KR102311558B1 - System and method for detecting structure damage using artificial intelligence, and a recording medium having computer readable program for executing the method - Google Patents

Abstract

A structure damage detection system using artificial intelligence, a method, and a recording medium recording a computer readable program for executing the method are disclosed. A structure damage detection system using artificial intelligence includes a damage detection unit, a damage detection learning unit, and a virtual image generation unit. The damage detection unit detects damage to the structure in the observed image using the deep neural network, the damage detection learning unit learns the deep neural network of the damage detection unit using the virtual image, and the virtual image generator generates a virtual image. At this time, the virtual image generating unit generates a virtual image using the deep neural network, outputs the virtual image to a preset virtual image discriminator, and improves the deep neural network of the virtual image generating unit using the determination result of the virtual image discriminator.

Description

SYSTEM AND METHOD FOR DETECTING STRUCTURE DAMAGE USING ARTIFICIAL INTELLIGENCE, AND A RECORDING MEDIUM HAVING COMPUTER READABLE PROGRAM FOR EXECUTING THE METHOD }

The present invention relates to a damage detection technology for the maintenance of a concrete structure, and more particularly, to a system and method for detecting damage to the concrete of a structure using a photographed image and an artificial intelligence algorithm.

SOC facilities require periodic inspection for maintenance even after completion. For example, in the case of railroad tunnels, according to the results of analysis of the status of each year, the proportion of tunnels older than 20 years accounts for 41.8% of the total, so periodic inspection is essential. 1 is a table showing the status of railway tunnels by elapsed years.

On the other hand, damage information is recorded according to the subjective judgment and experience of the inspector when evaluating the condition of a conventional SOC facility, and thus different evaluation results are derived for each operation, thereby lowering objectivity and reliability of the results.

Accordingly, as a paradigm change for facility safety and maintenance using the 4th industrial revolution technology, a facility maintenance system using smart new technologies such as big data, 3D underground maps, drones, and robots is being built. However, the developed inspection robots are biased toward information collection tools rather than objective evaluation tools.

Therefore, automatic state evaluation technology using deep learning using image information for objective evaluation is necessary, but if it is difficult to collect image data, it is difficult to learn the deep neural network for state evaluation, so it is difficult to perform accurate state evaluation of structures. There is an impossible problem.

US 1020180043747 A

The present invention has been devised to solve the above-mentioned conventional problems, so that effective image data for applying a deep learning-based state evaluation technology can be efficiently constructed, and damage to a structure can be detected based on the constructed data. The purpose is to provide a system and method for doing so.

In order to achieve the above object, a structure damage detection system using artificial intelligence according to the present invention includes a damage detection unit, a damage detection learning unit, and a virtual image generation unit. The damage detection unit detects damage to the structure in the observed image using the deep neural network, the damage detection learning unit learns the deep neural network of the damage detection unit using the virtual image, and the virtual image generator generates a virtual image.

At this time, the virtual image generating unit generates a virtual image using the deep neural network, outputs the virtual image to a preset virtual image discriminator, and improves the deep neural network of the virtual image generating unit using the determination result of the virtual image discriminator.

According to this configuration, effective image data for applying deep learning-based condition evaluation technology can be constructed with virtual images generated using deep neural networks, and damage to concrete of structures can be efficiently detected based on the constructed data. there will be

In this case, the damage to the structure may be damage to the concrete of the structure.

In addition, the structure damage detection system may further include a virtual image discriminator, and the virtual image discriminator can determine a virtual image using a deep neural network, and improve the deep neural network of the virtual image discriminator using a virtual image and a real image. have. According to such a configuration, it is possible to more efficiently generate a virtual image through competition between the virtual image generating unit and the virtual image determining unit.

In addition, the damage detection learning unit separates the real image for learning and the real image for experimentation from the real image, learns the damage detection algorithm using the virtual image for learning generated using the real image for learning, and detects damage using the real image for the experiment The performance of the algorithm can be verified. According to such a configuration, it is possible to effectively perform the learning of the damage detection learning unit and the verification of the learning performance using a small amount of real images.

In addition, the damage detection unit includes a feature extraction unit for extracting features of the damaged area from the observed image, a bounding box calculation unit for calculating a bounding box including the damaged area, a damaged area dividing unit for dividing the damaged area from the damaged area, and the damaged area It may include a damage probability calculation unit for probabilistically calculating whether the damage of the.

In addition, the method for detecting structural damage using artificial intelligence according to the present invention includes a virtual image generation step in which a virtual image generation unit generates a virtual image, and a damage detection learning step in which the damage detection learning unit learns a deep neural network of the damage detection unit using the virtual image. , and a damage detection step in which the damage detection unit detects damage to concrete in the observation image using a deep neural network.

In this case, the virtual image generating unit generates a virtual image using the deep neural network, outputs the virtual image to a preset virtual image discriminator, and improves the deep neural network of the virtual image generating unit by using the determination result of the virtual image discriminator.

In addition, a recording medium recording a computer readable program for executing the method is disclosed together.

According to the present invention, it is possible to efficiently construct image data for applying a deep learning-based condition evaluation technology with a virtual image generated using a deep neural network, and to effectively detect damage to concrete of a structure based on the constructed data. do.

In addition, it is possible to more effectively generate a virtual image through competition between the virtual image generating unit and the virtual image determining unit.

In addition, it is possible to effectively perform the learning of the damage detection learning unit and the verification of the learning performance using a small amount of real images.

1 is a table showing the status of railway tunnels by elapsed years.
Figure 2 is a schematic block diagram of a structure damage detection system using artificial intelligence according to an embodiment of the present invention.
3 to 6 are examples of concrete crack images, and FIGS. 7 to 10 are examples of concrete peeling/thin film images.
11 is a view for explaining the concrete surface damage detection technology of the damage detection unit of FIG.
12 is a diagram for explaining a learning method to advance the performance of a deep neural network for detecting damage to a concrete surface.
13 is a view for explaining a concrete surface damage image generation technology.
14 is a view showing the main techniques presented in the present invention.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

2 is a schematic block diagram of a structure damage detection system using artificial intelligence according to an embodiment of the present invention. In FIG. 2 , the structure damage detection system 100 using artificial intelligence includes a damage detection unit 110 , a damage detection learning unit 120 , a virtual image generating unit 130 , and a virtual image determining unit 140 , , the damage detection unit 110 again includes a feature extraction unit 112 , a bounding box calculation unit 114 , a damaged part division unit 116 , and a damage probability calculation unit 118 .

The damage detection unit 110 detects damage to the structure in the observation image using a deep neural network. In this case, the damage to the structure may be damage to the concrete of the structure. 3 to 6 are examples of concrete crack images, and FIGS. 7 to 10 are examples of concrete peeling/thin film images.

To this end, the feature extraction unit 112 extracts the features of the damaged area from the observed image, the bounding box calculation unit 114 calculates a bounding box including the damaged area, and the damaged area dividing unit 116 is the damaged area. The damaged part is divided in , and the damage probability calculation unit 118 probabilistically calculates whether the damaged part is damaged.

More specifically, the detection algorithm of the damage detection unit 110 consists of input, feature extraction, determination, and output four steps, and the feature extraction step and determination step consist of a deep neural network. 11 is a view for explaining a concrete surface damage detection technology of the damage detection unit 110 .

As shown in FIG. 11, the feature extraction deep neural network plays a role in creating a feature map that can indicate the characteristics of the damaged area in the input image, and the deep neural network in the judgment stage plays a role of representing the damaged area as a bounding box, whether or not there is damage It is a multi-task-based algorithm that simultaneously performs the role of probabilistically representing .

The damage detection learning unit 120 learns the deep neural network of the damage detection unit 110 using a virtual image. At this time, the damage detection learning unit 120 separates the real image for learning and the real image for experimentation from the real image, uses the virtual image for learning generated using the real image for learning to learn the damage detection algorithm, and uses the real image for the experiment. Thus, the performance of the damage detection algorithm can be verified.

More specifically, the learning of the detection algorithm is characterized in that the detection performance is advanced by performing it using the virtual generated concrete damage image. 12 is a diagram for explaining a performance enhancement learning method of a deep neural network for detecting damage to a concrete surface.

In FIG. 12, first, to learn the deep neural network, virtual data is generated using the training data by dividing the training data and the experimental data, and the learning of the artificial intelligence algorithm for detecting damage to the concrete surface is performed by adding the virtual data to the training data. carry out After that, after applying the experimental data to the learned artificial intelligence, the learning is terminated according to the detection performance result, or the previous procedure is repeated to add new virtual data again to re-perform the learning.

The virtual image generating unit 130 generates a virtual image. At this time, the virtual image generation unit 130 generates a virtual image using a deep neural network, outputs the virtual image to a preset virtual image determination unit 140 , and uses the determination result of the virtual image determination unit 140 . Thus, the deep neural network of the virtual image generator 130 is improved.

In this case, the virtual image determining unit 140 may determine a virtual image using the deep neural network, and improve the deep neural network of the virtual image determining unit 140 using the virtual image and the real image.

More specifically, the virtual image generator 130 is configured to generate an infinite damage image from a small amount of concrete damage image using a virtual image generation technology using a generative adversarial network (GAN).

13 is a view for explaining a concrete surface damage image generation technology. First, the generator deep neural network of the virtual image generator 130 generates a random number of a 1 by 100 array generated between 0 and 1 and uses it as an input to create a virtual damaged image as an output.

Next, the discriminator deep neural network of the virtual image discriminator 140 divides the real and the virtual damaged images into true or false with a probability of 50%. It learns to be false to the discriminator with the virtual image created through the generator, and then learns to be true to the discriminator again with the real image.

After that, by connecting the generator and the discriminator, the virtual image created by the generator is called true, and only the generator part is learned. By repeating this method sufficiently, the complete generator deep neural network generates countless concrete damage images. By putting two deep neural networks in competition, the discriminator continuously learns the result of distinguishing between the virtual image created by the generator and the actual damaged image.

In summary, in the present invention, in order to secure learning data necessary for deep learning-based state evaluation using image information on concrete damage (cracks, peeling/exfoliation, etc.), unsupervised learning, a generative adversarial neural network, is used to We present a technology to create a virtual image of damage that can occur in the image and an artificial intelligence object recognition technology to detect damage based on the created image information.

More specifically, we propose a concrete surface damage image generation technology, a concrete surface damage detection technology, and a deep neural network performance enhancement learning method for the maintenance of concrete structures. 14 is a view showing the main techniques presented in the present invention.

According to the present invention, it is expected to be able to reduce economic losses due to damage by preventing safety accidents in advance by periodically inspecting road tunnels, railroad tunnels, common areas, and the like.

It is expected to be very useful as it can be sufficiently applied not only to the Hyperloop/Urban Loop, but also to the currently operated reinforced concrete tunnel and underground pipeline.

Although the present invention has been described with reference to some preferred embodiments, the scope of the present invention should not be limited thereto, but should also extend to modifications or improvements of the above embodiments supported by the claims.

100: Structure damage detection system using artificial intelligence
110: damage detection unit
112: feature extraction unit
114: bounding box calculation unit
116: damaged part division
118: damage probability calculator
120: damage detection learning unit
130: virtual image generation unit
140: virtual image discrimination unit

Claims (8)

a damage detection unit that detects damage to a structure in an observation image using a deep neural network;
Damage detection learning unit for learning the deep neural network of the damage detection unit using a virtual image; and
As a structure damage detection system using artificial intelligence comprising a virtual image generator for generating the virtual image,
The virtual image generation unit,
Generate the virtual image using a deep neural network,
Outputs the virtual image to a preset virtual image determining unit,
The deep neural network of the virtual image generation unit is improved by using the determination result of the virtual image discrimination unit,
The damage detection learning unit,
Separating the real image for learning and the real image for experimentation from the real image,
Learning the damage detector using a plurality of virtual images generated by the virtual image generator to differentiate it from the real image for learning using the real image for learning and the real image for learning, respectively,
Verifies the performance of the damage detection unit using the actual image for the experiment,
The virtual image generation unit generates a new virtual image again according to the performance of the damage detection unit,
The virtual image determining unit,
Determining the virtual image using a deep neural network,
Using the virtual image and the real image to improve the deep neural network of the virtual image discrimination unit,
The deep neural network of the virtual image generating unit is learned using the determination result performed by the deep neural network of the virtual image discriminating unit learned using the virtual image and the real image,
The deep neural network of the virtual image discrimination unit,
Structure damage detection system using artificial intelligence, characterized in that the virtual image is recognized as false and the real image is learned to be recognized as true, and then learned through competitive learning in which the virtual image is again learned to be recognized as true.
The method according to claim 1,
The damage to the structure is a structure damage detection system using artificial intelligence, characterized in that the concrete damage to the structure.
The method according to claim 1,
Structure damage detection system using artificial intelligence, characterized in that it further comprises the virtual image determination unit.
delete delete The method according to claim 1, The damage detection unit,
a feature extracting unit for extracting features of the damaged region from the observed image;
a bounding box calculation unit for calculating a bounding box including the damaged area;
a damaged area dividing unit for dividing the damaged area in the damaged area; and
Structure damage detection system using artificial intelligence, characterized in that it comprises a damage probability calculator for probabilistically calculating whether the damaged part is damaged.
a virtual image generating step of generating the virtual image by a virtual image generating unit;
Damage detection learning step in which the damage detection learning unit learns the deep neural network of the damage detection unit using the virtual image; and
As a structure damage detection method using artificial intelligence, the damage detection unit includes a damage detection step of detecting concrete damage in an observation image using a deep neural network,
The virtual image generation unit,
Generate the virtual image using a deep neural network,
Outputs the virtual image to a preset virtual image determining unit,
The deep neural network of the virtual image generation unit is improved by using the determination result of the virtual image discrimination unit,
The damage detection learning unit,
Separating the real image for learning and the real image for experimentation from the real image,
Learning the damage detector using a plurality of virtual images generated by the virtual image generator to differentiate it from the real image for learning using the real image for learning and the real image for learning, respectively,
Verifies the performance of the damage detection unit using the actual image for the experiment,
The virtual image generation unit generates a new virtual image again according to the performance of the damage detection unit,
The virtual image determining unit,
Determining the virtual image using a deep neural network,
Using the virtual image and the real image to improve the deep neural network of the virtual image discrimination unit,
The deep neural network of the virtual image generating unit is learned using the determination result performed by the deep neural network of the virtual image discriminating unit learned using the virtual image and the real image,
The deep neural network of the virtual image discrimination unit,
The method for detecting structural damage using artificial intelligence, characterized in that the virtual image is recognized as false and the real image is learned to be recognized as true, and then is learned through competitive learning in which the virtual image is again learned to be recognized as true.
A recording medium in which a computer readable program for executing the method of claim 7 is recorded.

Images