Corrosion of reinforcement concrete structure is a main cause of structural damage that required to repair or replacement. A non-destructive testing (NDT) method to find out the location and degree of corrosion-induced damage is needed. The NDT of ground penetrating radar (GPR) method has been found to be useful in evaluating existing reinforced concrete structures for continued use or for repair. This paper presents the study the degree of the corrosion using GPR based on image processing technique and artificial neural networks. The galvanostatic method employs direct current (DC) power supply with 5% sodium chloride (NaCl) solution for accelerating reinforcement bar corrosion before rebars induce to the reinforced concrete 1 x 1 x 0.3 m mixture. The rebars were corroded in 4 degree of corrosion (no corrosion, low corrosion, middle corrosion, and high corrosion). The 2 GHz of GPR, were used for detection the corrosion in reinforced concrete slabs after 28-days of standars moist curing. The GPR data showed that the corrosion can be detected, however need to analyzed and interpretation by using image processing and artificial neural network (ANN).
Keywords: NDT, GPR, galvanostatic method corrosion, image processing, and ANN.
INTRODUCTION
Corrosion of the steel rebar is the greatest factor in limiting the life expectancy of reinforced concrete structures. Corrosion of the steel rebars is caused either due to diffusion of the chloride ions to the steel surface or due to carbonation of concrete. Corrosion of the steel rebars and the subsequent cracking of concrete due to the ingress of chloride ions to the steel surface is more predominant than that due to carbonation of concrete [1]. To assess the condition of corrosion in the concrete structure, a number of non-destructive testing (NDT) methods have been recently studied.
Some NDT methods widely used test to assess the likehood of rebar corrosion, like half-cell potencial. However, the test does not allow the detection of delamination in a direct manner. It provides an indication of the state of corrosion activity, and in some cases information on the possible presence of damage if this corrosion is in advanced stage. Because of these conditions, engineers prefer to use the ground penetrating radar (GPR) technique. This choice is justified by the fact that this electromagnetic technology makes it possible to collect the data in a fast way. Unfortunately, the GPR technique is not yet completely accepted by engineers because its reliability towards the detection of corrosion is not satisfactory [2].
The GPR is becoming more and more popular as a concrete inspection method. The GPR is significant technology for locating embedded targets in concrete. The NDT method of GPR allow a reliable and efficient inspection of the structural integrity of reinforced concrete [3]. However, the results of GPR is very difficult to interpret and may require the skills of an experienced operator and the use of lengthy manual post-processing and subjective expertise to produce a reliable end result [4-8].
Recent years, many automatic techniques have been developed for interpreting the GPR data. Neural network, signal and image processing techniques employs to provide a high resolution image, provide accurate depth and location information, and facilitate straightforward data interpretation. However, the success has so far been limited to straightforward cases such as buried object location [9, 10].
Therefore, in this work we investigate the degree of corrosion in concrete structures by proposing the usage of image processing technique to extract features differentiate corrosion and no corrosion concrete structures from GPR data and to get the best interpretation for corrosion detection. In addition, for the further process the usage of artificial neural networks for decision of data whether corrosion or no corrosion is applied.
literature
Ground Penetrating Radar (GPR)
The potency of non-destructive methods of ground penetrating radar (GPR) for involving the transmission of electromagnetic waves into a material is under investigation. The reflections of these waves at interfaces and objects within the material have been analyzed to determine the location or depth of these interfaces and buried objects, and to determine the properties of the material. Mostly GPR is utilized in reflection mode which a signal is emitted via an antenna into the structure below investigation. The arrival time and the simpleness of reflected signals caused by replacement in material properties is written and examined.
GPR Applications in image processing and ANN
Study done by [9] have developed system comprises a neural network classifier, a pattern recognition stage, and additional pre-processing, feature-extraction and image processing stages to provide a high-resolution image of the sub-surface in near real-time facilitating straight forward data interpretation and providing accurate depth and azimuth location information of the rebars. In addition, [11] had developed guidelines and recommendations for ground penetrating radar data acquisition and interpretation. By combining information extracted from various cues from within the data in a manner that minimizes the reliance on ready-made assumptions, rules of thumb and conjecture, it is possible to improve the reliability and accuracy of the final interpretation result.
In [12], the authors had achieved by subjecting ground penetrating radar radargrams to a series of image processing stages followed by a curve-ï¬tting procedure speciï¬cally developed for hyperbola. The ï¬tting technique was applied on a variety of real hyperbolic signatures that were
collected from a controlled test site. They had obtained the results indicating that this technique was fully capable of successfully estimating the depth and radius to within 10%, which validates the method and justify the used assumptions. Whereas, [13] had showed that the use of a multilayered perceptron (MLP) neural network approach could be quite effective in automating the identiï¬cation and location of embedded steel reinforcing bars from a GPR investigation. Accurate estimation of depth, or cover, requires a reliable knowledge of the dielectric properties of the concrete, and recent work using a specially developed wideband horn antenna for direct determination of in situ properties was also outlined.
In [14], the auhors had aimed at detecting and characterizing inclusions in concrete structures by inverting ground-penetrating radar data. Moreover, with 99.99% of the original variance the data needs only 139 dimensions. This dimensional reduction can make the artiï¬cial neural network (ANN) training easier and faster. The artiï¬cial neural network were trained to ï¬nd the buried inclusions characteristic and considering a non-homogenous host medium by inverting the pre-processed data. The results show that the expected maximum error was kept under 1%, which is a remarkable result, since the host medium is non-homogenous.
Materials and Methods
Sample Preparation and Data Acquisition
In this work the two reinforced concrete slabs dimension l = 1 m, w = 1 m, and h = 0.3 m have prepared. The concrete grade is C30. Portland cement, uncrushed sand, crushed limestone with a maximum aggregate size of 20 mm were used to prepared the concrete mixture. Details of the mixture proportions are given in table 1. The y-type reinforced bars (rebars) with length approximately 1 m are selected with diameter is 20 mm.
Table 1. Concrete Mixture Proportions
Concrete Mixtures
Quantity
Density (kg/m3)
2430
Cement Content (kg/m3)
380
W/C
0.5
Per m3
Per trial mix (0.3 m3)
Cement (Kg)
380
114
Water (Kg or L)
190
57
Fine Aggregate (Kg)
780
234
Coarse Aggregate (Kg)
1080
324
Galvanostatic Method for accelerating corrosion
At first, the rebars is immersed in a solution of 5% NaCl (0 day for no corrosion, 1 day for low corrosion, 3 days for middle corrosion, and 7 days for high corrosion) using direct current (DC) power supply. The direction of current was adjusted so that the corrosing rebars served as the anode, while a bar facing the corrosing rebars served as the cathode. The current of 10V (volt) and 1A (ampere) were applied in the corrosion process. During the corrosion process, the electric current should be kept constant. The process was continued until the rebars corroded with different degree. Last, the corroded rebars are induced to concrete mixture.
GPR Testing
The 2 GHz of GPR manufactured by IDS, were used for detection the corrosion in reinforced concrete slabs after 28-days of standars moist curing. The result could be proposed in a-scan, b-scan, c-scan, and 3D image as tabulated in figure 1. In this paper we used 3D image as GPR data for corrosion detection. The 3D image of GPR data showed that the corrosion can be detected only for high corrosion, however the GPR can not visualize for low and middle corrosion because the image is not clear, as shown in figure 2.
Image Processing Techniques
There are large number applications of image processing in diverse spectrum of human activites i.e. cancer detection in biomedical images, computer science, aggregate shape recognition in material, faulty component identification, remotely sensed scene interpretation. Imaging of concrete structures have presented many challenges due to the fact that concrete is a non-homogeneous material [15].
In this paper we use two image processing technique i.e. K-means clustering, and edge detection to detect and investigate corrosion and non corrosion image from GPR data of concrete structures.
K-Means Clustering
Segmentation using K-means clustering refers to the process of partitioning a digital image into multiple segments. The image segmentation results can be used to derive region-wide color and texture features, which in turn, together with the segment location, boundary shape, and region size, can be used to extract semantic information [16]. The goal of this technique is to simplify the representation of an image into something that is more meaningful and easier to analyze. The result
(a)
(b)
(c)
(d)
Figure 1. The results of GPR (a) Typical A-scan, (b) Typical B-scan, (c) C-Scan, and (d) 3D image
Figure 2. The 3D image of GPR
of image segmentation is a set of segments that collectively cover the entire image. Each of the pixels in a region are similar with respect to some color characteristics [17].
In this work the algorithm of color based segmentation using K-Means Clustering was applied using matlab toolbox. The data given by GPR is clustered by the k-means method, which aims to partition the points into k groups of color images.
Edge Detection
Edge detection is a kind of method of image segmentation based on range non-continuity. Edge always appears in two neighboring areas having different grey level. It exists between object and background, object and object, region and region, and between element and element [18]. This study uses Sobel edge detectors for convolution the GPR data images. These edge detectors have two masks producing edge image for detecting an edge in K-means clustering images of GPR data.
The Sobel operator counts difference using weighted for 4 neighborhoods [19]. The Sobel operator has the similar function as the Prewitt operator, but the edge detected by the Sobel operator is wider. Sobel operator can process those images with lots of noises and gray gradient well. We order that :
S= (a+ ca+a)- (a+ ca+ a), and
S= (a+ ca+a)- (a+ ca+ a)
With c = 2, mask sand s can be declared as :
S= dan S=
M= (1)
The direction of this point is:
(2)
The original image of GPR data, K-means clustering image and the edge detection drawing of Sobel operator gained using Matlab 7 simulation are shown in figure 3.
(a)
(b)
(c)
(d)
Figure 3. (a) GPR data, (b) K-means clustering image, (c) binary image and (d) Sobel edge image
Artificial Neural Network (ANN)
The most commonly Artiï¬cial Neural Networks (ANNs) used for pattern recognition and classiï¬cation [20] is a multilayer perceptron (MLP) networks. MLP networks with m outputs and n hidden nodes can be expressed as:
;
for 1 and 1
where w and w denote the weights of the connection between input and hidden layer and the connection between hidden and output layer, respectively. Band v denote the thresholds in hidden nodes and inputs that are supplied in the input layer, respectively; n and n are the number of input nodes and hidden nodes respectively. F(·) is an activation function that is normally selected as a sigmoidal function [21]. The MLP model is shown in Figure 4.
Figure 4. The MLP networks
The input layer, hidden layer, and output layer compiled up to down respectively. In this current study we use 4 features i.e. sum of white pixels, sum of black pixels, area of rebar, and perimeter of rebar of GPR data as input fed into MLP networks and 2 output i.e. corrosion and no corrosion. Four features were obtained from processed images which images segmented were taken no corrosion and corrosion separately with size 100x200 pixels as shown in Figure 5 and a part data of used features were tabulated in Table 2
Features extraction for sum of black pixels, white pixels and area was calculated based on binary image as shown in Figure 3(c). However, perimeter of object was calculated based on sobel edge image as shown in Figure 3(d).
(a)
(b)
Figure 5. (a) no corrosion segmentation, and (b) corrosion segmentation
Table 2. Part data of used features (pixels)
Conditions
Black
White
Area
Perimeter
Corrosion
7817
12183
14109
4060
Corrosion
7640
12360
14550
5498
Corrosion
7530
12470
14333
5488
No corrosion
12695
7305
19973
5693
No corrosion
12770
7230
20036
5652
No corrosion
12449
7551
19615
5734
*100x200=20000
Pixel/image
Before the features of GPR data was fed into MLP networks, the data was normalized and arranged using 10 fold cross validation. Data classification is tabulated in Table 3. Thus, ten datasets were used to show range of classification results for corrosion and no corrosion.
Table 3. GPR data classification
Trainset
Testset
Corrosion
90
10
No corrosion
27
3
Total
117
13
Total data
130
RESULTs
All data were obtained from GPR scan of sample in concrete laboratory civil engineering University Sains Malaysia. The allocations for training and testing has showed in Table 3 with ten datasets. As mentioned before, four features were extracted from each data image as input fed into ANNs. The MLP networks trained with Levenberg-Marquardt Backpropogation were used for training ten datasets.
For this result section, we presented the classification performances based on accuracy of ANNs for showing the capability of GPR data to detect corrosion and no corrosion in concrete structure. The result has showed that all of the ten datasets performed excellent classification with all datasets accuracies value 100% for training and testing as shown in Table 4.
Table 4. Performance of accuracies MLP trained with LM
Datasets
Epoch
Hidden
AccTrain
AccTest
Train1
42
1
100
100
Train2
8
1
100
100
Train3
5
1
100
100
Train4
14
1
100
100
Train5
9
1
100
100
Train6
9
1
100
100
Train7
8
1
100
100
Train8
8
1
100
100
Train9
17
1
100
100
Train10
10
1
100
100
According to hidden nodes and epoch nodes used by ANNs, we analyzed that these data were not complicated data due to high accuracy performance of ANNs classification. Most probably, these results were influenced by effective image processing technique for segmentation the object of interest and other object in images.
CONCLUSIONs AND DISCUSSIONs
This paper presented the observation of GPR data with two conditions of concrete structures. Image processing techniques were used to preprocess the image data to obtain features in order to classify the image into two class using artificial neural networks fast, clearly and easily.
The general accuracy results presented in Table 3 has proven to be effective for detecting the corrosion and no corrosion in concrete structures using GPR data. The main advantages of ANNs for classification is fast and better result than other statistic tools especially the complicated data. The authors of this paper believe that the used techniques could be used to classify the other damage images.
