the prediction model. A method for predicting the quality of ready-mixed concrete for predicting the quality of concrete. [Selection diagram] Figure 1

Description

  The present invention relates to a method for predicting quality of ready-mixed concrete.

Green concrete is produced by kneading the respective materials in a mixer for a predetermined time based on a composition adjusted to the desired quality of the green concrete.
On the other hand, even if the mixing and mixing time are the same, the quality of the produced ready-mixed concrete also fluctuates due to the fluctuation of the surface water ratio of the aggregate and the fluctuation of the outside temperature. Therefore, usually, it is judged whether or not it is possible to manufacture ready-mixed concrete having a desired quality by visually confirming a monitor image taken by a camera installed in the mixer and using the power load value of the mixer. The manufacturing conditions are adjusted accordingly.
However, since the above judgment is made by a human being on the basis of feeling and experience, the judgment may be different depending on the operator, and it is difficult to manufacture ready-mixed concrete of stable quality.

  As a system capable of performing quality control of concrete without being an expert, for example, in Patent Document 1, a measuring unit that measures vibration and sound generated from a mixer during kneading of high-fluidity concrete, and the above-mentioned measurement A high-fluidity concrete quality control system is provided, which comprises a spectrum analysis means for spectrally analyzing the vibrations and sounds, and a quality estimation means for estimating the quality of the high-fluidity concrete from the result of the spectrum analysis.

JP, 2003-177115, A

  An object of the present invention is to provide a method capable of predicting the quality of ready-mixed concrete in a short time and with high accuracy without depending on the human sense or experience.

The present inventor, as a result of diligent study to solve the above problems, as a prediction model created by machine learning using a plurality of combinations of input data including image data and output data regarding the quality of fresh concrete, image data, The present invention has been completed by finding that the above object can be achieved by a method of predicting the quality of ready-mixed concrete by using output data obtained by inputting input data including the above.
That is, the present invention provides the following [1] to [9].
[1] A method for predicting the quality of fresh concrete using a prediction model, wherein the prediction model is a combination of learning input data including image data and learning output data relating to the quality of fresh concrete. It is created by machine learning using a plurality of learning data, inputting the prediction input data including image data to the above prediction model, and outputting the prediction output data regarding the quality of fresh concrete from the above prediction model. A method for predicting the quality of ready-mixed concrete, characterized by predicting the quality of ready-mixed concrete using the output data for prediction.
[2] After the image data is mixed in the mixer for kneading the raw concrete material, water and the raw concrete material other than water are started to be kneaded to stabilize the power load value of the mixer, and then the image is taken. The method for predicting the quality of ready-mixed concrete according to the above [1], which is obtained.
[3] The image data is obtained from each of a plurality of images taken continuously in a mixer for kneading the raw concrete material with water and the raw concrete material other than water being kneaded. Obtained, a plurality of image data, for each of the plurality of image data, input prediction data including image data, input to the prediction model, from the prediction model output data for prediction of the quality of green concrete The method for predicting quality of fresh concrete according to the above [1] or [2], which predicts the quality of the fresh concrete by using the moving average value of the plurality of output data for prediction that has been output.
[4] The image data is obtained by cutting a range in which raw concrete material located near the stirring blade fixed to the rotating shaft of the mixer may be reflected from the image captured inside the mixer. The method for predicting quality of ready-mixed concrete according to the above [2] or [3].
[5] The image data is obtained from the image taken inside the mixer, and the following (1) to (3) of the kneading member of the kneading member including the rotating shaft of the mixer and the stirring blade fixed to the rotating shaft is used. The method for predicting quality of ready-mixed concrete according to any one of [2] to [4] above, wherein a range in which the material of ready-mixed concrete located in the vicinity of all parts may be reflected is cut out.
(1) At least a part of the rotating shaft (2) Tip part of the stirring blade (3) Fixed part of the rotating shaft and the stirring blade [6] The learning input data and the prediction input data are further described as above. The method for predicting quality of fresh concrete according to any one of the above [1] to [5], including data relating to a power load value of a mixer when a raw concrete material is mixed.

[7] The method for predicting quality of fresh concrete according to any one of [1] to [6], wherein the machine learning is learning using a convolutional neural network.
[8] The raw material according to any one of [1] to [7], wherein the learning output data and the prediction output data are data related to at least one of classification of slump and slump flow of fresh concrete. Concrete quality prediction method.
[9] The raw material according to any one of the above [1] to [7], wherein the learning output data and the prediction output data are data relating to a numerical value of at least one of slump and slump flow of green concrete. Concrete quality prediction method.
[10] The input data for learning and the input data for prediction further include data on mixing conditions of ready-mixed concrete, data on quality of desired ready-mixed concrete, data on cement, data on materials of ready-mixed concrete other than cement, The method according to any one of the above [1] to [9], which includes one or more kinds of data selected from data on kneading means and method, data on environment during kneading, and data on transportation of ready-mixed concrete. Method for predicting quality of ready-mixed concrete.

  By using the method for predicting quality of ready-mixed concrete of the present invention, it is possible to predict the quality of ready-mixed concrete in a short time and with high accuracy, and it is possible to efficiently and stably produce ready-mixed concrete of desired quality. .

In the first embodiment, the position (a) of a still image obtained by photographing the inside of the mixer from above the biaxial mixer and the position (a) of the image data cut out from the still image, and the position of the image data shifted by 5 pixels on the basis of the position ( It is a figure which shows b). It is a top view of a mixer (biaxial mixer). It is sectional drawing which shows the state which cut | disconnected in the direction perpendicular | vertical to the rotating shaft of the mixer (biaxial mixer) which mixes the raw concrete material. FIG. 9 is a diagram showing a still image obtained by photographing the inside of the mixer from above the biaxial mixer and a position of image data cut out from the still image in the second embodiment. In an Example, it is a figure which shows the relationship of the electric power load value at the time of starting mixing for 1 minute, and the electric power load value at the time of passing 2 minutes by a biaxial mixer.

The quality prediction method of green concrete of the present invention is a method for predicting the quality of green concrete by using a prediction model, and the prediction model is learning input data including image data and learning about the quality of green concrete. It is created by machine learning using multiple learning data, which is a combination of output data for prediction, and inputs prediction input data including image data to the prediction model, and outputs prediction for the quality of fresh concrete from the prediction model. Data is output, and the quality of green concrete is predicted using the output data for prediction.
The details will be described below.

The prediction model is created by machine learning.
Examples of learning methods used for machine learning include neural network, linear regression, decision tree, support vector regression, ensemble method, support vector machine, discriminant analysis, naive Bayes method, and nearest neighbor method. These may be used alone or in combination of two or more.
Among them, the neural network is preferable from the viewpoint that the quality can be predicted with higher accuracy.
The neural network is preferably a hierarchical neural network having one or more intermediate layers between the input layer and the output layer from the viewpoint that the quality can be predicted with higher accuracy.

Examples of the neural network include a convolutional neural network (CNN) such as a three-dimensional convolutional neural network (3DCNN) and a recurrent neural network (RNN). A memory (LSTM: Long Short-Term memory) neural network (an improved recursive neural network using LSTM) and the like can be mentioned.
Above all, a convolutional neural network (a neural network having a convolutional layer, a pooling layer or the like as an intermediate layer), which has excellent performance in the field of image recognition, is more preferable. According to the convolutional neural network, it is possible to detect a feature amount from image data, and use the feature amount to create a prediction model capable of performing classification or regression.
In the convolutional neural network, the number of layers formed by the combination of the convolutional layer and the pooling layer is preferably 2 or more, and more preferably 3 or more, from the viewpoint that prediction can be performed with higher accuracy.
Machine learning includes, for example, “TensorFlow” (“TENSORFLOW” is a registered trademark) which is a software library developed by Google, and “IBM Watson” (“IBM WATSON” which is a system developed by IBM). Is a registered trademark.) And the like.

The prediction model is created by machine learning using a plurality of learning data that is a combination of learning input data including image data and learning output data regarding the quality of green concrete.
Image data used as learning input data includes image data relating to the production of ready-mixed concrete. Specifically, image data showing how each material is mixed in a mixer for mixing the raw concrete material (hereinafter also simply referred to as “mixer”), and each material is mixed in the mixer. After obtaining the ready-mixed concrete, the image data showing how the ready-mixed concrete is put into the hopper from the mixer and the image data showing the appearance of the ready-mixed concrete being stirred in the drum of the track agitator (for example, the track agitator A monitor that displays a history of the power load value of the mixer when mixing the raw concrete material by shining light from the vicinity of the raw concrete input port using a light etc.) Examples of the image data include a photograph of a change in values over time, which is displayed visually using a graph or the like. These may be used alone or in combination of two or more.
Above all, from the viewpoint that the quality can be predicted with higher accuracy, image data obtained by photographing the state of kneading the respective materials in the mixer is preferable. The image data may be image data at any point from immediately after kneading each material to the end of kneading, but from the viewpoint of predicting quality early, when the degree of decrease in the power load value becomes moderate. The photographed one is preferable.

The number of image data used as learning input data is preferably 100 or more, more preferably 1,000 or more, still more preferably 10,000 or more, and further preferably from the viewpoint that the quality can be predicted with higher accuracy. It is 50,000 or more, particularly preferably 100,000 or more.
The image data may be moving image data or still image data. The image may be a two-dimensional image or a three-dimensional image.
The image data is captured by a camera appropriately installed in the mixer or around the mixer. For example, a camera is installed in the upper part of the mixer so that each of the mixed materials can be seen well.
For the purpose of making it easier to see the unevenness of each material being kneaded, light is applied from the side or diagonally above of each material being kneaded so that the shadow becomes darker. May be.
In addition, from the viewpoint that the quality of image data can be predicted with higher accuracy, the stirring blade (mixer blade) that is rotating in the mixer is located at a specific place that is arbitrarily determined in the mixer. It is preferable to do so. The arbitrarily specified specific place may be one place or two or more places.

The image data may be data obtained by cutting out an image in a specific range from an image taken in the mixer (for example, imaged so as to show the entire inside of the mixer).
From the viewpoint that the above specific range can be predicted with higher accuracy, it is possible to show the part where the behavior of the material is likely to appear when continuously shooting the state of mixing the raw concrete material in the mixer The range with a property is preferable. Hereinafter, description will be given with reference to FIGS.
As an example of the above range, a range in which the raw concrete material 3 located in the vicinity of the stirring blade 4 fixed to the rotating shaft 2 of the mixer 1 may be reflected from the image captured inside the mixer 1 (Fig. 2), refer to a portion surrounded by a rectangular frame composed of a thick line). This range is a range in which the movement (change) of the surface of the raw concrete material 3 reflected in an image (a plurality of still images continuous with time or moving images) when kneading the raw concrete material 3 is large. Therefore, it is suitable as learning input data.
Further, in the case where the mixing amount of the raw concrete material 3 is large, etc., from the viewpoint of obtaining the image data in which the surface movement of the raw concrete material 3 reflected in the image is large, the image data includes the stirring blade 4 when rotating. What is cut out when it comes to the front surface side (see the portion surrounded by a rectangular frame formed by a thick line in FIG. 3) is preferable.

In addition, as an example of a range in which a portion where the behavior of the material is likely to appear may be reflected, the following kneading member 5 of the kneading member 5 including the rotating shaft 2 of the mixer 1 and the stirring blade 4 fixed to the rotating shaft 2 is described below. For all the portions (1) to (3), there is a range in which the raw concrete material 3 located in the vicinity thereof may be reflected.
(1) At least a part of the rotating shaft 2 (2) Tip part of the stirring blade 4 (3) Fixed part of the rotating shaft 2 and the stirring blade 4 When there are a plurality of stirring blades 4, at least one stirring blade 4 It suffices if the tip portion and the fixed portion are reflected.
When the mixer 1 is a biaxial mixer, the raw concrete material 3 located between the one rotary shaft 2 and the other rotary shaft 2 and in the vicinity of a part of both rotary shafts 2 is reflected. A possible range is preferred.

When cutting out an image in a specific range as image data (used as input data for learning) from an image taken inside the mixer, a plurality of image data may be cut out from one image taken inside the mixer. .
The cut-out ranges of the plurality of image data may partially overlap each other. For example, with reference to the position where the image data is first cut out, another image data is moved from a position shifted in 1 to 10 pixel units in at least one of the vertical direction and the horizontal direction (vertical direction and horizontal direction). A plurality of image data may be obtained from one image by cutting out.

  The image captured in the mixer is a grayscale image in which 1 pixel is represented by a numerical value of 256 gradations from 0 to 255, or 1 pixel is R (from the viewpoint of making image data more suitable for machine learning such as convolutional neural network). You may convert red (red) G (green) and B (blue) into color images represented by numerical values of 256 gradations from 0 to 255.

Difference data obtained from two pieces of image data arbitrarily selected from two or more shot image data may be used as image data used as learning input data. For example, the difference data of the two image data captured when the mixer blades are positioned at two arbitrarily defined locations in the mixer may be used as the image data used as the learning input data.
In the present specification, difference data refers to image data obtained by comparing two image data and extracting only different parts.
Alternatively, a combination of a plurality of still images that are continuous over time may be used as image data. Such image data can be obtained by, for example, using commercially available image software, setting the opacity (or transparency) of each of a plurality of still images to about 30 to 70%, and then superimposing them. The number of still images to be superimposed is preferably 2 to 20 from the viewpoint of work efficiency and the like.
The image data is captured, for example, for 10 seconds with 10 frames per second. As a result, data of 100 still images can be obtained.

From the viewpoint that the quality can be predicted with higher accuracy, data regarding the power load value of the mixer during mixing (hereinafter, also simply referred to as “power load value”) may be used as the learning input data. Good.
Examples of data regarding the power load value include the power load value and its absolute value at a specific time point immediately after kneading each material until the end of kneading, and the maximum or minimum value of the power load value at the time of kneading. The amount of change in the power load value, the change pattern of the power load value, and the like are included. These may be used alone or in combination of two or more.
The number of data regarding the power load value may be one, but is preferably 2 or more, and more preferably 5 or more from the viewpoint that the quality can be predicted with higher accuracy.

Further, the image data (image data obtained by photographing the state of mixing the respective materials in the mixer) to be used as the input data for learning may be determined based on the data regarding the power load value. For example, from the viewpoint of predicting the quality with higher accuracy and at an early stage, when mixing of water and raw concrete material other than water is started and the power load value of the mixer becomes stable (power load It is preferable to use the captured image data after the time point when the change amount of the value becomes small).
The determination of the time when the power load value of the mixer becomes stable in the mixing of the raw concrete material varies depending on the water-cement ratio of the raw concrete, but for example, any one of the following conditions (1) to (3) It may be determined that the power load value of the mixer is stable after satisfying the condition.
(1) After the power load value has reached a value of about ± 20% from the power load value at the end of kneading (2) If the water-cement ratio of ready-mixed concrete is about 50-70%, water and other than water From the start of kneading of the raw concrete material (for example, when the raw concrete materials other than water are put into the mixer and kneaded, and then water is added to the mixer to start the kneading) After about a second has passed (3) If the water-cement ratio is made small (less than 50%) from the viewpoint of strength development, the above-mentioned period is delayed, and in high-strength concrete, water is mixed with raw concrete materials other than water. 1 to 10 minutes after the start of

  Further, based on the data on the power load value, two or more image data are captured, and the difference data obtained from the two image data arbitrarily selected from the captured two or more image data is input for learning. It may be used as image data used as data. For example, after the power load value becomes stable, shooting (for example, 30 (frame) shooting for 1 second) is performed for an arbitrary time, and two image data arbitrarily selected from the obtained plurality of image data are acquired. The obtained difference data may be used as image data used as learning input data.

Further, from the viewpoint of predicting the quality with higher accuracy, other data may be used as the learning input data.
Other data include data on mix conditions of ready-mixed concrete, data on quality of desired ready-mixed concrete, data on cement, data on materials of ready-mixed concrete other than cement, data on mixing method and method, at mixing time. The data on the environment and the data on the transportation of ready-mixed concrete. These may be used alone or in combination of two or more.
These various data include data related to numerical values and data related to classification.
Here, in the present specification, data regarding numerical values means data that can be expressed as specific numerical values, and data regarding classification is a specific formulation design, a specific type, a specific property, or It means data classified according to criteria such as a specific numerical range.

  Examples of data regarding the mixing conditions of green concrete include cement, fine aggregate, coarse aggregate, water, various admixtures (AE agent, water reducing agent, AE water reducing agent, high performance water reducing agent, and High-performance AE water reducing agent, superplasticizer, setting retarder, etc.), and mixing ratio of various admixtures (for example, the amount of admixture (100% by mass) with respect to 100% by mass of cement), and the items in the formulation table. , Water cement ratio, air amount, fine aggregate ratio, unit water amount, unit cement amount, unit fine aggregate amount, unit coarse aggregate amount, unit admixture amount, unit admixture amount and the like. These may be used alone or in combination of two or more.

Examples of data regarding the quality of the target fresh concrete include strength (nominal strength, compressive strength, bending strength, etc.), slump, slump flow, air content, chloride content in the design of the target fresh concrete. Crack resistance, dynamic elastic modulus, dynamic shear elastic modulus, dynamic Poisson's ratio, cured product void volume and void diameter distribution, durability, color tone and the like can be mentioned. These may be used alone or in combination of two or more.
Among them, slump and slump flow, which are more important as the quality of green concrete, are preferable.
The strength, slump, slump flow, air content, and chloride content of green concrete can be measured, for example, by the test method described in "JIS A 5308: 2019 (Ready Mixed Concrete)".

Examples of data on cement include the type, chemical composition, mineral composition, wet type f. Of cement used as a material for ready-mixed concrete. CaO, loss on ignition, Blaine specific surface area, particle size distribution, residual amount of sieving test, color tone, mineralogy and crystallographic properties of each mineral contained in cement, hemihydrate ratio of gypsum contained in cement, etc. , Data on whole cement; chemical composition, hydraulic modulus, residual amount of sieve test, Blaine specific surface area (fineness), loss on ignition, supply amount, auxiliary raw materials (special waste such as waste) Raw material) supply amount, blending silo storage amount (remaining amount), storage silo storage amount (remaining amount), and the current value of the cyclone located between the raw material mill and the blending raw material blending silo (cyclone rotation speed). , Which has a correlation with the speed of the raw material passing through the cyclone), or a time point before a predetermined time from the time of charging the kiln (for example, one time point 5 hours ago or 3 hours ago, Raw material of cement clinker at a plurality of time points such as four time points before, 5 hours before, and 6 hours before (a fine material for cement clinker extracted by a countercurrent air flow during transportation). Hereinafter, referred to as the cement clinker kiln raw material.) Chemical composition and its hydraulic hardness, and the chemical composition of the raw material obtained by mixing the cement clinker kiln raw material and auxiliary materials, hydraulic hardness, brane specific surface area, sieve Data on the raw material of cement clinker, such as test residual amount, decarboxylation rate, and water content; amount of cement clinker raw material inserted into kiln during firing of cement clinker, kiln rotation speed, outlet temperature, firing zone temperature , Cement clinker temperature, kiln average torque, O ₂ concentration, NO _X concentration, clinker cooler temperature, and preheater gas flow rate (pre-heat Data related to the firing conditions of cement clinker, such as those that correlate with the temperature of the water); crushing temperature, amount of water sprayed in the finishing mill, separator air flow, type of gypsum, amount of gypsum added, amount of cement clinker input, finishing Data on the grinding conditions of cement such as the number of revolutions of the mill, the temperature of the powder discharged from the finishing mill, the amount of powder discharged from the finishing mill, the amount of powder not discharged from the finishing mill; cement clinker, Mineral composition, chemical composition, wet f. CaO (free lime), and weight, crystallographic properties of each mineral contained in cement clinker (lattice constant, crystallite size, etc.), ratio of two or more minerals contained in cement clinker, Examples include data on cement clinker. These may be used alone or in combination of two or more.

Examples of data on raw concrete materials other than cement include the type, density, water absorption rate, water content, surface water rate, particle size distribution, maximum size, and grain of aggregate (fine aggregate or coarse aggregate). The shape, the type of admixture, the type of admixture and the like can be mentioned. These may be used alone or in combination of two or more.
Examples of data relating to the kneading means and method include the kind, type and capacity of the mixer, the kneading amount of the materials, and the kneading time. These may be used alone or in combination of two or more.
Examples of data regarding the environment during mixing are the temperature (outside air temperature, the temperature inside the mixer, the temperature of ready-mixed concrete), the temperature of each of the materials used such as mixing water, cement, and aggregate, and the storage location ( The temperature and humidity of (in the container) may be mentioned. These may be used alone or in combination of two or more.
Examples of data on the transportation of green concrete include the power load value in the drum of a truck agitator, the capacity of green concrete, the thickness of the driving member, the mass and temperature, the outside air temperature during transportation, and the transportation time (after mixing is finished, The time until the end of transportation (at the time of unloading), transportation distance, etc. can be mentioned. These may be used alone or in combination of two or more.

  Examples of output data for learning about the quality of ready-mixed concrete include slump, slump flow, air amount, unit water amount, temperature of fresh concrete immediately after mixing with a mixer or at the end of transportation using an agitator truck (at the time of unloading), Workability, separation status of each material that constitutes fresh concrete, fluidity, rheology-related values (plastic viscosity, yield value, etc.), strength after hardening (compressive strength, bending strength, Young's modulus, etc.) and durability (shrinkage, shrinkage, Freeze-thaw resistance, neutralization, etc.) and the like. These may be used alone or in combination of two or more.

Machine learning in the present invention is performed according to a conventionally known general method of machine learning using a plurality of learning data which is a combination of learning input data including image data and learning output data relating to the quality of green concrete. Be seen.
The learning input data and the learning output data used as the learning data are data such as images and measured values obtained when the fresh concrete is actually manufactured as samples for the learning data.
The number of samples for learning data varies depending on the types of input data for learning and output data for learning required, but is preferably 4 or more, more preferably from the viewpoint of predicting quality with higher accuracy. Is 6 or more, particularly preferably 8 or more.

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, from the viewpoint that the quality can be predicted with higher accuracy. 500 or more.
The number of times of learning is preferably 1,000 times or more, more preferably 8,000 times or more, and particularly preferably 10,000 times or more, from the viewpoint that the quality can be predicted with higher accuracy.
It is not necessary to use all the learning data obtained from the sample in one learning, and a plurality of learning data selected from all the learning data may be used.
The selection of the learning data used in one learning is not particularly limited, and the learning data arranged according to a specific condition (for example, the shooting order) may be selected in order from the top, or randomly. You may choose to. In addition, when there are a plurality of samples for learning data, it may be selected so that at least one learning data of each sample is included.
In addition, in the present specification, “machine learning” means that learning is performed only by a machine (in particular, a computer) without using human thinking.

In addition, a part of learning data (combination of input data for learning and output data for learning) (for example, 15 to 25% of all learning data) is not used in machine learning and is not used in machine learning. It may be used as test data for confirming the reliability of the created prediction model. Specifically, after creating a prediction model, input the prediction input data obtained from the test data to the prediction model, and obtain the prediction output data (prediction value) obtained from the prediction model and the test data. By comparing the measured values obtained, the reliability of the obtained prediction model can be estimated and measured.
Generally, machine learning is performed while appropriately changing the type of data used as learning data, the number of learning data, the number of times of learning, and the like until a prediction model with excellent reliability is obtained.

Prediction model created by machine learning, input the prediction input data including image data, after outputting the prediction output data about the quality of the fresh concrete from the prediction model, using the prediction output data of the fresh concrete The quality can be predicted.
From the viewpoint of being able to predict the quality with higher accuracy, as input data for prediction, data regarding the power load value of the mixer at the time of mixing, data regarding the mixing conditions of fresh concrete, and the quality of the desired fresh concrete Data, data on cement, data on raw concrete materials other than cement, data on means and method of kneading, data on environment at the time of kneading, data on transportation of green concrete, etc. may be used. These may be used alone or in combination of two or more. These data are data obtained in real time in the production of ready-mixed concrete.
Details of the image data and the like used as the prediction input data are the same as the image data and the like used as the learning input data described above.
The details of the output data for prediction regarding the quality of green concrete are the same as the above-mentioned output data for learning regarding the quality of green concrete.

According to the method for predicting the quality of ready-mixed concrete of the present invention, in the production of ready-mixed concrete, the prediction model created in advance, by inputting image data or the like of the ready-mixed concrete under production, the quality of the ready-mixed concrete obtained (for example, The slump of ready-mixed concrete) can be predicted with high accuracy and speed without relying on human judgment. The “quality of green concrete” includes the quality of the hardened body obtained by hardening the raw concrete.
Since the quality is predicted for each of the plurality of image data acquired at the time of manufacturing, it is possible to obtain a plurality of output data (for example, a predicted value of slump or the like) from the plurality of image data.
Further, the average value calculated from the obtained plurality of output data may be used as the prediction output data.
For example, the image data used as the input data for learning shows that a mixer for kneading raw concrete materials is mixed with water and raw concrete materials other than water, It is a plurality of image data obtained from each of the images, and for each of the plurality of image data, the prediction input data including the image data is input to the prediction model, and the prediction output regarding the quality of the ready-mixed concrete from the prediction model. After outputting the data, the quality of the fresh concrete may be predicted using the moving average value of the plurality of output data for prediction that has been output.
The moving average value may be an average value of all the prediction output data obtained using a plurality of image data, which is obtained from each of a plurality of images continuously (temporarily) photographed, Of a plurality of image data obtained from each of a plurality of images captured continuously, it was obtained using an arbitrary number (for example, 5) of image data obtained most recently from the time of predicting quality. , The average value of a plurality (arbitrary number) of prediction output data may be used.

Also, during the kneading, when outputting a predicted value for each of a plurality of image data obtained by shooting continuously, such as a plurality of output data obtained (for example, the predicted value of the slump etc.) If output data that is clearly different from other output data is obtained, the data may be excluded from the prediction output data.
As an example of the method of excluding the obtained output data from the prediction output data, the standard deviation σ of the data set is calculated, and the output data outside the numerical value range of ± σ or ± 2σ from the average value of the data set is calculated. Is determined as output data that is clearly different from other output data, and is excluded.
If the quality of the obtained ready-mixed concrete is not expected to meet the desired quality of the ready-mixed concrete, by taking appropriate measures such as changing the production conditions of the ready-mixed concrete, the ready-mixed concrete can be efficiently and stably Can be manufactured.

For example, if the predicted slump value obtained from the prediction model is significantly different from the target slump value (hereinafter also referred to as the “target value”), the cement or aggregate used as the raw concrete material Since it is possible that the quality of admixtures, etc. was abnormal or the measured values of these materials were incorrect, it is necessary to immediately review the manufacturing process. It is also assumed that the set value of the surface water of the aggregate is different from the actual value, so a correction value is calculated before it is discharged from the mixer, and an appropriate amount of additional water is added to the mixer based on the correction value. You can also pour water.
In addition, if the slump prediction is performed regularly and the obtained slump prediction value deviates slightly from the target value, it is possible that the surface water ratio of the aggregate fluctuates. You should take measures such as reviewing.

Further, the control system can be automated by connecting a computer for controlling the production of ready-mixed concrete to a computer used for implementing the method for predicting quality of ready-mixed concrete of the present invention. For example, there is a method of automatically setting the surface water based on the variation in the difference between the predicted value and the target value of the slump.
Furthermore, even if various data in the production of ready-mixed concrete in a plurality of factories are transmitted via the Internet and the quality prediction method of the ready-mixed concrete of the present invention is used to centrally manage and control in one place. Good.
Further, when the quality of the green concrete is predicted by using the quality prediction method of the present invention when the green concrete is manufactured in an actual fresh concrete factory, the real-time image may be used for the prediction. As a specific prediction method, there is a method in which a camera and a computer are connected to extract a kneading image during the production of ready-mixed concrete in real time while a computer incorporating a learning model created in advance is used to perform prediction in real time.
In addition, the input image for learning is the kneading image of the fresh concrete obtained when manufacturing the fresh concrete in the actual ready-mixed concrete factory, and the slump measured after kneading is used as the output data for learning. By using the learning data that is a combination of the data and the output data for learning, the learning model created in advance is retrained at any time to obtain the latest learning model, so that the accuracy of the predicted value can be further improved.

[Preparation of green concrete 1-26]
Cement, fine aggregate (mountain sand), coarse aggregate A (crushed stone No. 5), coarse aggregate B (crushed stone No. 6) are charged into a biaxial mixer, and after kneading, water is added to mix them. Then, ready-mixed concrete 1-26 was produced. The amount of each material was set to the unit amount shown in Table 1.
In addition, the ready-mixed concrete 1-9, 10-19, 20-26 made the target slump value 8 cm, 12 cm, and 18 cm, respectively, and designed the composition of each material so that this slump value could be obtained. It is a thing.
Here, the target slump value is set to 8 cm, and the target slump value is set to 12 cm with the ready-mixed concrete (fresh concrete 1 to 9) in which each material is mixed and designed so that the slump value can be obtained. Then, it is possible to obtain the slump value by setting the fresh concrete (fresh concrete 10 to 19) in which the respective materials are mixed and designed so as to obtain the slump value to B and setting the target slump value to 18 cm. The ready-mixed concrete (ready-mixed concrete 20 to 26) for which the mixing design of each material was performed in this manner was classified into C, respectively.
In addition, when producing the ready-mixed concrete, a video camera was installed above the biaxial mixer, and a moving image was taken when each material was mixed.
For each of the prepared ready-mixed concrete, the measured value of the slump was measured according to “JIS A 1101: 2014 (concrete slump test method)”. The results are shown in Table 1.
In the production of green concrete 1 to 26, a graph showing the relationship between the power load value 1 minute after the start of mixing and the power load value 2 minutes after the start of mixing is shown in FIG.

[Example 1]
Fresh concrete 1, 2, 3, 10, 11, 13, 21, 23, and 26 were selected as samples for learning data.
For each sample, water was poured into the mixer, and 2 minutes after mixing was started, 5 minutes from the moving image, the entire interior of the biaxial mixer was shown with 30 frames per second. 150 (30 × 5) still images were obtained. From each still image, it is the range of the center part of the biaxial mixer and the position where one rotation axis and the other rotation axis are located. When shooting while mixing, (1) rotation of both At least one of the tip portions of a plurality of stirring blades fixed to a part of the shaft, (2) the rotary shaft (either one or both of the two rotary shafts), or (3) the rotary shaft (two rotary shafts). (Either one or both of the above) and at least one of the fixed portions of the plurality of stirring blades, image data (256) in a range in which the raw concrete material located in the vicinity thereof may be reflected. (× 256 pixels) were cut out (see the portion surrounded by the line in FIG. 1A).
The power load value of the twin-screw mixer was already in a stable state 2 minutes after the kneading was started, and the power load value was stable until the kneading was completed.

Further, as shown in FIG. 1B, 256 × 256 from a position shifted in units of 5 pixels in the vertical direction and the horizontal direction (vertical direction and horizontal direction) with reference to the position where the image data is cut out. A total of 24 pieces of image data of pixels were cut out (image data having different cut-out positions). That is, 25 pieces of image data in total are cut out from one still image, and 3,750 (30 × 5 × 25) pieces of image data for one sample, 33,750 (3,750 × 9) pieces of image data in total. Got
The 25 pieces of image data were image data in a range in which the raw concrete material located in the vicinity of all the portions (1) to (3) could be reflected.
Then, each image data is reduced to an image size of 64 × 64 pixels, and 27,000 image data, which is 80% of 33,750 images used as learning input data, is used and left as learning test data. 6750 sheets, which is 20% of the above, were used.

The obtained 64 × 64 pixel image data is used as a grayscale image obtained by converting each pixel of the image data into 256 gradations, and the converted image data is used as learning input data and learning input data. The target slump value (8 cm, 12 cm, or 18 cm: data related to classification of slump of ready-mixed concrete) of the sample from which the image data was obtained was used as learning output data.
Machine learning of a prediction model was performed using learning data composed of a combination of the learning input data and the learning output data to obtain a learned prediction model.
"TensorFlow" was used for machine learning, and learning was performed using a 5-layer convolutional neural network. In addition, the convolutional layer and the pooling layer are counted as one "combination of the convolutional layer and the pooling layer".
The number of times of learning was 10,000, and the number of image data (learning input data) input in one learning was 50 (randomly selected).

As the samples for the verification data, green concrete 4, 8, 9, 16, 17, 19, 20, 24, and 25, which were not used for learning, were selected.
For each sample, water was poured into the mixer, and 2 minutes after mixing was started, 5 minutes from the moving image, the entire interior of the biaxial mixer was shown with 30 frames per second. 150 still images were obtained. From one still image, image data (256 × 256 pixels) in a range in which the raw concrete material located near the portions (1) to (3) may be reflected was cut out. Next, a total of 8 pieces of image data of 256 × 256 pixels are shifted from the position shifted by 10 pixels in the vertical direction and the horizontal direction (vertical direction and horizontal direction) based on the position where the image data is cut out. (Image data in which the cut-out positions are all different) Cut out. That is, a total of 9 pieces of image data were cut out from one still image, and 1,350 pieces of image data were obtained for each sample, and a total of 12,150 pieces of image data were obtained.
Image data for verification is input as prediction input data to the learned prediction model, and from the prediction model, as output data for prediction related to quality of fresh concrete, data related to classification of slump of fresh concrete (A, B or C) was output.
It should be noted that the data regarding the classification of the slump is whether the slump value of 8 cm, 12 cm, or 18 cm is the target (A, B, or C) of the composition design of the sample from which the image data for verification is obtained. To predict.
Correct answer rate of 12,150 prediction output data obtained from the image data (the correct prediction rate (that is, the predicted classification (A, B or C) and the actual sample classification (A, B or C) The ratio of the same as C)) was 86%.

[Example 2]
Pour water into the mixer and start mixing for 1 minute. From the moving image for 15 seconds from the point where 15 minutes have passed, the number of frames per second is set to 30 and the entire interior of the biaxial mixer is reflected 450 sheets. After extracting the still image of No. 1, after obtaining 101,250 (30 × 15 × 25 × 9) image data from the still image, a learned prediction model is created in the same manner as in Example 1. The prediction model was used to output data regarding slump classification.
The correct answer rate of 12,150 prediction output data obtained from the image data was 88%.
Further, comparing Example 1 and Example 2, in Example 1, the image taken from the point 2 minutes after starting the kneading was used, whereas in Example 2, the kneading was performed. The image taken from 1 minute after the start of is used.
In connection with this, it can be seen from FIG. 4 that the power load value at the time point 2 minutes after the kneading was started and the power load value at the time point 1 minute were the same. The power load value was stable until the kneading was completed.
The correct answer rate (88%) of the prediction output data of Example 2 is equal to or higher than the correct answer rate (86%) of the predictive output data of Example 1, and Example 2 is compared with Example 1. , It is possible to predict the quality of ready-mixed concrete at an earlier stage.
[Example 3]
A learned prediction model was created in the same manner as in Example 2 except that the number of times of learning was set to 20,000, and the prediction model was used to output data regarding slump classification.
The accuracy rate of 12,150 prediction output data obtained from the image data was 92%.
From Examples 1 to 3, it can be seen that according to the present invention, classification of fresh concrete (classification by slump) can be predicted with high accuracy using image data.

[Example 4]
Fresh concrete 2-4, 6-8, 10-12, 14-15, 18, 21, and 23 were selected as samples for learning data.
For each sample, water was poured into the mixer, and 15 minutes after the kneading was started, one minute passed, and from the moving image for 15 seconds, the number of frames per second was set to 30, and the entire interior of the biaxial mixer was shown. We obtained 450 still images. In the same manner as in Example 1, 25 pieces of image data were cut out from each still image, and 11,250 sheets were obtained for each sample, and a total of 157,500 (11,250 × 14) pieces of image data were obtained.
The power load value of the twin-screw mixer was already in a stable state 1 minute after the kneading was started, and the power load value was stable until the kneading was completed.

The obtained image data was used as learning input data, and the actual measurement value of the slump of the sample from which the image data used as learning input data was obtained was used as learning output data.
Machine learning of a prediction model was performed using learning data composed of a combination of the learning input data and the learning output data to obtain a learned prediction model.
“TensorFlow” was used for machine learning, and learning was performed using a 6-layer convolutional neural network. In learning, the least squares method was used as the error function.
The number of times of learning was 500,000, and the number of image data (learning input data) input in one learning was 50 (randomly selected).

Fresh concrete 5, 9, 16, 20, 22, and 26 were selected as samples for verification data.
For each sample, water was poured into the mixer, and 15 minutes after the kneading was started, one minute passed, and from the moving image for 15 seconds, the number of frames per second was set to 30, and the entire interior of the biaxial mixer was shown. We obtained 450 still images. From all the above (1) to (3) parts, the 256 × 256 pixel image data including the range where the raw concrete material located in the vicinity may be reflected is cut out from one still image. It was Next, a total of 8 pieces of image data of 256 × 256 pixels are shifted from the position shifted by 10 pixels in the vertical direction and the horizontal direction (vertical direction and horizontal direction) based on the position where the image data is cut out. (Image data in which the cut-out positions are all different) Cut out. That is, a total of 9 pieces of image data were cut out from one still image, and 4,050 pieces of image data were obtained per sample, that is, a total of 24,300 pieces of image data were obtained.

For each image data for verification, input the image data as prediction input data to the above-mentioned learned prediction model, and from the prediction model, as prediction output data regarding the quality of fresh concrete, the predicted value of slump (of the fresh concrete The data related to the numerical value) was output.
The measured value of the slump of the sample for which the image data was obtained was compared with the predicted value of the slump obtained from the image data, and the case where the predicted value was within ± 1 cm of the measured value was regarded as the correct answer. The accuracy rate of the predicted value of 24,300 slumps obtained from the image data (the ratio of the predicted value within the range of ± 1 cm of the actual measured value) was 91%.
Further, for each of the samples for verification data (fresh concrete 5, 9, 16, 20, 22, and 26), 4,050 pieces of image data obtained from the samples were input to the prediction model. The average of the predicted values of slump was calculated. Next, the measured value of the slump of the sample was compared with the above average value, and the case where the average value was within ± 1 cm of the measured value was regarded as the correct answer. The accuracy rate of the average value of the predicted values of the six slumps obtained from the sample for the verification data (the rate of the average value being within ± 1 cm of the measured value) was 100%.
From Example 4, it can be seen that according to the present invention, the numerical value of the slump of ready-mixed concrete can be predicted with high accuracy using image data.

[Example 5]
In the production of ready-mixed concrete that is actually performed in a ready-mixed concrete factory, the quality of ready-mixed concrete was predicted using the method for predicting quality of ready-mixed concrete described below.
Raw concrete materials include cement (normal Portland cement, medium heat Portland cement, or low heat Portland cement), fine aggregate (made by mixing crushed sand and mountain sand), coarse aggregate (crushed stone 2005), and An admixture (at least one selected from an AE agent, an AE water reducing agent, and a high-performance AE water reducing agent) was used.
The type and amount of each material is such that the water-cement ratio of the fresh concrete is in the range of 34 to 55%, the slump value of the obtained raw concrete is in the range of 15 to 25 cm, and the slump flow of the obtained raw concrete is obtained. The value was appropriately determined so as to fall within the range of 42 to 63 cmm. Further, the kneading of the material, biaxially using a mixer (capacity 5 m ^3), kneading time is in the range of 60 to 120 seconds, the total amount in the range of 2.00～4.25M ³ of each material Within

70 batches of ready-mixed concrete with various blends were produced under the above-mentioned conditions, and learning data were obtained from the batches.
Specifically, when producing ready-mixed concrete, a video camera was installed above the biaxial mixer, and a moving image of kneading and mixing each material was taken using the camera. The captured moving image was displayed on the monitor in the monitoring room by the video signal distributor and stored as moving image data in the SD card of the recording device.
The moving image was photographed for 15 seconds from 15 seconds before the end of the mixing (immediately before discharging the fresh concrete from the biaxial mixer) to the end of the mixing. The number of frames per second of the moving image was set to 30, and 450 (30 × 15) still images showing the entire interior of the biaxial mixer were obtained. From each still image, image data (512 × 512 pixels) was cut out from a range in which the raw concrete material located near the stirring blade fixed to the rotating shaft in the biaxial mixer might be reflected.
The above range is a range in which a plurality of continuous still images are confirmed over time, and it is determined that the movement of the surface of the green concrete reflected in the image is large (see a portion surrounded by a white line in FIG. 4). The range was in the vicinity of the stirring blade.
The power load value of the biaxial mixer was already in a stable state at the start of capturing a moving image, and the power load value was stable until the kneading was completed.

In addition, a total of 24 pieces of 512 × 512 pixel image data are shifted from the position shifted by 5 pixels in the vertical and horizontal directions (vertical direction and horizontal direction) with reference to the position where the image data is cut out. (Image data in which the cut-out positions are all different) It was cut out (see FIG. 1B). That is, a total of 25 pieces of image data are cut out from one still image, 11,250 (30 × 15 × 25) sheets per batch, and a total of 787,500 (11,250 × 70) pieces of image data ( 70 batches of image data) were obtained.
Note that the image data of 25 sheets is the image data of the range in which the raw concrete material located near the stirring blade fixed to the rotating shaft in the twin-screw mixer may be reflected in all the above-mentioned parts. Met.
Next, each image data was reduced to an image size of 64 × 64 pixels, and 630,000 image data corresponding to 80% of 787,500 images were used as learning input data. In addition, 157,500 sheets corresponding to the remaining 20% were used as test data for confirming the reliability of the prediction model obtained after machine learning.

The produced fresh concrete was discharged from the mixer and loaded into a truck agitator. Then, the fresh concrete discharged from the truck agitator was sampled, and the actual measurement value of 70 batches of slump was measured in accordance with “JIS A 1101: 2014 (concrete slump test method)”. The output data for learning or the test data was used.
Further, for the above-mentioned green concrete, the measured value of the slump flow for 70 batches was measured according to "JIS A 5308: 2014 (Ready Mix Concrete)", and the measured value was used as output data for learning or test data. .
A combination of the learning input data (image data) and the learning output data (actual values of slump and slump flow) of the batch in which the image data used as the learning input data was obtained was used as the learning data.
After storing the moving image data in the SD card in another computer, machine learning of the prediction model was performed using the above learning data using the computer to obtain a learned prediction model.
“TensorFlow” was used for machine learning, and learning was performed using a 7-layer convolutional neural network. In learning, the least squares method was used as the error function.
The number of times of learning was 500,000, and the number of image data (learning input data) input in one learning was 50 (randomly selected).

As verification data, 26 batches of fresh concrete were newly produced under the same conditions as the above-mentioned 70 batches, and a moving image was taken in the same manner as the 70 batches of fresh concrete described above, and 1 batch was taken from the moving image. Therefore, 450 (30 × 15) still images were obtained.
Similar to the above-described 70 batches, one image data is cut out from each still image, and further, 10 points in the vertical direction and the horizontal direction (up and down direction and left and right direction) based on the position where the image data is cut out. A total of 8 pieces of image data of 512 × 512 pixels (image data having different cut-out positions) were cut out from positions shifted in pixel units. That is, a total of 9 pieces of image data are cut out from one still image, and 4,050 (30 × 15 × 9) sheets per batch, 105,300 (4,050 × 26) pieces of image data ( 26 batches of image data) were obtained.
The image data was image data in a range in which the raw concrete material located in the vicinity of all the portions (1) to (3) could be reflected.
Next, each image data was reduced to an image size of 64 × 64 pixels, and the obtained image data was used as verification data.
In addition, the actual measurement value of slump and the actual measurement value of slump flow of the fresh concrete after production were measured in the same manner as the above-mentioned 70 batches of fresh concrete.

The 26 batches for which the verification data were obtained were classified into A to D shown in Table 2 based on the measured values of the slump obtained from each batch.
The image data obtained from the 26 batches is input as prediction input data to the learned prediction model, and from the prediction model, as prediction output data regarding the quality of fresh concrete, the predicted value of slump (the numerical value of fresh concrete is calculated). Data).
For batches classified into each classification shown in Table 2, the measured value of slump and the predicted value of slump were compared, and the predicted value of slump in each classification was a value within ± 2.0 cm of the measured value. As a correct answer (within an allowable range), the correct answer rate (number of correct batches / total number of classified batches) of each classification was calculated.
In addition, the accuracy rate of each classification was calculated in the same manner except that the correct value (within the allowable range) was set when the predicted value of the slump was a value within ± 2.5 cm of the measured value. Table 2 shows the results.

Further, the 26 batches were classified into E to H shown in Table 3 based on the measured values of the slump flow obtained from each batch.
The image data obtained from the 26 batches is input as prediction input data to the learned prediction model, and from the prediction model, as a prediction output data regarding the quality of fresh concrete, a predicted value of slump flow (of the fresh concrete The data related to the numerical value) was output.
For batches classified into each classification shown in Table 3, the measured values of slump flow and predicted values of slump flow are compared, and in the batches classified into classifications E to G, the predicted value of slump flow is ± of the measured value. The correct answer (within the allowable range) was a batch with a numerical value within 7.5 cm, and the predicted slump flow in the batch classified into class H was a correct value within ± 2.5 cm of the measured value. As (within the allowable range), the correct answer rate of each classification was calculated. The results are shown in Table 3.

  It can be seen from Tables 2 to 3 that the accuracy rate is 72 to 99%, and that according to the present invention, the numerical values of slump and slump flow of green concrete can be predicted with high accuracy using image data.

1 Mixer 2 Rotating shaft 3 Raw concrete material 4 Stirring blade 5 Kneading member

Claims (9)

A method for predicting the quality of ready-mixed concrete using a prediction model, comprising:
The prediction model is a mixer for kneading the raw concrete material, in a state of kneading water and raw concrete material other than water, learning input data including photographed image data and raw concrete It is created by machine learning using multiple learning data that is a combination of learning output data related to quality,
Input the prediction input data including the image data to the prediction model for each of a plurality of image data captured continuously in a state where the water and the raw concrete material other than water are mixed in the mixer. Then, after outputting the output data for prediction of the quality of the fresh concrete from the above prediction model, the quality of the fresh concrete is predicted by using the moving average value of the output data for prediction output. Concrete quality prediction method.   The image data is taken in the mixer for mixing the raw concrete material, after the mixing of water and the raw concrete material other than water is started and the power load value of the mixer is stabilized, The method for predicting quality of ready-mixed concrete according to claim 1. The image data is an image of the inside of the mixer, and a range in which the raw concrete material located near the stirring blade fixed to the rotating shaft of the mixer may be reflected is cut out. Item 3. A method for predicting the quality of ready-mixed concrete according to Item 1 or 2 . From the image obtained by photographing the inside of the mixer, the image data includes all the following parts (1) to (3) of the kneading member including the rotating shaft of the mixer and a stirring blade fixed to the rotating shaft. The method for predicting quality of fresh concrete according to any one of claims 1 to 3 , wherein a range in which the material of fresh concrete located in the vicinity is reflected is cut out.
(1) At least a part of the rotating shaft (2) Tip part of the stirring blade (3) Fixed part of the rotating shaft and the stirring blade The learning input data and the prediction for the input data, further, ready-mixed concrete according to any one of claims 1 to 4 containing data related to the power load value of the mixer during kneading of the material of the raw concrete Quality prediction method. The machine learning, ready-mixed concrete quality prediction method according to any one of claims 1 to 5, which is a learning using a neural network convolution. The learning output data and the output data for the prediction, the data relating to at least one of classification of raw concrete slump and slump flow, raw concrete quality prediction according to any one of claims 1 to 6 Method. The quality prediction of ready-mixed concrete according to any one of claims 1 to 7 , wherein the learning output data and the prediction output data are data relating to numerical values of at least one of slump and slump flow of fresh concrete. Method. The input data for learning and the input data for prediction are further data on mixing conditions of ready-mixed concrete, data on quality of desired ready-mixed concrete, data on cement as a material of ready-mixed concrete, materials of ready-mixed concrete other than cement about the data, mixing means and data relating to the method of raw concrete material, claims comprising one or more data selected regarding environment during kneading of the raw concrete material data, and from data relating to the transportation of ready-mixed concrete Item 9. A method for predicting the quality of green concrete according to any one of items 1 to 8 .