JP7455660B2 - How to diagnose or predict concrete deterioration - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act 74th Cement Technology Conference Lecture Abstracts 2020, Publication date: May 18, 2020

The present invention relates to a method for diagnosing or predicting concrete deterioration.

As the Ministry of Land, Infrastructure, Transport and Tourism has announced that periodic inspections of social infrastructure will be legalized, social interest in the maintenance and management of social infrastructure is increasing. In connection with this, the importance of diagnostic technology for concrete structures is also increasing.
The causes of deterioration of concrete structures are wide-ranging, including alkaline aggregate reaction (ASR), salt damage, and corrosion of reinforcing steel due to carbonation. Regardless of the type of deterioration, the cost of maintaining concrete structures can be reduced by diagnosing, preventing, and repairing the cause at an early stage.

There are a wide variety of methods for diagnosing the deterioration of concrete structures, depending on the type of deterioration to be diagnosed. However, even if diagnosis is carried out, the progression of deterioration is often only noticed when concrete spalling is discovered during periodic inspections.
Furthermore, although it is possible to easily diagnose the type of concrete deterioration by checking for the occurrence of cracks in concrete structures and the presence of rust, the characteristics of each type of deterioration are manifested in changes in appearance. It takes time. Furthermore, it is difficult to quantify the timing at which changes in appearance appear. Furthermore, if cracks actually occur in concrete, this may lead to a decrease in the load-bearing capacity of the concrete structure.
Therefore, it is necessary to diagnose the type of concrete deterioration at an early stage and take countermeasures.

As a method for early detection of concrete deterioration, Patent Document 1 describes a method for detecting concrete deterioration using a pattern appearing in an image of the maximum principal strain distribution obtained through the following steps (A) and (B). A method for early detection of concrete deterioration is described.
(A) Image acquisition step of acquiring digital images of the concrete surface to be acquired over time (B) Calculating strain using the digital image correlation method based on the digital image, and based on the strain, distribution of the maximum principal strain Maximum principal strain distribution acquisition process to obtain

JP 2019-74339 Publication

An object of the present invention is to create a predictive model that can diagnose or predict the deterioration of concrete, and to provide a method that can diagnose or predict the deterioration of an actual structure based on the predictive model.

As a result of intensive studies to solve the above problems, the inventors of the present invention discovered the steps of producing a test specimen, acquiring an initial image of the surface of the test specimen, and acquiring a surface strain image of the surface of the specimen. A process of creating a strain distribution on the surface of the specimen using a digital image correlation method, and a process of creating learning data that is a combination of learning input data including strain distribution data and learning output data including data regarding concrete deterioration. The present invention has been completed by discovering that the above object can be achieved by a method for creating a predictive model for concrete deterioration, which includes a step of creating a predictive model for diagnosing or predicting concrete deterioration using machine learning.
That is, the present invention provides the following [1] to [14].

[1] A method for creating a predictive model for concrete deterioration, which comprises creating a predictive model for predicting concrete deterioration by performing the following steps (A) to (E).
(A) A test body production process in which a concrete test body is prepared as a learning sample. (B) An initial image acquisition process of the test body in which an initial image of the surface of the above test body is acquired. (C) Surface strain has become apparent. Surface strain image acquisition step (D) of acquiring a surface strain image of the surface of the test body Using a digital image correlation method, the strain distribution on the surface of the test body is determined from the initial image and the surface strain image. Data creation step (E) for creating strain distribution of the test specimen: Through machine learning using learning data that is a combination of learning input data including the above strain distribution data and learning output data including data related to concrete deterioration. , a machine learning process that creates a predictive model to predict concrete deterioration.

[2] In the above step (D), the following steps (d-1) to (d-5) are performed to obtain a plurality of strain distribution data consisting of image data. How to create a predictive model for deterioration.
(d-1) Using a digital image correlation method, calculate the strain value at each coordinate on the surface of the test piece from the initial image and the surface strain image to obtain a plurality of strain value data. Creation step (d-2) After sorting the plurality of strain numerical data obtained in step (d-1) in ascending order based on the strain numerical values, among the plurality of rearranged strain numerical data, strain Randomly select one strain numerical data from the strain numerical data located in 10 to 30% of the total number of the plurality of strain numerical data, counting from the data with the minimum value, and calculate the strain of the selected strain numerical data. The value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the numerical value is the lower limit of the strain value, and from the strain numerical data located in 70 to 90% of the total number of the plurality of strain numerical data A numerical range in which one piece of numerical strain data is arbitrarily selected, and the upper limit of the strain value is the value obtained by adding the value obtained by multiplying the interquartile range by 1.5 to the strain numerical value of the selected numerical strain data. Setting step (d-3) Strain numerical data selection step (d-4) Step (d- Image data normalization step (d-5) in which the strain numerical data selected in step 3) is normalized to obtain one piece of image data having gradation; multiple pieces of image data are cut out from the above one piece of image data. Strain distribution data creation step [3] in which a plurality of strain distribution data consisting of image data is obtained [3] The following step (B-1) is performed between the above step (B) and the above step (C). The method for creating a predictive model for concrete deterioration according to [1] or [2].
(B-1) A test piece in which a process is performed on the test piece for which an initial image has been acquired to make surface strain apparent on the surface of the test piece, and surface strain is made obvious on the surface of the test piece. manifestation process

[4] By performing the following steps (F) to (I) using the prediction model obtained by the concrete deterioration prediction model creation method described in any one of [1] to [3] above, A concrete deterioration diagnosis or prediction method for diagnosing or predicting deterioration of target concrete.
(F) Initial image acquisition step of the target concrete, which acquires an initial image of the surface of the concrete that is the target of diagnosis or prediction. (G) Target concrete, which acquires a surface strain image of the surface of the concrete where surface strain has become apparent. Surface strain image acquisition step (H) Using a digital image correlation method, create strain distribution data on the concrete surface from the initial image and the surface strain image (I) Diagnostic or Prediction process

[5] In the above step (H), the following steps (h-1) to (h-5) are performed to obtain a plurality of strain distribution data consisting of image data. How to create a predictive model for deterioration.
(h-1) Using the digital image correlation method, calculate the strain value at each coordinate on the surface of the test piece from the initial image and the surface strain image to obtain a plurality of strain value data. Creation step (h-2) After sorting the plurality of strain numerical data obtained in step (h-1) in ascending order based on the strain numerical values, among the plurality of rearranged strain numerical data, strain Randomly select one strain numerical data from the strain numerical data located in 10 to 30% of the total number of the plurality of strain numerical data, counting from the data with the minimum value, and calculate the strain of the selected strain numerical data. The value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the numerical value is the lower limit of the strain value, and from the strain numerical data located in 70 to 90% of the total number of the plurality of strain numerical data A numerical range in which one piece of numerical strain data is arbitrarily selected, and the upper limit of the strain value is the value obtained by adding the value obtained by multiplying the interquartile range by 1.5 to the strain numerical value of the selected numerical strain data. Setting step (h-3) Strain numerical data selection step (h-4) of selecting strain numerical data having a strain numerical value greater than or equal to the lower limit value and less than or equal to the upper limit value from the plurality of numerical strain data data. Image data normalization step (h-5) in which the strain numerical data selected in step 3) is normalized to obtain one image data having gradation. Multiple image data are cut out from the above one image data. Strain distribution data creation step [6] where a plurality of strain distribution data consisting of image data is obtained [6] The following step (F-1) is performed between the above step (F) and the above step (G). The method for diagnosing or predicting concrete deterioration according to [4] or [5].
(F-1) Processing is performed on the concrete for which the initial image has been acquired to make surface strain obvious on the surface of the concrete, and surface strain is made obvious on the surface of the concrete, making the target concrete obvious. Processing process

[7] The above-mentioned [4] wherein the data regarding the deterioration of concrete output in the diagnosis or prediction step includes data regarding the type of deterioration of the target concrete, and the type of deterioration of the concrete is diagnosed or predicted based on the data. The method for diagnosing or predicting concrete deterioration according to any one of ~[6].
[8] The data related to the deterioration of concrete output in the diagnosis or prediction step includes data related to the progress of deterioration of the target concrete, and based on the data, the progress of deterioration of the target concrete is diagnosed or predicted [4] ] - [6] The method for diagnosing or predicting concrete deterioration according to any one of [6].
[9] The data regarding the deterioration of concrete output in the above diagnosis or prediction process includes data regarding the repair method for deterioration of the target concrete, and based on the data, the optimal repair method for the deterioration of the target concrete is diagnosed or predicted. The method for diagnosing or predicting concrete deterioration according to any one of [4] to [6] above.

[10] A method for creating a predictive model for concrete deterioration, which comprises creating a predictive model for predicting concrete deterioration by performing the following steps (J) to (L).
(J) A test body preparation process (K) in which a concrete test body is prepared as a learning sample; A surface crack image acquisition process of the test body (in which a surface crack image of the surface of the above-mentioned test body in which surface cracks have become apparent is acquired) L) Develop a predictive model for predicting concrete deterioration through machine learning using learning data that is a combination of learning input data including surface crack image data and learning output data including data regarding concrete deterioration. Create a machine learning process [11] By performing the following steps (M) to (N) using the prediction model obtained by the concrete deterioration prediction model creation method described in [10] above, A method for diagnosing or predicting concrete deterioration that diagnoses or predicts deterioration of concrete.
(M) Obtaining a surface crack image of the concrete surface to be diagnosed or predicted in which surface cracks have become apparent. (N) Diagnostic input including data of the surface crack image. Diagnosis or prediction step of inputting data into the above prediction model and outputting diagnostic output data including data regarding concrete deterioration from the above prediction model to diagnose or predict concrete deterioration.

[12] The above-mentioned [11] wherein the data regarding the deterioration of concrete output in the diagnosis or prediction step includes data regarding the type of deterioration of the target concrete, and the type of deterioration of the concrete is diagnosed or predicted based on the data. A method for diagnosing or predicting concrete deterioration as described in .
[13] The data related to the deterioration of concrete output in the diagnosis or prediction step includes data related to the progress of deterioration of the target concrete, and the progress of deterioration of the target concrete is diagnosed or predicted based on the data [11] ] Method for diagnosing or predicting concrete deterioration.
[14] The data regarding the deterioration of concrete output in the above diagnosis or prediction step includes data regarding the repair method for deterioration of the target concrete, and based on the data, the optimal repair method for the deterioration of the target concrete is diagnosed or predicted. The method for diagnosing concrete deterioration according to the above [11].

According to the concrete deterioration prediction model creation method of the present invention, a prediction model that can diagnose or predict concrete deterioration at an early stage is created, and the deterioration of concrete (for example, a concrete structure) can be diagnosed or predicted. I can do it.

It is an image of the strain distribution of test body A produced in Example 1. It is an image of the strain distribution of test body B produced in Example 1. 2 is an image of the strain distribution of concrete that is the target of diagnosis in Example 1. This is an image showing the distribution of surface cracks.

The concrete deterioration prediction model creation method of the present invention is a method of creating a prediction model for predicting concrete deterioration by performing the following steps (A) to (E).
(A) A test body production process in which a concrete test body is prepared as a learning sample. (B) An initial image acquisition process of the test body in which an initial image of the surface of the above test body is acquired. (C) Surface strain has become apparent. Surface strain image acquisition step (D) of acquiring a surface strain image of the surface of the test body Using a digital image correlation method, the strain distribution on the surface of the test body is determined from the initial image and the surface strain image. Data creation step (E) for creating strain distribution of the test specimen: Through machine learning using learning data that is a combination of learning input data including the above strain distribution data and learning output data including data related to concrete deterioration. , a machine learning process to create a predictive model for predicting concrete deterioration Each process will be explained in detail below.

[(A) Test body production process]
This step is a step of producing a concrete test body as a learning sample.
The concrete test specimen is used to obtain learning data used in the machine learning process (process (E)) described below. The above-mentioned test specimen may be one in which concrete has already deteriorated, or may be one in which concrete has not deteriorated.
The above-mentioned test body may be produced after appropriately adjusting the material composition, manufacturing method, curing conditions, service conditions, etc. of the above-mentioned test body so that the desired deterioration easily occurs.
In addition, test specimens with different material compositions, manufacturing methods, curing conditions, service conditions, etc., test specimens with different degrees of deterioration progress, and after deterioration has occurred, various repair methods are used to repair the deterioration that has occurred. It is preferable to prepare test specimens under various conditions, such as a specimen under various conditions.
Furthermore, the number of types of deterioration occurring in one concrete test specimen may be one type or two or more types.
The concrete test specimen may be one in which the deterioration factor (deterioration cause) is unknown, but from the viewpoint of being able to predict the deterioration of various types of concrete with higher accuracy, it is necessary to use one in which the deterioration factor is clear, or Preferably, the deterioration factor has been identified through various analyses.
Alternatively, a core may be taken from a concrete structure in service and used as the concrete test specimen. Moreover, it is preferable to identify the deterioration factor by analyzing the core before using it as a concrete test specimen.

The above test specimens were tested using alkaline aggregate reaction (ASR), delayed formation of ettringite (DEF), etc., in order to create a more accurate predictive model by obtaining learning data (described later) regarding the deterioration of various types of concrete. It is preferable to prepare a plurality of samples for each type of concrete deterioration, such as frost damage, fire damage, salt damage, and reinforcing steel corrosion due to carbonation.
The shape of the test specimen is not particularly limited, and may be one that imitates the shape of a concrete structure that is actually manufactured.
The above-mentioned test specimen may be one manufactured to obtain learning data, but a part (core portion) or the whole of an actually manufactured concrete structure may be used as the test specimen.

Examples of concrete include, but are not limited to, ordinary concrete, watertight concrete, hot concrete, cold concrete, mass concrete, fluidized concrete, high fluidity concrete, high strength concrete, low heat generation concrete, expansive concrete, low shrinkage concrete, and fiber reinforced concrete. , lightweight concrete, and polymer concrete. Further, the concrete may be unreinforced concrete, reinforced concrete, or prestressed concrete.

[(B) Initial image acquisition process of test specimen]
This step is a step of acquiring an initial image (usually a digital image) of the surface of the concrete test specimen produced in step (A).
Although the means for acquiring the initial image is not particularly limited, examples thereof include a line sensor scanner and the like. In addition, depending on the acquisition method, in the process of acquiring surface strain images of the test specimen described later, from the viewpoint of making surface strains more obvious and enabling analysis using the digital image correlation method to be performed with higher accuracy, concrete A geometric pattern (marker) may be drawn on the surface of the test specimen using a spray or the like.
Furthermore, in order to obtain a good image, the surface portion of the concrete test specimen may be polished in advance. In addition, when drawing a geometric pattern on the surface part of a concrete test body, it is preferable to perform the said polishing before drawing a geometric pattern (marker).

[(B-1) Test specimen revealing treatment process]
This step (B-1) is a step that can be added arbitrarily, and is a step performed between step (B) and step (C).
This step is a process for making surface strain apparent on the surface of the concrete specimen whose initial surface image was obtained in step (B) (hereinafter also referred to as "revealing treatment"). This is a step in which surface strain is made apparent on the surface of the test specimen.
By performing this process, surface strain can be made obvious on the surface of a concrete specimen with a small degree of surface strain (no obvious surface strain has occurred), or a concrete test in which surface strain has become obvious. By increasing the degree of manifestation of surface strain on the body, it is possible to create a prediction model that can predict concrete deterioration with higher accuracy.
Moreover, a prediction model can be created in a short period of time after creating a concrete test specimen.
The manifestation treatment is not particularly limited as long as it can increase and manifest the strain (surface strain) appearing on the surface of the concrete test specimen, but for example, heating means such as a heat gun can be used. methods to apply temperature changes (temperature history) to the test specimen using physical means, methods to apply stress to the test specimen by physical means, methods to leave the test specimen still in an environment that accelerates the deterioration of the test specimen, etc. can be mentioned.
The revealing treatment does not require excessive treatment that would destroy the surface of the concrete test specimen (for example, generate surface strain on the order of several thousand μm); It is sufficient to make it large enough so that it does not cause any damage.

[(C) Surface strain image acquisition process of test specimen]
This step is a step of acquiring a surface strain image of the surface of a concrete specimen in which surface strain has become apparent (occurred).
Note that the presence or absence of surface strain manifestation (occurrence) can be determined by creating data on the strain distribution on the surface of the concrete specimen using a digital image correlation method.
The concrete test specimen in which surface strain has become apparent refers to the concrete specimen prepared in step (A) in which surface strain has already become apparent (e.g., the concrete specimen that is in service and has been collected from a concrete structure in which surface strain has become apparent). Concrete test specimen), a concrete specimen prepared in step (A) (with no apparent surface strain), or a test specimen in which surface strain has naturally become apparent due to aged deterioration, or step (B- This is a test specimen in which surface strain has become apparent on the surface of the test specimen by performing the manifestation treatment in 1).
The means for acquiring the surface strain image is similar to that used in acquiring the initial image. Furthermore, the surface strain image includes a range where surface strain has become apparent, and is acquired from the same range as the location where the initial image was acquired.

[(D) Test specimen strain distribution creation process]
This step uses a digital image correlation method to create data on the strain distribution on the surface of the concrete specimen from the initial image obtained in step (B) and the surface strain image obtained in step (C). It is a process.
The digital image correlation method calculates the amount of movement of each position on the specimen based on the distribution of brightness values of digital images (initial image and surface strain image) acquired before and after surface strain becomes apparent, and calculates the strain value. (maximum principal strain).

（式（１）中、Ｍはサブセットのｘ，ｙ方向の画素数、ｆ（ｘ，ｙ）は顕在化処理前の初期画像の座標（ｘ，ｙ）におけるサブセット内の輝度値、ｇ（ｘ^＊，ｙ^＊）は顕在化処理後の表面ひずみ画像の座標（ｘ^＊，ｙ^＊）におけるサブセット内の輝度値を示す。） Specifically, the strain value is calculated through the following calculation processes (i) to (ii).
(i) In the initial image, find the luminance value distribution within a subset centered at an arbitrary position.
(ii) Search for a subset of the initial image that has a brightness value distribution that has the highest correlation with the brightness value distribution of the surface strain image, take the center point as the position after the point of interest has been displaced, and The amount of displacement toward the center point is calculated, and the amount of displacement is further converted into a strain value. Note that the correlation between the subsets before and after the manifestation process is expressed using the correlation coefficient R of the following formula (1).

(In formula (1), M is the number of pixels in the x and y directions of the subset, f(x, y) is the brightness value within the subset at the coordinates (x, y) of the initial image before the visualization process, and g(x ^* , y ^* ) indicates the brightness value within the subset at the coordinates (x ^* , y ^* ) of the surface strain image after the visualization process.)

（上記式（２）中、ｕおよびｖはサブセット画像の中心における変位部分であり、Δｘ，Δｙはサブセットの中心から点（ｘ，ｙ）までの距離である。）
以上の計算は、市販の画像解析用ソフトウェア（例えば、ｄｉｇｉｔａｌ：Ｃｏｒｒｅｌａｔｅｄ ｓｏｌｕｔｉｏｎｓ社製）を用いて行うことができる。
デジタル画像相関法によって得られたひずみ数値を、各座標にプロットすることで、各座標と該座標のひずみ数値を含むデータを、ひずみ分布のデータとして作成することができる。 However, in reality, the subset may not be rectangular because the surface strain image itself after the visualization process is deformed with respect to the subset before the visualization process, which is set to be a rectangle. In this case, in order to correct this, assuming that the displacement gradient inside the subset is constant, the following formula (2) is used for the coordinates (x, y) and (x ^* , y ^* ) before and after the manifestation process. use

(In the above formula (2), u and v are the displacement portions at the center of the subset image, and Δx and Δy are the distances from the center of the subset to the point (x, y).)
The above calculation can be performed using commercially available image analysis software (eg, digital: manufactured by Correlated Solutions).
By plotting the strain values obtained by the digital image correlation method at each coordinate, data including each coordinate and the strain value of the coordinate can be created as strain distribution data.

Furthermore, in step (D), by performing the following steps (d-1) to (d-5), a plurality of pieces of strain distribution data consisting of image data can be obtained.
Note that the strain distribution data includes data including strain values and image data created based on the strain values.
[(d-1) Numerical strain data creation process]
This step uses a digital image correlation method to calculate the strain values at each coordinate on the surface of the concrete specimen from the initial image obtained in step (B) and the surface strain image obtained in step (C). This is the process of obtaining a plurality of numerical strain data.
Here, "numerical strain data" is data consisting of the coordinates of the surface of the concrete test body and the numerical value of strain at the coordinates.
Further, the number of coordinates for calculating the strain value is usually a value of the resolution of the initial image and the surface strain image, and the strain value is calculated for each pixel of the image.

[(d-2) Numerical range setting process]
In this step, after sorting the plurality of strain numerical data obtained in step (d-1) in ascending order based on the strain numerical values, the strain numerical value is the smallest among the rearranged plurality of strain numerical data. Randomly select one strain numerical data from strain numerical data located in 10 to 30% (preferably 15 to 25%, more preferably 25%) of the total number of multiple strain numerical data, counting from the data. , the lower limit of the strain value is the value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the strain value of the selected numerical strain data, and 70 to 90 of the total number of multiple numerical strain data. % (preferably 75 to 85%, more preferably 75%), arbitrarily select one strain numerical data from the strain numerical data located at This is a step in which the value obtained by adding the value multiplied by .5 is set as the upper limit value of the strain value.
Here, "strain numerical data located in 10 to 30% (or located in 70 to 90%) of the total number of multiple numerical strain data, counting from the data with the minimum strain numerical value" means in ascending order Among the plurality of numerical strain data sorted, the total number is increased by 0. Strain numerical data located at the number from the number multiplied by 1 (or 0.7) (round down or round up) to the number multiplied by 0.3 (or 0.9) to the above total number (round down or round up) means.
For example, if there are 100 pieces of strain numerical data, the strain numerical data located at 10 to 30% of the total number of multiple strain numerical data are the 10th to 30th strain numerical data counting from the strain numerical data with the minimum strain value. It is located data.
Further, the "interquartile range" is the value obtained by subtracting the first quartile from the third quartile. Note that when the data is arranged in order of magnitude, the value located at 25% from the bottom (here, the strain value) is the first quartile, the value located at 50% is the second quartile, The value located at 75% is called the third quartile).

[(d-3) Numerical strain data selection process]
In this step, from a plurality of strain numerical data, the strain having a numerical value of strain that is greater than or equal to the lower limit value (defined in step (d-2)) and less than or equal to the upper limit value (defined in step (d-2)) is determined. This is the process of selecting numerical data.
In this step, a prediction model with higher accuracy is created by excluding data having a strain value less than the above-mentioned lower limit value and data having a strain value exceeding the above-mentioned upper limit value from the plurality of strain value data. I can do it.
Here, the words "more than the lower limit value" and "below the upper limit value" are used for convenience to divide into two categories based on a specific value, so in the present invention, they are respectively " The terms ``exceeding the lower limit'' and ``less than the upper limit'' may be substituted.

[(d-4) Image data normalization step]
This step is a step of normalizing the distortion numerical data selected in step (d-3) to obtain one sheet of image data having gradations.
For example, the strain value included in the selected strain value data is converted to a brightness level of 256 levels (0 to 255) based on the size of the value, and colors corresponding to the brightness are plotted at the coordinates where the strain value is obtained. In this way, one image data of the surface of the concrete specimen can be obtained.

[(d-5) Strain distribution data creation process]
This step is a step of cutting out a plurality of image data from the single image data obtained in step (d-4) to obtain a plurality of pieces of strain distribution data made up of the image data.
An example of a method for cutting out multiple pieces of image data from one piece of image data is to cut out a specific range (for example, a rectangular range of 10 to 100 pixels on a side) of one piece of image data, and then One example is a method of repeatedly cutting out a range of positions that are preferably moved by 1 to 50 pixels in at least one direction, up, down, left, or right from the position where the image was cut out.
The number of image data cut out from one image data is preferably 10 or more, more preferably 100 or more, particularly preferably 500 or more from the viewpoint of creating a predictive model with higher accuracy. Further, from the viewpoint of image resolution and processing capacity, the above number is preferably 5,000 or less, more preferably 3,000 or less, particularly preferably 2,000 or less.
By cutting out multiple pieces of image data from one piece of image data and using each piece of image data as strain distribution data, multiple pieces of learning data can be obtained from one concrete specimen, resulting in higher accuracy. Can create predictive models.

[(E) Machine learning process]
This process involves machine learning using learning data that is a combination of learning input data including data on the strain distribution on the surface of the concrete specimen obtained in step (D) and learning output data including data on concrete deterioration. This is the process of creating a predictive model for predicting concrete deterioration.
In addition to the strain distribution data described above, the learning input data may include data obtained from the test specimen regarding basic information on concrete, data regarding the service environment of concrete, data regarding concrete structures, etc. .
Data regarding basic information on concrete includes the period of use of concrete (when concrete was installed at the site), data on cement (type, quality, etc.), data on aggregate (type, quality, etc.), and information on materials other than cement and aggregate. Data on materials used, mix of materials used in concrete (unit amount, content), data on kneading of materials used, data on design drawings (dimensions, shape, etc. of concrete specimens), data on parts used (materials) Examples include location of use, shape, material, dimensions, etc.), nominal strength, etc.

Examples of data related to the environment in which concrete is used include solar radiation conditions at the location where the concrete is installed, presence or absence of rain exposure, data related to annual temperature, data related to annual humidity, annual amount of antifreeze sprayed, etc.
Data on concrete structures includes data on residual strain generated inside concrete due to earthquakes, etc., data on past repairs (type of repair method, data on materials used during repair, area to be repaired, etc.). etc.), crack area, etc.
These data may be used alone or in combination of two or more.

The data related to concrete deterioration used as output data for learning include data related to types of concrete deterioration, data related to progress of concrete deterioration, data related to repair methods for concrete deterioration, and the like.
Data regarding the types of concrete deterioration include types of concrete deterioration (e.g., alkaline aggregate reaction (ASR), delayed formation of ettringite (DEF), frost damage, fire damage, fatigue, drying shrinkage, repeated temperature changes, chemical Concrete deterioration (deterioration caused by structural erosion, earthquakes, salt damage, or corrosion of reinforcing steel due to carbonation), and the extent of such deterioration of concrete (size (range) of deterioration, amount of strain, crack area) etc.) etc. If multiple types of deterioration have occurred in the concrete test specimen, the data regarding the type of concrete deterioration should include the type of deterioration, the proportion of each deterioration in the total deterioration occurring in the test specimen, and the data on the type of deterioration of the concrete. The degree is mentioned.
Data regarding the progress of deterioration of concrete include the degree of progress of deterioration of concrete (to what stage has the deterioration progressed), etc.
Data on repair methods for concrete deterioration include crack repair methods, section repair methods, surface coating methods, surface impregnation methods, spalling prevention methods, electrochemical repair methods, and structural renewal methods that are suitable for concrete deterioration. Repair methods include.
These data may be used alone or in combination of two or more.

Machine learning is performed according to a conventionally known general machine learning method using learning data that is a combination of the above-described learning input data and learning output data.
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, and the like. 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 make predictions with higher accuracy. The neural network is preferably a hierarchical neural network having one or more intermediate layers between an input layer and an output layer from the viewpoint of being able to make predictions with higher accuracy. Among these, a convolutional neural network (a neural network having a convolution layer, a pooling layer, etc. as an intermediate layer), which has excellent performance in the field of image recognition, is more suitable.
In addition, cross-validation (e.g., stratified k-fold cross-validation) is performed using a portion (e.g., 20 to 30%) of multiple training data as test data to confirm the reliability of the obtained prediction model. You may go.
Machine learning is performed while appropriately changing the type of data used as learning data, the number of pieces of learning data, the number of times of learning, etc., until a highly reliable predictive model is obtained.
Note that in this specification, "machine learning" refers to learning solely by machines (especially computers) without human thought.

By performing the following steps (F) to (I) using the prediction model created by the above-described method, it is possible to diagnose or predict the deterioration of the concrete to be diagnosed.
Each step will be explained in detail below.
[(F) Initial image acquisition process of target concrete]
This step is a step of acquiring an initial image of the surface of the concrete to be diagnosed.
[(F-1) Target concrete revealing treatment process]
This step is a step that can be added arbitrarily, and is a step performed between step (F) and step (G).
This process is a process in which the concrete for which the initial image of the surface was obtained in step (F) is subjected to processing to make surface strain apparent on the surface of the concrete. .
[(G) Surface strain image acquisition process of target concrete]
This step is a step of acquiring a surface strain image of the surface of concrete where surface strain has become apparent.
[(H) Target concrete strain distribution creation process]
This step is a step of creating data on the strain distribution on the concrete surface from the initial image obtained in step (F) and the surface strain image obtained in step (G) using a digital image correlation method. be.

Furthermore, in step (H), by performing the following steps (h-1) to (h-5), a plurality of pieces of strain distribution data consisting of image data can be obtained.
(h-1) Using the digital image correlation method, calculate the strain value at each coordinate on the surface of the test piece from the initial image and the surface strain image to obtain a plurality of strain value data. Creation step (h-2) After sorting the plurality of strain numerical data obtained in step (h-1) in ascending order based on the strain numerical values, among the plurality of rearranged strain numerical data, strain Randomly select one strain numerical data from the strain numerical data located in 10 to 30% of the total number of the plurality of strain numerical data, counting from the data with the minimum value, and calculate the strain of the selected strain numerical data. The value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the numerical value is the lower limit of the strain value, and from the strain numerical data located in 70 to 90% of the total number of the plurality of strain numerical data A numerical range in which one piece of numerical strain data is arbitrarily selected, and the upper limit of the strain value is the value obtained by adding the value obtained by multiplying the interquartile range by 1.5 to the strain numerical value of the selected numerical strain data. Setting step (h-3) Strain numerical data selection step (h-4) of selecting strain numerical data having a strain numerical value greater than or equal to the lower limit value and less than or equal to the upper limit value from the plurality of numerical strain data data. Image data normalization step (h-5) in which the strain numerical data selected in step 3) is normalized to obtain one image data having gradation. Multiple image data are cut out from the above one image data. Steps (F) to (H), (F-1), (h-1) to (h-5) are concrete test steps. The steps are the same as those in steps (B) to (D), (B-1), and (d-1) to (d-5) described above, except that concrete, which is the target of diagnosis, is used instead of the body. .

[(I) Diagnosis or prediction process]
This step inputs diagnostic input data including strain distribution data into the prediction model created in step (E), outputs diagnostic output data including data regarding concrete deterioration from the prediction model, This is the process of diagnosing or predicting the deterioration of concrete.
The input data for diagnosis is the same as that used as the input data for learning described above.
Further, the diagnostic output data is the same as that mentioned above as the learning output data. These may be used alone or in combination of two or more.
The number of input data for diagnosis may be one or more. For example, in the step (h-5) described above, when a plurality of strain distribution data consisting of image data is obtained, for each of the plurality of strain distribution data, the diagnostic input data including the strain distribution data is converted into a predictive model. A plurality of pieces of diagnostic output data may be obtained by inputting the prediction model into the prediction model and outputting the diagnostic output data from the prediction model. Furthermore, when a plurality of pieces of diagnostic output data are obtained, for example, the most frequently used type of diagnostic output data may be used to diagnose or predict concrete deterioration.

According to the concrete deterioration diagnosis method (steps (F) to (I)) described above, diagnostic input data including strain distribution data regarding the concrete to be diagnosed (for example, an existing concrete structure) is Deterioration of target concrete can be diagnosed or predicted from diagnostic output data including data regarding concrete deterioration, which is obtained by inputting it into a prediction model.
For example, data related to the type of concrete deterioration is output as data related to concrete deterioration, and based on this data, the type of deterioration occurring in the concrete to be diagnosed (for example, alkaline aggregate reaction (ASR)) , delayed formation of ettringite (DEF), deterioration caused by freezing damage, fire damage, salt damage, or corrosion of reinforcing steel due to carbonation), and the extent of concrete deterioration (size of deterioration). can.
If multiple types of deterioration occur in the concrete to be diagnosed, diagnose the main causes (factors) of concrete deterioration based on the type of deterioration, the degree of each deterioration, and the degree of each deterioration. be able to. The type of each deterioration and the degree of each deterioration can be determined from the absolute value of the amount of strain caused by each deterioration.

Additionally, as data regarding the deterioration of concrete, data regarding the progress of deterioration of the concrete subject to diagnosis is output, and based on this data, the progress of deterioration occurring in the concrete subject to diagnosis is diagnosed or predicted. be able to. The progress of deterioration includes not only the degree of progress of deterioration that is currently occurring, but also the prediction of the progress of deterioration of concrete in the future.
When predicting the progress of deterioration of concrete in the future, it is preferable to use data regarding the environment in which concrete is used as input data for diagnosis. Data regarding the service environment of concrete (data such as solar radiation conditions and rain exposure at the concrete installation location, data regarding annual temperature, data regarding annual humidity, annual amount of anti-freezing agent sprayed, etc.) Not only past data from the time of construction to the present is used, but also data (expected data) related to the concrete service environment from the present to the point at which the prediction is desired. Moreover, from the viewpoint of diagnosing or predicting with higher accuracy, it is preferable to use data regarding the above-mentioned basic information of concrete as input data for diagnosis.

In addition, as data on concrete deterioration, data on repair methods for concrete deterioration that is the target of diagnosis is output, and based on this data, the optimal repair method for the target concrete deterioration is diagnosed (the optimal repair method is selected). )can do.
Examples of repair methods include crack repair methods, cross-section repair methods, surface coating methods, surface impregnation methods, spalling prevention methods, electrochemical repair methods, and structural renewal.

In addition to the above-mentioned method for creating a predictive model for predicting concrete deterioration (steps (A) to (E) above), the method for creating a predictive model for concrete deterioration according to the present invention includes the following steps: One method is to create a prediction model for predicting concrete deterioration by performing (J) to (L).
(J) A test body preparation process in which a concrete test body is prepared as a learning sample (K) An image acquisition process of surface cracks on the test body in which a surface crack image of the surface of the above-mentioned test body in which surface cracks have become apparent is acquired. (L) A predictive model for predicting concrete deterioration through machine learning using learning data that is a combination of learning input data including the surface crack image data and learning output data including data regarding concrete deterioration. The machine learning process for creating the following is a detailed explanation of each process.

[(J) Test body production process]
This step is a step of producing a concrete test body as a learning sample.
The concrete test specimen is used to obtain learning data used in a machine learning process (process (L)) described below. The above-mentioned test specimen may be one in which surface cracks have already become apparent (occurred) due to the progression of deterioration of the concrete, or it may be one in which the deterioration of the concrete has not progressed.
The above-mentioned test body may be produced after appropriately adjusting the material composition, manufacturing method, curing conditions, service conditions, etc. of the above-mentioned test body so that the desired deterioration easily occurs.
In addition, test specimens with different material compositions, manufacturing methods, curing conditions, service conditions, etc., test specimens with different degrees of deterioration progress, and after deterioration has occurred, various repair methods are used to repair the deterioration that has occurred. It is preferable to prepare test specimens under various conditions, such as a specimen under various conditions.
Furthermore, the number of types of deterioration occurring in one concrete test specimen may be one type or two or more types.
The concrete test specimen may be one in which the deterioration factor (deterioration cause) is unknown, but from the viewpoint of being able to predict the deterioration of various types of concrete with higher accuracy, it is necessary to use one in which the deterioration factor is clear, or Preferably, the deterioration factor has been identified through various analyses.
Alternatively, a core may be taken from a concrete structure (actual structure) in service and used as the concrete test specimen. Moreover, it is preferable to identify the deterioration factor by analyzing the core before using it as a concrete test specimen.

The above test specimens were tested using alkaline aggregate reaction (ASR), delayed formation of ettringite (DEF), etc., in order to create a more accurate predictive model by obtaining learning data (described later) regarding the deterioration of various types of concrete. It is preferable to prepare a plurality of samples for each type of concrete deterioration, such as frost damage, fire damage, salt damage, and reinforcing steel corrosion due to carbonation.
The shape of the test specimen is not particularly limited, and may be one that imitates the shape of a concrete structure that is actually manufactured.
The above-mentioned test specimen may be one manufactured to obtain learning data, but a part (core portion) or the whole of an actually manufactured concrete structure may be used as the test specimen.
The concrete is not particularly limited and is the same as the concrete in step (A) described above.

[(K) Surface crack image acquisition process of test specimen]
This step is a step of acquiring a surface crack image of the surface of a concrete test specimen in which surface cracks have become apparent.
In this specification, "surface cracks have become apparent" means that surface cracks have occurred (are occurring), preferably, surface cracks have clearly occurred (to the extent that they are visible). (happening).
A concrete test specimen with surface cracks is a concrete test specimen prepared in step (J) that has already developed surface cracks (for example, a concrete test specimen that is in service and is taken from a concrete structure in which surface cracks have become apparent). Concrete specimen prepared in process (J) (with no surface cracks), concrete specimen with natural surface cracks due to aged deterioration, or process (J) ) Tests can be performed on concrete test specimens (with no apparent surface cracks) prepared in This is a concrete test specimen in which surface cracks were intentionally created by applying stress to the specimen or by leaving the specimen in an environment that promotes deterioration of the specimen.
The means for acquiring surface crack images may be any means that can recognize the distribution of cracks generated on the surface, such as a digital camera.
Moreover, the surface crack image is acquired based on the range in which surface cracks are distributed (occurred).

[(L) Machine learning process]
In this process, concrete is improved by machine learning using learning data that is a combination of learning input data including surface crack image data obtained in step (K) and learning output data including data regarding concrete deterioration. This is the process of creating a predictive model for predicting deterioration.
The surface crack image data may be data of the entire image obtained in step (K), or data of an image obtained by cutting out (selecting) a specific range from the image obtained in step (K). It may be. When cutting out an image, a plurality of images may be cut out to generate a plurality of image data. An example of a method for cutting out a plurality of images is a method similar to the method for cutting out a plurality of image data from one sheet of image data described in step (d-5).
In addition to the surface crack image data described above, the learning input data may include data related to basic information on concrete, data related to the service environment of concrete, data related to concrete structures, etc. obtained from the test specimen. good.
The data regarding the basic information of concrete, the data regarding the service environment of concrete, and the data regarding concrete structures include the same data as the data regarding basic information of concrete in the above-mentioned step (E).
These data may be used alone or in combination of two or more.

The data regarding the deterioration of concrete used as the output data for learning include data similar to the data regarding the deterioration of concrete used as the output data for learning in the above-mentioned step (E).
Machine learning is performed according to a conventionally known general machine learning method using learning data that is a combination of the above-described learning input data and learning output data.
Examples of the learning method used for machine learning include the same data regarding concrete deterioration used as the learning output data in step (E) described above.

By using the prediction model obtained by steps (J) to (L) and performing the following steps (M) to (N), it is possible to diagnose or predict the deterioration of the concrete to be diagnosed.
(M) An image acquisition process for surface cracks of the target concrete, in which a surface crack image is acquired of the surface of the concrete to be diagnosed or predicted, in which surface cracks have become apparent. (N) A diagnosis or prediction process for inputting diagnostic input data including data on the surface crack image into the prediction model, and outputting diagnostic output data including data on concrete deterioration from the prediction model, thereby diagnosing or predicting concrete deterioration. Step (M) is the same as step (K) described above, except that the concrete to be diagnosed is used instead of a concrete test specimen.

[(N) Diagnosis or prediction process]
In this step, diagnostic input data including surface crack image data obtained in step (M) is input into the prediction model created in step (L), and data regarding concrete deterioration is included from the prediction model. This is a process of diagnosing or predicting concrete deterioration by outputting diagnostic output data.
The input data for diagnosis is the same as that used as the input data for learning described above.
The surface crack image data may be data of the entire image obtained in step (M), or data of an image obtained by cutting out (selecting) a specific range from the image obtained in step (M). It may be. When cutting out an image, a plurality of images may be cut out to generate a plurality of image data. An example of a method for cutting out a plurality of images is a method similar to the method for cutting out a plurality of image data from one sheet of image data described in step (d-5).

There may be one piece of diagnostic input data, or there may be more than one piece of diagnostic input data. For example, in the case where a plurality of surface crack images are obtained from the surface crack images acquired in step (M), the diagnostic input data including the surface crack image data is used as a prediction model for each of the plural surface crack image data. A plurality of pieces of diagnostic output data may be obtained by inputting the input data into the prediction model and outputting the diagnostic output data from the prediction model. Furthermore, when a plurality of pieces of diagnostic output data are obtained, the most frequently used type of diagnostic output data may be used to diagnose or predict concrete deterioration.
Further, the diagnostic output data is the same as that mentioned above as the learning output data. These may be used alone or in combination of two or more.

According to the concrete deterioration diagnosis method (steps (M) to (N)) described above, diagnostic input data including surface crack image data regarding the concrete to be diagnosed (for example, an existing concrete structure) Deterioration of the target concrete can be diagnosed or predicted from diagnostic output data including data regarding concrete deterioration obtained by inputting the above into a prediction model.
For example, data related to the type of concrete deterioration is output as data related to concrete deterioration, and based on this data, the type of deterioration occurring in the concrete to be diagnosed (for example, alkaline aggregate reaction (ASR)) , Delayed Formation of Ettringite (DEF), Deterioration caused by freezing damage, fire damage, salt damage, or corrosion of reinforcing steel due to carbonation), and the extent of concrete deterioration (size of deterioration) can be diagnosed or predicted. can.
If multiple types of deterioration have occurred in the concrete to be diagnosed, the main causes (factors) of concrete deterioration can be diagnosed or Can be predicted. The type of each deterioration and the degree of each deterioration can be determined from the absolute value of the amount of cracks caused by each deterioration.

Additionally, as data regarding the deterioration of concrete, data regarding the progress of deterioration of the concrete subject to diagnosis is output, and based on this data, the progress of deterioration occurring in the concrete subject to diagnosis is diagnosed or predicted. be able to. The progress of deterioration includes not only the degree of progress of deterioration that is currently occurring, but also the prediction of the progress of deterioration of concrete in the future.
When predicting the progress of deterioration of concrete in the future, it is preferable to use data regarding the environment in which concrete is used as input data for diagnosis. Data regarding the service environment of concrete (data such as solar radiation conditions and rain exposure at the concrete installation location, data regarding annual temperature, data regarding annual humidity, annual amount of anti-freezing agent sprayed, etc.) Not only past data from the time of construction to the present is used, but also data (expected data) related to the concrete service environment from the present to the point at which the prediction is desired. Moreover, from the viewpoint of diagnosing or predicting with higher accuracy, it is preferable to use data regarding the above-mentioned basic information of concrete as input data for diagnosis.

In addition, as data on concrete deterioration, data on repair methods for concrete deterioration that is the target of diagnosis is output, and based on this data, the optimal repair method for the target concrete deterioration is diagnosed (the optimal repair method is selected). )can do.
Examples of repair methods include crack repair methods, cross-section repair methods, surface coating methods, surface impregnation methods, spalling prevention methods, electrochemical repair methods, and structural renewal.

EXAMPLES Hereinafter, the present invention will be specifically explained with reference to Examples, but the present invention is not limited to these Examples.
[Example 1]
As test specimen A, for the purpose of causing deterioration due to alkaline aggregate reaction (ASR), when kneading raw materials, the amount of Na ₂ Oeq (total alkali) in cement was 1.2% by mass. A flat plate measuring 400 mm wide x 400 mm long x 50 mm thick was prepared by hardening concrete to which 1 mol/liter aqueous sodium hydroxide (NaOH) solution was added.
After acquiring an initial image of the upper surface of the test piece A using a line sensor scanner, the test piece A was left standing in an environment of 40±2° C. and a relative humidity of 95% or more for 5 weeks. This is a treatment for promoting the alkaline aggregate reaction (ASR) and exposing surface distortion caused by deterioration due to the alkaline aggregate reaction.
Using a line sensor scanner, a surface strain image of test piece A in which surface strain became apparent was acquired.
Next, using the digital image correlation method, data on the strain distribution of the specimen A was created from the initial image and the surface strain image. The obtained strain distribution data is shown in Figure 1.
Among the strain distribution data (image) created from test specimen A, the strain distribution data with an area that is 10% of the entire image area is used as input data for learning, and data regarding concrete deterioration (deterioration Type: Deterioration due to alkaline aggregate reaction (ASR)) was used as the output data for learning.

In addition, as test specimen B, a flat plate of 400 mm width x 400 mm length x 50 mm thickness was prepared in the same manner as test specimen A except that no sodium hydroxide (NaOH) aqueous solution was added, and drying shrinkage was caused in test specimen B. Ta.
After acquiring an initial image of the upper surface of the test piece B using a line sensor scanner, the surface of the test piece B was heated using a heater to perform surface strain revealing treatment. After heating at 20°C for 4 hours, the temperature was raised to 90°C at a rate of 20°C per hour, and then 90°C was maintained for 12 hours. Thereafter, the temperature was lowered to room temperature at a rate of 8.75°C in 1 hour.
Using a line sensor scanner, a surface strain image of specimen B in which surface strain became apparent was acquired.
Next, using the digital image correlation method, data on the strain distribution of the specimen B was created from the initial image and the surface strain image. The obtained strain distribution data is shown in FIG.
Among the strain distribution data (image) created from test specimen B, the strain distribution data with an area that is 10% of the entire image area is used as input data for learning, and data regarding concrete deterioration (deterioration Type: Deterioration due to drying shrinkage) was used as the output data for learning.
A predictive model was created by performing machine learning using a convolutional neural network and learning data (a combination of learning input data and learning output data) obtained from test specimens A to B.

As the concrete to be predicted, a flat plate measuring 400 mm wide x 400 mm long x 50 mm thick was prepared to simulate deterioration due to alkali aggregate reaction (ASR).
After acquiring an initial image of the upper surface of the concrete using a line sensor scanner, the surface of the concrete was heated using a heater to perform surface strain revealing treatment. Specifically, after pre-curing at 20°C for 4 hours, the temperature was raised at a temperature increase rate of 20°C per hour until it reached 90°C, then 90°C was maintained for 12 hours, and then for 1 hour. The revealing treatment was carried out by lowering the temperature to room temperature at a temperature lowering rate of 8.75°C.
A line sensor scanner was used to obtain surface strain images of concrete with visible surface strain.
Next, the strain distribution of the concrete was created from the initial image and the surface strain image using a digital image correlation method. The obtained strain distribution is shown in FIG. 3.
Among the strain distribution data (image) created from concrete, strain distribution data of an area that is 10% of the area of the entire image is used as diagnostic input data, and this data is input into a prediction model, Diagnosis output data including data on concrete deterioration was output from the prediction model to diagnose concrete deterioration, and it was found that combined deterioration due to drying shrinkage and alkaline aggregate reaction (ASR) had occurred, and alkaline aggregate reaction It was diagnosed that the main cause was

[Example 2]
The above test specimen A was left standing for 6 weeks in an environment of 40±2° C. and a relative humidity of 95% or more. This is a treatment for promoting the alkaline aggregate reaction (ASR) and exposing surface cracks caused by deterioration due to the alkaline aggregate reaction.
A location where surface cracks were widely distributed was selected by visual inspection, and a surface crack image including the location was acquired.
Among the surface crack images obtained from test specimen A, image data (surface crack image data) in an area that is 10% of the entire area of the image is used as input data for learning, and data regarding concrete deterioration is used as learning input data. (Type of deterioration: deterioration due to alkaline aggregate reaction (ASR)) was used as output data for learning.
Machine learning was performed using a convolutional neural network and learning data obtained from test body A (a combination of learning input data and learning output data) to create a predictive model.
As the concrete to be predicted, a flat plate measuring 400 mm wide x 400 mm long x 50 mm thick was prepared to simulate deterioration due to alkali aggregate reaction (ASR).
A location where surface cracks were widely distributed was selected by visual inspection, and a surface crack image including the location was acquired. The obtained surface crack image is shown in FIG.
Among surface crack images created from concrete, image data (surface crack image data) in an area that is 10% of the entire area of the image is used as input data for diagnosis, and this data is used as a prediction model. When the prediction model outputs diagnostic output data including data on concrete deterioration and diagnosed concrete deterioration, it was diagnosed that the main cause was alkaline aggregate reaction (ASR).

[Example 3]
Using the concrete used in test body A of Example 1, two reinforcing bars (D10) were arranged in parallel inside the test body with an interval, 400 mm wide x 400 mm long x 50 mm thick. A test specimen C-1 was prepared.
In addition, a test specimen C-2 of 400 mm width x 400 mm length x 50 mm thickness, in which two reinforcing bars (D16) are arranged parallel to each other at intervals inside the test specimen using the above concrete; Four reinforcing bars (D10) were arranged inside the test specimen in a square shape, that is, two bars at a time in parallel and vertically and horizontally, measuring 400 mm wide x 400 mm long x 50 mm thick. A test body C-3 and a test body C-4 each measuring 400 mm in width x 400 mm in length x 50 mm in thickness without using reinforcing bars were prepared.
After acquiring an initial image of the top surface of each test piece using a line sensor scanner, each test piece was left standing in an environment of 40±2° C. and a relative humidity of 95% or higher.
Using a line sensor scanner, surface strain images of the test specimens in which surface strain became apparent were obtained after 3 weeks, 5 weeks, 7 weeks, and 9 weeks.
Next, using the digital image correlation method, the strain value at each coordinate was calculated for each pixel from the initial image and the surface strain image.
After sorting the obtained strain numerical data (187 x 342 matrix data) in ascending order of strain numerical value, from the strain numerical value located at 25% from the minimum strain numerical value, 1 in the interquartile range. A value was calculated by subtracting the value multiplied by .5, and the calculated value was taken as the lower limit of the strain value.
Further, from the strain value located at 75% from the minimum value of the strain value, a value was calculated by adding a value obtained by multiplying the interquartile range by 1.5, and the calculated value was taken as the upper limit of the strain value.
Next, from among the obtained numerical strain data, data in which the strain numerical values contained in the data are less than the above lower limit value and exceed the above upper limit value are excluded, the remaining numerical strain data is normalized, and the numerical strain data is was converted into image data expressed in 256 levels (0 to 255) of brightness.
An image of 64×64 pixels was cut out from the above image data. By cutting out a total of 1,000 images by moving 7 pixels vertically and 5 pixels horizontally based on the position of the image, data on 1,000 strain distributions can be obtained from one surface strain image. Ta.
That is, by obtaining 4,000 pieces of strain distribution data from one test piece, a total of 16,000 pieces of strain distribution data were obtained.

Using the concrete used in test body B of Example 1, two reinforcing bars (D10) were arranged in parallel inside the test body with an interval, 400 mm wide x 400 mm long x 50 mm thick. A test specimen D-1 was prepared.
In addition, a test specimen D-2 of 400 mm width x 400 mm length x 50 mm thickness, in which two reinforcing bars (D16) are arranged parallel to each other at intervals inside the test specimen using the above concrete; Four reinforcing bars (D10) were arranged inside the test specimen in a square shape, that is, two bars at a time in parallel and vertically and horizontally, measuring 400 mm wide x 400 mm long x 50 mm thick. A test body D-3 and a test body D-4 each measuring 400 mm in width x 400 mm in length x 50 mm in thickness without using reinforcing bars were prepared.
After obtaining an initial image of the top surface of each specimen using a line sensor scanner, each specimen was subjected to drying shrinkage in the same manner as specimen B in Example 1 to reveal surface strain. processed.
Using a line sensor scanner, surface strain images of the test specimens in which surface strain became apparent were obtained after 1 week, 2 weeks, 4 weeks, and 6 weeks, respectively.
In the same manner as the above-mentioned test specimens C-1 to C-4, 1,000 pieces of strain distribution data were obtained from one surface strain image.
That is, by obtaining 4,000 pieces of strain distribution data from one test piece, a total of 16,000 pieces of strain distribution data were obtained.

In addition, when kneading raw materials for the purpose of causing deterioration due to delayed formation of ettringite, Same as test specimens C-1 to C-4 described above, except that potassium sulfate (K ₂ SO ₄ ) was added in an amount of 2% by mass in terms of SO ₃ based on the mass of cement (100% by mass). Test specimens E-1 to E-4 were prepared in the same manner.
After acquiring an initial image of the top surface of each test piece using a line sensor scanner, each test piece was left standing in an environment of 40±2° C. and a relative humidity of 95% or higher.
Using a line sensor scanner, surface strain was observed after 1 week, 3 weeks, 5 weeks, 7 weeks (excluding specimens without reinforcing steel), and 9 weeks. Surface strain images of each specimen were obtained.
In the same manner as the above-mentioned test specimens C-1 to C-4, 1,000 pieces of strain distribution data were obtained from one surface strain image.
In other words, by obtaining data on 5,000 or 4,000 strain distributions (obtained from test specimens that do not use reinforcing bars) from one test specimen, a total of 19,000 strain distributions can be obtained. I got the data.

Using the data of a total of 51,000 strain distributions described above, machine learning was performed to create a prediction model.
In addition, 12,000 strain data obtained from test specimens C-1 to C-4 (deterioration due to alkali aggregate reaction (ASR)), test specimens D-1 to D-4 (deterioration due to drying shrinkage) 12,000 pieces of strain data obtained from test specimens E-1 to E-4 (where deterioration occurred due to delayed formation of ettringite) The strain data of Machine learning was performed by using the 4,000 pieces of strain data obtained from test specimens E-1 to E-4 (5,000 pieces of strain data obtained from test specimens E-1 to E-4) as verification data.

As the concrete to be predicted, a flat plate measuring 400 mm wide x 400 mm long x 50 mm thick was prepared to simulate deterioration due to alkali aggregate reaction (ASR).
After acquiring an initial image of the upper surface of the concrete using a line sensor scanner, it was left standing for 9 weeks in an environment of 40±2° C. and relative humidity of 95% or more.
Using a line sensor scanner, surface strain images of the concrete in which surface strain became apparent after 3, 5, 7, and 9 weeks were obtained, respectively.
Next, in the same manner as the above test body A, 1,000 strain distribution data were obtained from each of the test bodies after 3, 5, 7, and 9 weeks.
For each of the 1,000 pieces of strain distribution data, input the strain distribution data as input data for diagnosis into a prediction model, output diagnostic output data including data regarding concrete deterioration from the prediction model, 1. Data on deterioration (types of deterioration) of 000 concretes were obtained. The accuracy rate of the data was calculated.

In addition, a flat plate measuring 400 mm wide x 400 mm long x 50 mm thick, which simulates drying shrinkage, was used as the concrete to be predicted. After obtaining data on the deterioration of 1,000 pieces of concrete, we calculated the accuracy rate.
Furthermore, except for using a flat plate measuring 400 mm wide x 400 mm long x 50 mm thick, which simulates deterioration due to delayed formation of ettringite, as the concrete to be predicted, a flat plate that simulates deterioration due to alkaline aggregate reaction (ASR) is used. After obtaining data on the deterioration of 1,000 pieces of concrete in the same manner as above, the accuracy rate was calculated.
The results are shown in Table 1.

[Example 4]
Calculate the value by subtracting the value obtained by multiplying the interquartile range by 1.5 from the strain value located 10% from the minimum value of the strain value, and use the calculated value as the lower limit of the strain value, and Same as Example 3 except that a value is calculated by adding a value obtained by multiplying the interquartile range by 1.5 from the strain value located at 90% from the value, and the calculated value is used as the upper limit of the strain value. A predictive model was created using the predictive model, and the prediction model was used to evaluate concrete We outputted data regarding the deterioration of the data and calculated the accuracy rate. The results are shown in Table 1.

From Table 1, the accuracy rate of Example 3 (in which a strain value located at 25% and a strain value located at 75% from the minimum value of strain values were selected in step (d-2)) was It can be seen that this is higher than the accuracy rate of Example 4 (in which the strain values located at 10% and 90% of the minimum strain value were selected in step (d-2)).

Claims (7)

A predictive model for predicting concrete deterioration is created by performing the following steps (A) to (C), (d-1) to (d-5), and (E). How to create a predictive model for concrete deterioration.
(A) A test body production process in which a concrete test body is prepared as a learning sample. (B) An initial image acquisition process of the test body in which an initial image of the surface of the above test body is acquired. (C) Surface strain has become apparent. A surface strain image acquisition step of the test piece, which acquires a surface strain image of the surface of the test piece.
(d-1) Using a digital image correlation method, calculate the strain value at each coordinate on the surface of the test piece from the initial image and the surface strain image to obtain a plurality of strain value data. Creation process
(d-2) After sorting the plurality of strain numerical data obtained in step (d-1) in ascending order based on the strain numerical values,
One strain numerical data from among the strain numerical data located in 15 to 30% of the total number of the plurality of strain numerical data, counting from the data with the minimum strain numerical value among the plurality of rearranged numerical strain data. is arbitrarily selected, and the value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the strain value of the selected numerical strain data is set as the lower limit value of the strain value, and
One strain numerical data is arbitrarily selected from the strain numerical data located in 70 to 85% of the total number of the plurality of strain numerical data, and the strain numerical value of the selected strain numerical data is set to 1.5 in the interquartile range. A numerical range setting process in which the upper limit of the strain value is the value obtained by adding the values multiplied by
(d-3) Strain numerical data selection step of selecting strain numerical data having a strain numerical value greater than or equal to the lower limit value and less than or equal to the upper limit value from the plurality of numerical strain data data.
(d-4) Image data normalization step of normalizing the strain numerical data selected in step (d-3) to obtain one piece of image data having gradation.
(d-5) Strain distribution data creation step of cutting out a plurality of image data from the one sheet of image data to obtain a plurality of strain distribution data on the surface of the test piece made up of image data.
(E) Develop a predictive model for predicting concrete deterioration through machine learning using learning data that is a combination of learning input data including the above strain distribution data and learning output data including data regarding concrete deterioration. Machine learning process to create The method for creating a predictive model for concrete deterioration according to claim 1 , wherein the following step (B-1) is performed between the step (B) and the step (C).
(B-1) A test piece in which a process is performed on the test piece for which an initial image has been acquired to make surface strain apparent on the surface of the test piece, and surface strain is made obvious on the surface of the test piece. manifestation process The following steps (F) to (G), (h- 1 ) to (h-5), and A concrete deterioration diagnosis or prediction method that diagnoses or predicts deterioration of target concrete by performing (I).
(F) Initial image acquisition step of the target concrete, which acquires an initial image of the surface of the concrete that is the target of diagnosis or prediction. (G) Target concrete, which acquires a surface strain image of the surface of the concrete where surface strain has become apparent. Surface strain image acquisition process
(h-1) Using the digital image correlation method, calculate the strain value at each coordinate on the surface of the test piece from the initial image and the surface strain image to obtain a plurality of strain value data. Creation process
(h-2) After sorting the multiple strain numerical data obtained in step (h-1) in ascending order based on the strain numerical values,
One strain numerical data from among the strain numerical data located in 15 to 30% of the total number of the plurality of strain numerical data, counting from the data with the minimum strain numerical value among the plurality of rearranged numerical strain data. is arbitrarily selected, and the value obtained by subtracting the value obtained by multiplying the interquartile range by 1.5 from the strain value of the selected numerical strain data is set as the lower limit value of the strain value, and
One strain numerical data is arbitrarily selected from the strain numerical data located in 70 to 85% of the total number of the plurality of strain numerical data, and the strain numerical value of the selected strain numerical data is set to 1.5 in the interquartile range. A numerical range setting process in which the upper limit of the strain value is the value obtained by adding the values multiplied by
(h-3) Strain numerical data selection step of selecting strain numerical data having a strain numerical value greater than or equal to the lower limit value and less than or equal to the upper limit value from the plurality of numerical strain data.
(h-4) Image data normalization step in which the strain numerical data selected in step (h-3) is normalized to obtain one piece of image data having gradation.
(h-5) Strain distribution data creation step of cutting out multiple pieces of image data from the single image data and obtaining multiple pieces of strain distribution data on the concrete surface made up of image data.
(I) Input diagnostic input data including data on the strain distribution to the prediction model, output diagnostic output data including data on concrete deterioration from the prediction model, and diagnose or predict concrete deterioration. diagnostic or predictive process The method for diagnosing or predicting concrete deterioration according to claim 3 , wherein the following step (F-1) is performed between the step (F) and the step (G).
(F-1) Processing is performed on the concrete for which the initial image has been obtained to make surface strain apparent on the surface of the concrete, and surface strain is made apparent on the surface of the concrete to make the target concrete visible. chemical treatment process The concrete according to claim 4 , wherein the data regarding the deterioration of concrete output in the diagnosis or prediction step includes data regarding the type of deterioration of the target concrete, and the type of deterioration of the concrete is diagnosed or predicted based on the data. A method for diagnosing or predicting the deterioration of. According to claim 4 or 5 , the data regarding the deterioration of concrete outputted in the diagnosis or prediction step includes data regarding the progress of deterioration of the target concrete, and the progress of deterioration of the target concrete is diagnosed or predicted based on the data. Methods for diagnosing or predicting concrete deterioration as described. A claim in which the data regarding the deterioration of concrete outputted in the diagnosis or prediction step includes data regarding a repair method for deterioration of the target concrete, and based on the data, the optimal repair method for deterioration of the target concrete is diagnosed or predicted. The method for diagnosing or predicting concrete deterioration according to 4 or 5 .

Images