The durability performance of reinforced concrete (RC) building structures is significantly affected by the corrosion of the steel reinforcement due to chloride penetration, thus, the chloride ion diffusion coefficient should be investigated through experiments or theoretical equations to assess the durability of an RC structure. This study aims to predict the chloride ion diffusion coefficient of concrete, a heterogeneous material. A convolutional neural network (CNN)-based regression model that learns the condition of the concrete surface through deep learning, is developed to efficiently obtain the chloride ion diffusion coefficient. For the model implementation to determine the chloride ion diffusion coefficient, concrete mixes with w/c ratios of 0.33, 0.40, 0.46, 0.50, 0.62, and 0.68, are cured for 28 days; subsequently, the surface image data of the specimens are collected. Finally, the proposed model predicts the chloride ion diffusion coefficient using the concrete surface image data and exhibits an error of approximately 1.5E−12

In the construction industry, the demand for reinforced concrete (RC) structures with high durability, is increasing [

Chloride ions, through penetration, eventually reach the surface of steel reinforcement and corrode the steel reinforcement, thereby reducing its durability. Therefore, it is important to measure the time period within which chlorides can reach the surface of steel reinforcement. The chloride ion diffusion coefficient represents the chloride ion diffusion rate that implies the time elapsed between the penetration of chloride ions into concrete and corrosion of the steel reinforcement. It represents the resistance of an RC structure to chloride penetration and is used as the main parameter to determine the durability performance of an RC structure [

In concrete, various pores and microcracks are present owing to the air bubbles generated during hardening or material separation. Generally, chloride ions diffuse into the concrete through pores and microcracks. Therefore, chloride penetration occurs through the microcracks at the interface between the aggregate and cement, for example, penetration through the concrete defects, such as the relatively large cracks or the numerous pores in hardened cement. Therefore, the condition and characteristics of the concrete surface are important factors in determining the chloride ion diffusion coefficient.

This study aims to propose a feasibility model to predict the chloride diffusion coefficient based on the concrete surface images. This study proposes a method for estimating the chloride ion diffusion coefficient by analyzing the material properties appearing on the surface of a concrete specimen using the images of the surface. In this method, a neural network that utilizes a convolution filter was applied to predict the chloride ion diffusion coefficient. The performance of the proposed model was verified by comparison with the experimental results collected through the NT build 492 experiment.

Chlorides penetrate into concrete through micropores, aggregates, interfacial transition zones (ITZs), and cracks on the concrete surface [

According to the literature, numerical data can be estimated by connecting the image data and numerical values. Regression is used to predict the target value (chloride diffusion coefficient) of this study by inputting the surface information of heterogeneous materials to the CNN of the proposed method. The concrete surface image data include features in the form of the height × width pixel matrix through image conversion. Each pixel includes features that appear on the concrete surface, and the convolution kernel extracts the features in the image by dividing the image into small blocks. Through the convolution operation (

An image can be expressed using

where

Over the last 10 years, CNNs have been intensively investigated and applied in a wide range of research fields [

Cement, a major component of concrete, generates hydration products through a continuous hydration reaction; further, the hydration products reduce the pores. Such reduction in the pores and increase in the hydration products are essential factors for determining the chloride ion diffusion coefficient because such factors essentially act as defense mechanisms for effectively resisting the chlorides introduced from outside. Therefore, in this study, six different concrete mixes were subjected to water curing for 28 days, and a comparative test was conducted.

Design | W/C | S/a | Weight per unit volume (kg/m^{3}) |
Ad. | Slump | Air | |||
---|---|---|---|---|---|---|---|---|---|

strength (MPa) | (%) | (%) | (B%) | (mm) | content (%) | ||||

W | C | Sand | Gravel | ||||||

21 | 62 | 50.0 | 185 | 300 | 874 | 921 | 0.4 | 150 | 4.8 |

27 | 46 | 47.0 | 178 | 390 | 795 | 945 | 0.4 | 150 | 4.6 |

30 | 40 | 47.0 | 165 | 410 | 803 | 955 | 0.4 | 150 | 4.5 |

18 | 68 | 51.0 | 170 | 250 | 932 | 944 | 0.5 | 150 | 4.9 |

24 | 50 | 48.0 | 165 | 330 | 852 | 973 | 0.5 | 150 | 4.6 |

40 | 33 | 45.0 | 160 | 480 | 749 | 965 | 0.6 | 150 | 4.4 |

The chloride ion diffusion coefficient was obtained using the NT Build 492 test method [_{2}) and combined with a rubber container. Thereafter, an electrical potential difference is applied by filling the anode with a 0.3 M NaOH aqueous solution and a cathode with a 10% NaCl aqueous solution. In this instance, the first applied voltage is 30 V. The current value is immediately measured upon the application of the voltage to correct the applied voltage (U). After examining whether a valid initial current value (I_{0}) is obtained, an accelerated test is conducted during the corresponding time (t). After the application of the potential difference, the specimen is disassembled and split in the axial direction. A 0.1 N silver nitrate (AgNO_{3}) aqueous solution is sprayed onto the split surface to examine the chloride ion penetration depth. When the silver chloride precipitate is clearly visible after approximately 15 min, the chloride ion penetration depth is accurately measured up to a unit of 0.1 mm using a vernier caliper, and the average value is obtained. The chloride ion diffusion coefficient of the specimen can be estimated by substituting the test conditions and measurement results into

where D is the chloride ion diffusion coefficient (m^{2}/s), F is the Faraday constant (9.648 × 10^{4} J/V·mol), R is the gas constant (8.314 J/K·mol), L is the thickness of the specimen (m), t is the test duration (s), _{3}; 0.07 N), z is the valence of the ion (z = 1 for chloride ion), U is the voltage difference between the anode and cathode (V), T is the temperature of the solution (K),

Design strength | Chloride ion diffusion coefficient (m^{2}/s): 28 days |
||
---|---|---|---|

(MPa (W/C)) | a | b | c |

18 (68%) | 9.76E−12 | 1.18E−11 | 9.92E−12 |

21 (62%) | 6.89E−12 | 8.21E−12 | 7.33E−12 |

24 (50%) | 9.88E−12 | 7.09E−12 | 8.83E−12 |

27 (46%) | 5.39E−12 | 4.29E−12 | 4.88E−12 |

30 (40%) | 4.62E−12 | 5.56E−12 | 5.50E−12 |

40 (33%) | 3.37E−12 | 4.17E−12 | 4.74E−12 |

In this study, image processing was utilized to estimate the chloride ion diffusion coefficient. The prediction of the chloride ion diffusion coefficient of concrete through image processing has been performed in the following order: feature extraction, feature selection, and post-processing. Among these steps, feature extraction is the most important and difficult step. As mentioned previously, concrete is a heterogeneous material, and the chloride ion diffusion coefficient is affected by the complex spatial distribution of aggregates, cement hydration products, and pores. Because the components are heterogeneous and have irregular shapes, there are limitations in the manual extraction of all the features that affect the compressive strength. To overcome these limitations, an attempt is made to utilize a deep convolution neural network (DCNN) that has exhibited an excellent prediction performance.

Therefore, it is necessary to capture the images and photographs of the concrete specimens fabricated in a laboratory environment, calculate the chloride ion diffusion coefficient through a test, and train the model by labeling the calculated chloride ion diffusion coefficient on the captured data. The dataset constructed as images was acquired using a DSLR camera sensor. The size of the image obtained using the DSLR camera was 6000 × 3376 pixel.

For a large image size, a large amount of information is available; however, it may interfere with learning owing to the high dimensionality and overlapping pixels. If the image size is extremely small, the performance of the DCNN deteriorates owing to a significant reduction in the amount of information. Therefore, determining the optimal input image size is very important during the design of a DCNN algorithm for its performance and learning. Because the performance of an algorithm significantly depends on the image size, multiple studies have been conducted to determine the optimal image size for a new image dataset. Therefore, in this study, an experiment was performed using the following three image dimensions: 224 × 224, 112 × 112, and 84 × 84.

This study intends to perform an experiment on the chloride ion diffusion coefficient. Therefore, the images of the concrete specimens were obtained using the DSLR camera, and the chloride ion diffusion coefficient was used as the target value for each image. The chloride ion diffusion coefficient obtained through the NT Build 492 described above, was a real value ranging from 3.37E–12 to 1.18E–11. Therefore, for the proposed model, each chloride ion diffusion coefficient was replaced with values ranging from 1.18 to 11.8, to be used as the output value. Consequently, 9,726 images were used as the training (7,780) and validation (1,946) data, and 4,169 images were used as the test data.

In this study, a DCNN algorithm structure that can predict the chloride ion diffusion coefficient, is designed; the diffusion coefficient is an important index for the performance assessment of a concrete structure and is determined using the DSLR data obtained by capturing the images of concrete specimens. Determining the optimal structure for learning features that facilitate predicting the chloride ion diffusion coefficient, is extremely important for the performance and training of the algorithm. In addition, multiple studies have been conducted to determine the optimal structure from various types of data, and each structure exhibits a significant difference in the performance. In the field of deep learning, it has been experimentally proven that a deep neural network structure can demonstrate a high accuracy if sufficient data are available. However, given that the number of datasets is not sufficient in practical environments, the performance of DCNN models could be influenced not only on the architecture of networks but also by the image size of the dataset. Because the amount of information in data could affect to recognize the features extracted from the surface images of heterogeneous materials such as concrete. Thus, this study experimented with the various DCNN models with different image sizes.

The structure of a DCNN is determined by various factors, such as the number of layers, type of each layer, number of output neurons for each layer, activation function, and weight connection. The layer types include the fully connected layer, convolutional layer, normalization layer, and pooling layer. In addition, there are various parameters for each layer that cause a variation in the overall structure. Therefore, the structure of the DCNN is determined based on numerous combinations. Because there is still no established theoretical method to determine the optimal structure of a DCNN, the optimal structure must be determined based on the empirical trial-and-error method. In this study, three approved DCNN structures (VGGNet, Inception and ResNet) were used, and simpler CNN were experimented for confirming the most suitable architecture to solve the regression problem using the image data. The experiments were performed by changing the structures according to the experimental data to determine the optimal structure.

The VGGNet architectures used in this study are adopted 16 parameter layers from the reported study [

To verify the feasibility of the presented algorithms, the input and output of the DCNN model were defined, and the data of the DCNN model were preprocessed and amplified for matching the experimental data with the target value. The algorithms were modified according to the input size of the concrete data, and the algorithm structures were optimized for the concrete data such that the algorithms could efficiently track the performance target value. For the chloride ion diffusion coefficient, a single output characterization model was constructed by applying the above several algorithms, and the model was evaluated. To predict the chloride ion diffusion coefficient as a real number, the mean square error loss shown in

When the parameter of the deep learning-based model is

A representative method to minimize this loss function is the gradient descent that facilitates determining the optimal parameter

In this study, the initial learning rate was set to 0.00001. Further, Keras, a deep learning framework, was used to train the chloride ion diffusion coefficient prediction model in this study.

The conjecture of this study is that a DCNN that can learn data-based features by itself will be able to learn features in concrete images that may facilitate predicting the chloride ion diffusion coefficient. To verify this conjecture, an experiment was performed by constructing a single DCNN for the chloride ion diffusion coefficient, using the image data. Initially, the image dimension was fixed at 84 × 84, and ^{4} iterations to estimate the chloride ion diffusion coefficient of concrete. It was found that both the train loss and validation loss decreased as the number of iterations increased. This implied that the proposed model was well-trained without overfitting. Therefore, the experimental results confirmed that the DCNN could learn the patterns included in the image data for predicting the chloride ion diffusion coefficient.

In this study, the prediction of chloride ion diffusion coefficient of a concrete structure was attempted using the simply modified VGG16 [

Predicted chloride ion diffusion coefficient | Measurements | ||||
---|---|---|---|---|---|

Image size | Model name | RMSE | MAPE | MAE | R^{2} |

Customized_CNN | 0.6905 | 8.7591 | 0.5367 | 0.8898 | |

VGG16 [ |
0.4435 | 4.9942 | 0.3148 | 0.9531 | |

ResNet50 [ |
0.2778 | 2.4262 | 0.1503 | 0.9786 | |

InceptionV3 [ |
0.4072 | 2.1768 | 0.1411 | 0.9483 | |

112 × 112 | Customized_CNN | 0.7141 | 8.7762 | 0.5590 | 0.8841 |

VGG16 [ |
0.4776 | 5.3630 | 0.3396 | 0.9462 | |

ResNet50 [ |
0.1686 | 1.7651 | 0.1019 | 0.9914 | |

InceptionV3 [ |
0.3519 | 1.9379 | 0.1180 | 0.9614 | |

224 × 224 | Customized_CNN | 1.0611 | 13.0731 | 0.8421 | 0.7467 |

VGG16 [ |
0.2477 | 3.0313 | 0.1819 | 0.9851 | |

ResNet50 [ |
0.0972 | 0.9706 | 0.0585 | 0.9958 | |

InceptionV3 [ |

When the several experimental metrics of the various models were compared through

In this study, an algorithm for predicting the chloride ion diffusion coefficient based on the concrete images was designed using deep learning, and it was verified for the actual application. With the aging of domestic concrete structures, the importance of assessing the durability of concrete also increases. Among the basic indices for assessing the durability of concrete, the chloride ion diffusion coefficient, an important parameter for assessing the durability of concrete, generally utilizes two methods (the natural diffusion test and accelerated diffusion test methods). Considering that the natural diffusion test is time-intensive, various accelerated diffusion test methods have recently been used to reduce the overall experimental time. Although recent studies have been conducted using image processing techniques, there are limitations in directly extracting features from the surface of concrete, a heterogeneous material. This study proposes a model that can automatically extract the unique features appearing on a concrete surface using a CNN and predict the chloride ion diffusion coefficient based on such features. In addition, the validity of the model was verified through the application of the concrete chloride ion diffusion coefficient dataset constructed in a laboratory environment.

The surface image data of the concrete specimens were input to the proposed model, and the architecture was implemented through a combination of CNN and regression. For the optimization of the model, the method of minimizing the loss (

This study suggests the applicability of deep learning to the field of facility maintenance by proposing a method for estimating the chloride ion diffusion coefficient of concrete using images. This study has the limitations of the inappropriate representation of the various concrete characteristics in a real environment because the data that were collected were those of the laboratory-refined concrete specimens. Therefore, in the future, it will be necessary to conduct research on the estimation of the chloride ion diffusion coefficient under various deterioration phenomena of the concrete surface, such as microcracks and spalling. Moreover, only image data were used in this study; if various other variables, such as concrete mix information, geographic environment information, and time information, are used for learning, it will be possible to construct a more accurate and sophisticated model.

The authors would like to thank the National Research Foundation of Korea for funding this research project.