JP7760126B2 - Compaction assessment system and compaction assessment method - Google Patents

Description

Article 30, paragraph 2 of the Patent Act applies. ・Deep learning-based automatic compaction determination system taking into account the fresh properties of concrete (March 1, 2021) ・DIA2021 Dynamic Image Processing Practical Use Workshop 2021 (March 4, 2021) ・CS14-54 Test judgment using an AI automatic concrete compaction determination system (August 2, 2021) ・76th Annual Academic Lecture Meeting, 2021 National Convention of the Japan Society of Civil Engineers (September 10, 2021) ・Attempt to develop an AI-based automatic concrete compaction determination system (November 17, 2021) ・2nd AI and Data Science Symposium (November 17, 2021)

This invention relates to technology related to the compaction of fresh concrete, and more specifically to a compaction assessment system that can automatically assess the compaction level while taking into account the properties of fresh concrete, such as slump and air content, and a compaction assessment method using this system.

Along with steel, concrete is one of the most important construction materials, and is used in a variety of structures, including civil engineering structures such as dams, tunnels, and bridges, as well as architectural structures such as apartment buildings and office buildings. These concrete structures are sometimes manufactured in advance in factories and transported to the desired location, but in the case of civil engineering and architectural structures, they are often constructed directly at the desired location (site). In either case, concrete structures are constructed by pouring concrete (fresh concrete), which is a mixture of cement, water, aggregate, etc., into formwork, waiting for the concrete to harden, and then removing the formwork.

Fresh concrete is poured into the formwork by being poured through a chute from an agitator truck, by being dropped from a hose by a concrete pump truck, or in some cases by a worker splashing it with a shovel. The fresh concrete poured into the formwork is then compacted using a vibrator (vibrating vibrator). When the vibrating vibrator vibrates the concrete, it liquefies, and as it liquefies further, air bubbles within the concrete rise and escape to the outside, and the aggregate and mortar within the concrete are rearranged, resulting in the concrete being compacted.

Traditionally, determining whether concrete is sufficiently compacted has been left to the operator's judgment, based on various factors such as subtle changes in the concrete surface and the compaction time. The Standard Specifications for Concrete Construction also state that the degree of compaction should be determined when "a line of cement paste appears at the contact point between the concrete and the sheathing," and "the concrete shows no visible reduction in volume, the surface is nearly level, and a gloss appears on the surface." In other words, determining the compaction of concrete, and ultimately the quality of concrete structures, relies on the experience and knowledge of the vibrator operator. However, there are only a limited number of operators who can accurately determine compaction, and with the chronic labor shortage in recent years, it is becoming increasingly difficult to secure such operators in a timely manner.

As a result, several technologies have been proposed to objectively evaluate the compaction of concrete without relying on the operator's qualitative judgment. For example, Patent Document 1 proposes a technology that continuously acquires images of concrete being compacted and uses machine learning to determine the compaction level from those images.

Japanese Patent Application Laid-Open No. 2022-015003

The technology described in Patent Document 1 is extremely useful in that it allows for objective judgments without relying on the qualitative judgment of the operator. By using the technology described in Patent Document 1, it is not necessary to secure an operator who can properly judge the compaction of concrete, and since the judgment results are recorded, it is also possible to fulfill post-mortem verification and accountability.

Insufficient compaction can result in unfilled areas (junk marks), large air bubbles (pockmarks), or uneven distribution of constituent materials, while excessive compaction can lead to material separation due to differences in the density of the constituent materials. Both insufficient and excessive compaction can impair the quality of the hardened concrete expected when a concrete structure is designed, so it is necessary to determine when compaction is complete in an appropriate and immediate manner (in real time).

It is also known that the time required for compaction (the time until compaction is complete) varies greatly depending on the properties of the fresh concrete being compacted (hereinafter referred to as "fresh properties"). For example, concrete with a high slump (i.e., soft properties), which is one type of fresh property, generally requires a shorter compaction time, while concrete with a low slump (i.e., hard properties) tends to require a longer compaction time. In other words, the fresh properties of concrete are an extremely important factor when judging the compaction of that concrete.

However, conventional technology that uses machine learning to determine compaction from images only studies images to generate a trained model, and does not take into account the fresh properties of the concrete. As a result, while compaction can be determined relatively accurately for concrete with a high slump, it is not possible to determine compaction very accurately for concrete with a low slump, resulting in variations in compaction determination depending on the fresh properties.

The objective of the present invention is to solve the problems inherent in conventional technology by providing a compaction assessment system and method that can assess the compaction of fresh concrete while taking into account the fresh properties of the concrete.

The present invention was developed based on an unprecedented concept, focusing on the generation of a trained model through machine learning of concrete images marked with compaction completion/incompletion status and fresh properties.

The compaction assessment system of the present invention is a system that assesses the compaction of fresh concrete using concrete images captured of the surface of fresh concrete, and is equipped with an image classification means and an assessment means. The image classification means is a means for assigning a "judgment label (completed label/incomplete label)" to a concrete image based on the results output by a trained model by inputting concrete images and fresh property data, and the assessment means is a means for assessing compaction based on the number of completed labels assigned by the image classification means. Note that the "judgment label" referred to here does not refer to a final assessment of compaction, but rather to a provisional evaluation result for each concrete image. The fresh property data is data indicating the properties of fresh concrete, including at least slump and air content. The trained model is generated by machine learning concrete images labeled as compacted or incomplete and fresh property data. The image classification means then assigns a completed label when the trained model determines that the compaction of the fresh concrete related to the input concrete image is complete, and assigns an incomplete label when the trained model determines that the compaction of the fresh concrete is incomplete.

The compaction determination system of the present invention can also have a trained model composed of an image encoder, a fresh property encoder, and a classifier. The image encoder converts the input concrete image into an m-dimensional image feature vector (m is a natural number). The fresh property encoder converts the input fresh property data into an n-dimensional property data feature vector (n is a natural number). The classifier converts the image feature vector and property data feature vector into an m+n-dimensional integrated feature vector by concatenating them, and outputs the likelihood of whether the compaction of the fresh concrete is complete or incomplete based on the integrated feature vector.

The compaction assessment system of the present invention can also determine that compaction is complete when, among the assessment labels assigned to each concrete image that makes up an "image group," there are more completed labels than incomplete labels. Note that an "image group" here refers to a collection of multiple concrete images arranged in chronological order of their capture.

The compaction assessment system of the present invention may further include an input means. This input means allows an operator to input fresh property data.

The compaction assessment system of the present invention can also use fresh property data that further includes one or more of the following data: slump flow, N-type penetration, and electrical conductivity.

The compaction assessment system of the present invention may further include an image segmentation means. This image segmentation means is a means for segmenting a concrete image into a plurality of divided concrete images. In this case, the image classification means assigns a completed/incomplete label to each divided concrete image, and the assessment means assesses the compaction for each region corresponding to the divided concrete image.

The compaction assessment method of the present invention is a method for assessing the compaction of fresh concrete using the compaction assessment system of the present invention, and is a method comprising an image acquisition process, an image classification process, and an assessment process. In the image acquisition process, fresh concrete is photographed at predetermined intervals to acquire multiple concrete images. In the image classification process, the image classification means assigns a completed/incomplete label to each of the multiple concrete images, and in the assessment process, the assessment means assesses the compaction state.

The compaction assessment system and compaction assessment method of the present invention have the following effects.
(1) It is possible to objectively and instantly (in real time) determine whether concrete compaction is complete without relying on the judgment of the operator operating the vibration vibrator.
(2) Since the compaction is judged taking into account the fresh properties of the concrete, the compaction of the concrete can be judged more appropriately than in the prior art.
(3) The present invention can be applied to various cases, such as when a vibration vibrator is operated not only by a person but also when a vibration vibrator is attached to a construction machine or the like.

A model diagram that schematically shows multiple images that make up a video or a series of still images. 1 is a block diagram showing the main configuration of the compaction determination system of the present invention. (a) is a model diagram that schematically shows the network structure of the trained model, and (b) is an explanatory diagram that shows the contents of each layer of the trained model. 1A is an image diagram schematically showing an unsegmented concrete image, and FIG. 1B is an image diagram schematically showing a concrete image made up of a plurality of segmented concrete images. FIG. 10 is a model diagram illustrating a determination process using a voting function. (a) is an image showing the transparency of concrete in the early stages of compaction, and (b) is an image showing the transparency of concrete after compaction has progressed to a certain extent. (a) is an image showing concrete in the early stages of compaction with a white and black frame, and (b) is an image showing concrete that has already been compacted to a certain extent with a white and black frame. A flow chart showing the main processing flow of the compaction determination system of the present invention. 1 is a flow chart showing the main processing flow of the compaction determination method of the present invention. (a) is a graph showing the results of a compaction evaluation of fresh concrete using a conventional model, (b) is a graph showing the results of a compaction evaluation of fresh concrete using a model of the present invention, and (c) is a legend for explaining this figure.

An example implementation of the compaction assessment system and compaction assessment method of the present invention is described below with reference to the accompanying drawings.

1. Overview The present invention is a technology for determining compaction based on images (hereinafter referred to as "concrete images") of the surface of fresh concrete undergoing compaction. This technology can be used in the construction of various concrete structures, including cast-in-place concrete work, precast concrete production in factories, and precast concrete production at construction sites (so-called site PCs). More specifically, the technology sequentially captures images of the concrete as it changes during compaction, and uses the resulting concrete images to determine whether compaction of the fresh concrete is complete or incomplete. Therefore, to avoid missing the point at which compaction is complete, it is preferable to capture concrete images at a relatively short interval. It is therefore preferable to capture the target object as a video or a series of still images, as shown in Figure 1. While both video and continuous still images are composed of images (frames) captured at extremely short intervals, video generally refers to images captured at 24 to 60 frames per second (i.e., 24 to 60 fps). Therefore, continuous still images are defined herein as images captured at a rate of less than 24 fps or more than 60 fps.

2. Compaction Determination System The compaction determination system of the present invention will be described in detail. The compaction determination method of the present invention is a method for determining the compaction of fresh concrete using the compaction determination system of the present invention. Therefore, the compaction determination system of the present invention will be described first, and then the compaction determination method of the present invention will be described.

Figure 2 is a block diagram showing the main components of the compaction assessment system 100 of the present invention. As shown in this figure, the compaction assessment system 100 of the present invention is configured to include an image classification means 101 and an assessment means 102, and can also be configured to include an input means 103, an image segmentation means 104, an image acquisition means 105, an image selection means 106, a model generation means 107, an image analysis means 108, a display control means 109, a display means 110, an image storage means 111, and a trained model storage means 112.

Of the main elements that make up the compaction assessment system 100, the image classification means 101, assessment means 102, input means 103, image segmentation means 104, image selection means 106, model generation means 107, image analysis means 108, and display control means 109 can be manufactured as dedicated components, or a general-purpose computer device can be used. This computer device is equipped with a processor such as a CPU, memory such as ROM and RAM, input means such as a mouse and keyboard, and a display, and can be configured as a personal computer (PC), server, tablet PC such as an iPad (registered trademark), or mobile device including a smartphone. When using a computer device equipped with a display, it is recommended that the display be used as the display means 110.

The image storage means 111 and trained model storage means 112 can be implemented as a general-purpose computer storage device, or as a database server. If implemented as a database server, they can be placed on a local network (LAN: Local Area Network), or can be a cloud server that stores data via the internet. The image acquisition means 105 can acquire video and continuous still images, and can be implemented using a digital video camera, digital camera, smartphone, tablet PC, or the like. The video and continuous still images (concrete images) acquired by the image acquisition means 105 are stored in the image storage means 111 (Figure 2).

Below, we will provide a detailed explanation of each of the main elements that make up the compaction assessment system 100 of the present invention.

(Model generation means)
The model generation means 107 is a means for generating a "trained model" by machine learning numerous concrete images and fresh property data (Figure 2). The fresh property data is data indicating fresh properties (properties of fresh concrete) and is a dataset including slump and air content. It can also be a dataset including slump flow, N-type penetration, and electrical conductivity (one or more of these). Meanwhile, each concrete image used for training is labeled "compacted" or "incomplete" by a person (e.g., an experienced technician). For convenience, concrete images labeled "compacted" or "incomplete" are referred to as "concrete images with a judgment." The trained model generated by the model generation means 107 is stored in the trained model storage means 112 (Figure 2).

By the way, judged concrete images can be generated by using video or a series of still images taken of the compaction work of fresh concrete, and attaching a human judgment to each concrete image (compacted/not compacted). Of course, when performing machine learning, a combination of judged concrete images and fresh property data related to the concrete contained in those images (hereinafter, this combination will be referred to as "training data") is used. Therefore, it is a good idea to compact two or more types of concrete with different fresh properties, obtain video or a series of still images of the compaction, and then use the resulting large amount of training data related to two or more types of concrete for learning.

The machine learning used to generate trained models can employ various conventional machine learning techniques, including deep learning techniques such as CNN (Convolutional Neural Network) and NIN (Network In Network). NIN incorporates a small-scale multilayer perceptron (MLP) instead of a simple convolutional filter, allowing for more complex representations than conventional CNNs. Furthermore, introducing a Global Average Pooling (GAP) layer instead of a fully connected layer reduces the weight parameters, resulting in reduced memory consumption and faster operation.

(Input means)
The input means 103 is a means by which fresh property data can be input by an operator, and can be, for example, a pointing device (mouse, touch panel, pen tablet, touchpad, trackpad, trackball, etc.), a keyboard, etc. Furthermore, concrete images are input to the trained model under the control of the image selection means 106. Specifically, the image selection means 106 selectively controls whether the concrete images are automatically input via wireless or wired communication means at the time of acquisition, or whether the concrete images stored in the image storage means 111 are read and input.

(Image analysis means)
The image analysis means 108 is a means for outputting the likelihood of compaction completion or incompletion for a concrete image. Here, likelihood refers to the probability of compaction completion or incompletion, and is output as, for example, "compaction completion is 75% and incompletion is 25%." To output this likelihood, a trained model generated by the model generation means 107 is used. That is, by inputting fresh property data and a concrete image into the trained model, the likelihood of compaction completion or incompletion for that concrete image is output (Figure 2).

Figure 3 shows an example of a trained model generated by the model generation means 107, where (a) is a model diagram that schematically illustrates the network structure, and (b) is an explanatory diagram illustrating the contents of each layer. As shown in Figure 3(a), the trained model can be configured to include an image encoder, a fresh property encoder, and a classifier. Of these, the image encoder mainly passes (processes) concrete images and is configured with an image convolution layer, a pooling layer, a flattening layer, and a fully connected layer, and converts the input concrete image into an m-dimensional (m is a natural number) feature vector (hereinafter referred to as "image feature vector").

The fresh property encoder mainly passes (processes) fresh property data and is composed of a fully connected layer, converting the input fresh property data into an n-dimensional feature vector (hereinafter referred to as the "property data feature vector") (n is a natural number). The classifier, on the other hand, is composed of a concatenation layer, a fully connected layer, and an output layer, and outputs the likelihood of fresh concrete compaction being complete or incomplete based on an m+n-dimensional feature vector (hereinafter referred to as the "integrated feature vector") that combines the image feature vector and property data features.

For example, in Figure 3, a concrete image (input size: 270 x 270 x 3) is input to the MLP convolution layer. The flattening layer flattens the output (output size: 384), and the fully connected layer of the image encoder converts it into a 512-dimensional image feature vector. Meanwhile, the fully connected layer of the fresh property encoder converts the input fresh property data (input size: 5) into a 256-dimensional property data feature vector. The concatenation layer of the classifier then concatenates the image feature vector and property data feature vector into a 768-dimensional (512 + 256) integrated feature vector. The output layer then outputs the likelihood of compaction completion or incompletion based on the feature vector (128 dimensions) converted by the fully connected layer of the classifier. The input size of the MLP convolution layer 1 shown in Figure 3(b), "270 x 270 x 3," is the number of horizontal pixels x the number of vertical pixels x the number of channels, with the number of channels being the three RGB bands of the image. In this example, the fresh property data consists of five types: slump and air volume, slump flow, N-type penetration, and electrical conductivity, so the input size of the fully connected layer of the fresh property encoder is set to "5".

(Image Classification Means)
The image classification means 101 is a means for assigning a "determination label" to an input concrete image based on the output result of the image analysis means 108 (i.e., the trained model) (FIG. 2). Specifically, if the likelihood of compaction completion is greater than the likelihood of compaction incomplete, the image is assigned a "completed label." Conversely, if the likelihood of compaction incomplete is greater than the likelihood of compaction completion, the image is assigned an "incomplete label." As mentioned above, the "determination label" here does not refer to a final determination of compaction, but rather to a provisional evaluation result for each concrete image.

The compaction assessment system 100 of the present invention can be configured to process a single concrete image as a single region (i.e., all pixels are targeted), or it can be configured to process a single image as consisting of multiple images (hereinafter referred to as "divided concrete images"). Figure 4(a) shows an undivided concrete image (i.e., all pixels), while Figure 4(b) shows a concrete image consisting of multiple (8 x 6 in the figure) divided concrete images.

When processing a single concrete image as a single region, as shown in Figure 4(a), the image analysis means 108 outputs the likelihood of compaction being complete or incomplete for that concrete image, and the image classification means 101 assigns a judgment label (completed label/incomplete label) to that concrete image based on the output result. On the other hand, when processing as consisting of multiple segmented concrete images, as shown in Figure 4(b), the image segmentation means 104 first segments the concrete image into multiple segmented concrete images. Then, by inputting the segmented concrete images and fresh property data into the trained model, the image analysis means 108 outputs the likelihood of compaction being complete or incomplete for that segmented concrete image, and the image classification means 101 assigns a completed label or incomplete label to that segmented concrete image based on the output result. For example, in the case of Figure 4(b), a judgment label (completed label/incomplete label) is assigned to each of the 48 segmented concrete images.

(Judgment means)
The determination means 102 is a means for making a final determination as to whether compaction is complete or incomplete for an area of fresh concrete corresponding to a concrete image (hereinafter referred to as a "concrete area") (Fig. 2). Of course, in the case where the concrete image is not divided as shown in Fig. 4(a), the determination is made for the concrete area corresponding to the entire concrete image, while in the case where the concrete image is divided as shown in Fig. 4(b), the determination is made for each concrete area corresponding to the divided concrete image.

When the image classification means 101 assigns a completion label even once, the determination means 102 can immediately determine that the concrete region has been compacted. However, just as even experienced technicians tend to be cautious in their judgments before and after the concrete has been compacted, a single timing alone may result in an incorrect determination. In particular, when machine learning techniques are used, chattering, in which two classes of results alternate, is known to occur. For this reason, a final determination is generally made after waiting for the results to stabilize. Therefore, the determination means 102 may be configured to determine whether compaction is complete based on the number of completion labels assigned. For example, the determination means 102 may be configured to determine whether compaction is complete for a concrete region when the cumulative number of assigned completion labels reaches a predetermined number. Alternatively, the determination of whether compaction is complete may be made using a voting function. The determination procedure using the voting function is described below with reference to Figure 5.

Figure 5 is a model diagram explaining the judgment process using the Voting function. In this diagram, each concrete image is represented by a line segment (vertical line), and the concrete images are arranged in order of the time of shooting (i.e., in chronological order) starting from the image at which shooting began (the leftmost image in the diagram). Furthermore, the image classification means 101 distinguishes between "concrete images labeled as completed" and "concrete images labeled as incomplete" by the shade of the line segment; that is, "concrete images labeled as completed" are shown with pale (thin) line segments, and "concrete images labeled as incomplete" are shown with darker line segments.

When making a judgment using the voting function, attention is focused on a set (hereinafter referred to as "image group GP") consisting of multiple concrete images arranged in chronological order of capture time, as shown in Figure 5. A judgment label (completed label/incomplete label) is assigned to each concrete image that makes up image group GP. Furthermore, in the case where concrete images are not divided, as shown in Figure 4(a), image group GP is composed of concrete images, while in the case where concrete images are divided, as shown in Figure 4(b), image group GP is composed of divided concrete images. Then, when there are more completed labels than incomplete labels among the judgment labels assigned to image group G, the judgment means 102 judges that compaction is complete for that concrete area.

For example, in the case of Figure 5, an image group GP is composed of seven concrete images captured consecutively. Furthermore, the determination means 102 performs a determination every time two frames of concrete images are acquired. At the seventh determination, image group GP7 contains two "concrete images labeled as completed." At the eighth determination, image group GP8 contains three "concrete images labeled as completed." However, since these are fewer than the "concrete images labeled as incomplete," the determination means 102 determines that compaction is incomplete for that concrete area. At the ninth determination, image group GP9 contains five "concrete images labeled as completed," and the determination means 102 determines that compaction is complete for that concrete area. The point in time at which compaction is complete can be the time associated with the last concrete image in image group GP9, which is determined to be compaction complete, as shown in the figure, or it can be set to the time associated with any concrete image included in image group GP, such as the first concrete image. Furthermore, the number of concrete images comprising image group GP can be designed arbitrarily, and the timing at which the determination means 102 makes its determination can also be designed arbitrarily.

(Display control means)
The display control means 109 is a means for displaying a concrete image on a display means 110 such as a display in accordance with the determination result of the determination means 102 (FIG. 2). That is, when the determination means 102 changes its determination from "compaction incomplete" to "compaction complete," the display control means 109 changes the display specifications of the concrete image related to that determination. Examples of the display specifications include the transparency of the image, the line color or line type of the frame line that surrounds the image (hereinafter referred to as the "image frame"), a display color for coloring the image, or a combination of these.

For example, in Figure 6, the display control means 109 controls the "image transparency" as a display specification, i.e., the concrete image is displayed with low transparency while compaction is incomplete, and is displayed with high transparency once the determination means 102 determines that compaction is complete. As a result, in Figure 6(a), since the concrete is in the early stages of compaction, many of the divided concrete images are colored black (i.e., low transparency), while in Figure 6(b), since compaction has progressed to a certain extent, most of the divided concrete images, except for the four corners, are displayed in a transparent state. Also, in Figure 7, the display control means 109 controls the "image frame display color" as a display specification, i.e., the image frame is displayed in white while compaction is incomplete, and is displayed in black once the determination means 102 determines that compaction is complete. As a result, in Figure 7(a), the concrete is in the early stages of compaction, so many of the divided concrete images have white image frames, while in Figure 7(b), compaction has progressed to a certain extent, so most of the divided concrete images, except for the four corners, are displayed with black image frames. Note that Figures 6 and 7 show examples of divided concrete images, but of course, in cases where the concrete image is not divided, the display control means 109 will control and display the entire concrete image.

(Example of use)
An example of using the compaction assessment system 100 of the present invention will be described with reference to Figure 8. Figure 8 is a flow diagram showing the main processing flow of the compaction assessment system 100 of the present invention, with the central column showing the processing to be performed, the left column showing information necessary for that processing, and the right column showing information resulting from that processing.

First, a trained model is generated by the model generation means 107 (Step 211 in Figure 8). It is recommended that the trained model be generated in advance before construction of the concrete structure begins. Once construction of the concrete structure begins, an image acquisition means 105 and a computer (PC or tablet PC) are placed at the concrete pouring site. This computer should be configured with an image classification means 101, a determination means 102, an input means 103, an image segmentation means 104, an image analysis means 108, and a display control means 109.

When fresh concrete is poured into the formwork and compaction work begins, for example using a high-frequency vibration vibrator, concrete images are acquired as video or continuous still images using the image acquisition means 105. The operator then uses the input means 103 to input fresh property data for the concrete being constructed, and the concrete images are input into the trained model under the control of the image selection means 106.

When a concrete image is input, the image segmentation means 104 segments the concrete image into multiple segmented concrete images (Step 221 in Figure 8). Next, the image analysis means 108 outputs the likelihood of compaction being complete or incomplete for each segmented concrete image (Step 222 in Figure 8), and the image classification means 101 assigns a judgment label (completed label/incomplete label) to the segmented concrete image based on the output result (Step 223 in Figure 8).

After a certain amount of time has passed (for example, when seven frames of concrete images have been acquired), an image group GP is set up using multiple divided concrete images (Step 224 in Figure 8), and the determination means 102 determines whether compaction is complete for the concrete areas corresponding to the divided concrete images. Once the determination means 102 determines that compaction is complete, the display control means 109 changes the display specifications and displays the divided concrete images on the computer's display means 110.

3. Compaction Determination Method Next, the compaction determination method of the present invention will be described with reference to Figure 9. Note that the compaction determination method of the present invention is a method for determining the compaction of fresh concrete using the compaction determination system 100 described up to this point, and therefore, we will avoid any explanation that overlaps with the content described for the compaction determination system 100, and will only describe content that is unique to the compaction determination method of the present invention. In other words, content not described here is the same as that described in "2. Compaction Determination System."

Figure 9 is a flow diagram showing the main steps of the compaction assessment method of the present invention. As shown in this figure, first, formwork is assembled based on the design plan and the required rebar is placed. Once these preparations are complete, fresh concrete is poured into the formwork (Step 10 in Figure 9). Next, compaction work begins using a high-frequency vibration vibrator or the like (Step 20 in Figure 9), and concomitantly, concrete images are acquired as video or continuous still images using the image acquisition means 105 (Step 30 in Figure 9).

Meanwhile, the operator uses the input means 103 to input fresh property data for the concrete being constructed, and the concrete image is input into the trained model under the control of the image selection means 106. When the concrete image is input, the image segmentation means 104 segments the concrete image into multiple segmented concrete images, the image analysis means 108 outputs the likelihood of compaction being complete or incomplete for each segmented concrete image, and the image classification means 101 assigns a judgment label (completed label/incomplete label) to the segmented concrete image based on the output result (Step 40 in Figure 9).

After a certain amount of time has passed, an image group GP consisting of multiple divided concrete images is set, and the determination means 102 determines whether compaction is complete for the concrete area corresponding to the divided concrete image (Step 50 in Figure 9). If the determination means 102 determines that compaction is incomplete for the concrete area (No in Step 60 in Figure 9), compaction work continues at the same location. On the other hand, if the determination means 102 determines that compaction is complete (Yes in Step 60 in Figure 9), the high-frequency vibration vibrator is moved and compaction work is performed at another location, and when compaction completion is confirmed for the entire construction area, the compaction work is terminated (Step 70 in Figure 9).

4. Verification Results The inventors evaluated and verified the compaction of fresh concrete using a trained model trained only on concrete images (hereinafter referred to as the "conventional model") and a trained model of the present invention trained on concrete images and fresh property data (hereinafter referred to as the "invention model"). Figure 10 is a graph showing the verification results, with (a) showing the results using the conventional model and (b) showing the results using the invention model. In Figures 10(a) and 10(b), the vertical axis represents the "number (No.) of the divided concrete image" and the horizontal axis represents the frame number (i.e., time) of the captured concrete image. As shown in Figure 10(c), two bars are displayed, one above the other, for each number of the divided concrete image. The upper bar represents the incomplete or completed state, and the lower bar represents the correct or incorrect judgment result, using white (correct) or black (incorrect). Comparing FIGS. 10(a) and (b), it is clear that the evaluation results of the conventional model output more incorrect answers, which confirms the effectiveness of the model of the present invention.

The compaction assessment system and method of the present invention can be used in the construction of a variety of concrete structures, including cast-in-place concrete work, the production of precast concrete in factories, and the production of precast concrete (so-called site PC) at construction sites. Given that the present invention provides properly compacted, high-quality concrete structures, it can be said to be an invention that not only has industrial applications but is also expected to make a significant contribution to society.

100 Compaction assessment system of the present invention 101 Image classification means (of compaction assessment system) 102 Assessment means (of compaction assessment system) 103 Input means (of compaction assessment system) 104 Image division means (of compaction assessment system) 105 Image acquisition means (of compaction assessment system) 106 Image selection means (of compaction assessment system) 107 Model generation means (of compaction assessment system) 108 Image analysis means (of compaction assessment system) 109 Display control means (of compaction assessment system) 110 Display means (of compaction assessment system) 111 Image storage means (of compaction assessment system) 112 Learned model storage means (of compaction assessment system) GP Image group

Claims (7)

A system for determining the compaction of fresh concrete using a concrete image obtained by photographing the surface of the fresh concrete,
an image classification means for assigning a completed label or an incomplete label to the concrete image according to a result output by a trained model by inputting the concrete image and fresh property data;
a determination means for determining compaction according to the number of completion labels assigned by the image classification means,
The fresh property data is data indicating properties of fresh concrete, and includes at least data on slump and air content,
The trained model is generated by machine learning the concrete images to which compaction is completed or incomplete and the fresh property data,
The image classification means assigns the completed label when the trained model determines that compaction of the fresh concrete related to the input concrete image is completed, and assigns the incomplete label when the trained model determines that compaction of the fresh concrete is incomplete.
A compaction assessment system characterized by: The trained model is composed of an image encoder, a fresh feature encoder, and a classifier;
the image encoder converts the input concrete image into an m-dimensional (m is a natural number) image feature vector;
The fresh property encoder converts the input fresh property data into an n-dimensional property data feature vector (n is a natural number);
the classifier converts the image feature vector and the property data feature vector into an (m+n)-dimensional integrated feature vector by concatenating the image feature vector and the property data feature vector, and outputs a likelihood of completion or incompletion of compaction of the fresh concrete based on the integrated feature vector.
2. The compaction determination system according to claim 1. the determination means determines that compaction is complete when the number of completed labels assigned to an image group consisting of a plurality of the concrete images arranged in order of photographing time is greater than the number of incomplete labels assigned to the image group.
3. The compaction determination system according to claim 1 or 2. Further provided with an input means by which an operator can input the fresh property data,
The compaction determination system according to any one of claims 1 to 3. The fresh property data further includes one or more data of slump flow, N-type penetration, and electrical conductivity.
5. The compaction determination system according to claim 1, wherein the compaction determination system is a compaction determination system for determining the compaction degree of a groundwater. further comprising image division means for dividing the concrete image into a plurality of divided concrete images;
the image classification means assigns the completed label or the incomplete label to each of the divided concrete images;
The determining means determines compaction for each region corresponding to the divided concrete image.
The compaction determination system according to any one of claims 1 to 5. A method for determining the compaction of fresh concrete using the compaction determination system according to any one of claims 1 to 6,
an image acquisition step of photographing the fresh concrete at predetermined intervals to acquire a plurality of images of the concrete;
an image classification step of assigning the completed label or the incomplete label to each of the plurality of concrete images by the image classification means;
A determination step of determining compaction by the determination means.
A method for determining compaction.