Construction and Building Materials 297 (2021) 123785

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

Machine learning model for predicting the crack detection and pattern
recognition of geopolymer concrete beams
Aravind N a, Nagajothi S b,⇑, Elavenil S b
a
Department of Civil and Environmental Environment, National University of Science and Technology, Oman
b
School of Civil Engineering, Vellore Institute of Technology, Chennai Campus, Chennai 127, India

h i g h l i g h t s

 Crack acquisition of experimented geopolymer concrete beams with BFRP & GFRP bars.
 Well-equipped image pre-processing python packages used to collect crack patterns.
 Automated quality check using machine learning python packages.
 Support Vector Machine (SVM) is used to classify the failure patterns.
 SVM compared with other five machine learning classifiers using confusion matrix.

a r t i c l e i n f o a b s t r a c t

Article history: One of the major challenges in the construction industry is the detection of cracks in concrete structures
Received 10 January 2021 and identification of failure types of these structures that lead to their degradation. Manual quality
Received in revised form 19 April 2021 checks are prone to human error, and require longer response time and specialist experience and knowl-
Accepted 27 May 2021
edge. Therefore, visualizing the cracks and identifying failures in concrete structures using computer
techniques is now a preferred option. The present work focuses on identifying the cracks using image
processing and failure pattern recognition technique by employing suitable machine learning algorithms,
Keywords:
and validating the techniques using Python programming. For this purpose, M30 grade geopolymer and
Geopolymer
Pre-processing
conventional concrete beams were cast using Basalt Fibre Reinforced Polymer/Glass Fibre Reinforced
Classifiers Polymer and Steel bars. The beams were subjected to four point static bending test by varying the shear
Confusion matrix span to the effective depth ratio. The experimental images were used for image processing and failure
Machine learning pattern recognition in Python language. Employing six machine learning classifiers, the failures in the
Python structures were classified into three classes namely, flexure, shear, and compression. The machine learn-
ing classifiers were also adopted to determine the confusion matrix, accuracy, precision, and recall scores.
It was found that among the six classifiers used, the support vector classifier gave the best performance
with 100% accuracy in identifying the failure patterns.
Ó 2021 Elsevier Ltd. All rights reserved.

1. Introduction nosilicate materials are fly ash, Ground Granulated Blast furnace
Slag, slag cement, slag, metakaolin and silica fume. Alkali solution
Concrete is the main base material for constructing a building is a mixture of sodium silicate and sodium hydroxide along with
with cement being the major ingredient of concrete. Cement man- water or a mix of potassium silicate and potassium hydroxide
ufacturing industries release large amounts of carbon dioxide into along with water. It is used for casting of structural members,
the atmosphere. Geopolymer concrete is an innovative and sus- pavement construction, brick/ block manufacturing work, etc.
tainable construction material which consists of three components When buildings are exposed to aggressive environments in the
namely, a source of aluminosilicate materials, fine and coarse long term, damages occur which cause degradation of the build-
aggregate, and an alkali activating solution. Examples of alumi- ings. The factors involved in degradation of concrete are salt ero-
sion, frost damage, dry shrinkage, earthquakes, and rain water
[1]. The degradation levels and carrying capacity of reinforced
⇑ Corresponding author. concrete structures are mainly influenced by the width, length,
E-mail addresses: [email protected] (N Aravind), naga.jothis2014phd1138@ type, and the number of cracks in the structures [2]. Hence, the
vit.ac.in (S Nagajothi), [email protected] (S Elavenil).

https://doi.org/10.1016/j.conbuildmat.2021.123785
0950-0618/Ó 2021 Elsevier Ltd. All rights reserved.
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

assessment of cracks in concrete structures is important for inspec- in concrete bridges. Deep fully convolution neural network (FCN)
tion, diagnosis, maintenance, and safety life prediction. Crack method (using VGG16) was used by Dung and An [22] for detecting
detection by visual inspection requires more experience, special- the cracks in concrete. Fujita et al. [23] used two pre-processing
ist’s knowledge, and is tedious, time consuming, and subjective. methods namely, subtraction and line extraction, and a threshold
Currently, automatic image-based crack detection methods are process to separate the cracks from the background in order to
used along with visual inspection. Many automatic crack detection assess their accuracy and firmness [24].
methods have been proposed during the past few years [3,4]. Due Manufactured sand (M sand) is used currently due to the scar-
to irregularities in crack size, shape, unevenly elucidated condi- city of river sand. Steel corrosion is a major problem and requires
tions, and blemishes, crack segmentation from the crack images expensive maintenance, repair, retrofitting, and rehabilitation in
is very difficult. Hence, crack detection is very important, specifi- reinforced concrete structures. This problem is overcome by using
cally for the variety of cracks in complicated backgrounds. In this reinforced concrete structures such as Fibre-Reinforced Polymer
study, a crack detection technique is proposed which employs (FRP) bars that are non-corrosive, light weight, and environment-
pre-processing of images for extracting crack patterns and classi- friendly in nature. This research mainly focuses on studying the
fiers for segregating the class of crack patterns. experimental shear behaviour of geopolymer concrete beams using
The paper is organized as follows: Section 2 presents the rele- BasaltFibre-Reinforced Polymer (BFRP) and Glass Fibre-Reinforced
vant review of literature, section 3 describes the materials and Polymer (GFRP) bars by varying the shear span to the effective
methods, section 4 presents the crack analysis, section 5 presents depth ratio, crack detection using image processing, and analytical
results and discussion, and section 6 concludes the paper. failure pattern recognition using machine learning algorithms in
python script.

2. Related work
3. Materials and methods
Several research works have been carried out in the field of
crack detection in geopolymer concrete and the techniques used This section first details the materials and methods used in the
to measure the crack dimensions. Most recently, Amer Hassan study. Next, the experimental setup is described. Finally, the anal-
et al. have performed an intensive study on geopolymer concrete. ysis methodology using image processing and machine learning is
Based on their study, it has been concluded that geopolymer con- presented.
crete is a sustainable construction material thus presenting a new
technology that is of significance in the construction industry [5]. 3.1. Materials used
Also, as a popular material it has garnered significant attention in
the construction industry due to its useful by-product materials, The materials used in the study for making geopolymer and
eco-friendly (low ‘embodied energy’ and low ‘embodied CO2 emis- conventional concrete are described in the following sub-sections.
sion’) nature, non-requirement for water curing, economic viabil-
ity, and conventional concrete-like properties [6,7,8]. Geopolymer 3.1.1. Geopolymer and conventional concrete
concrete has better compressive strength, develops minor cracks, For making geopolymer concrete, aluminosilicate materials, i.e.
and undergoes only slight damage in the mass at elevated temper- fly ash and ground granulated blast furnace slag (GGBS), coarse
atures as compared to Ordinary Portland Cement (OPC) concrete aggregate, manufactured sand, alkali activator solutions (sodium
[9,10]. It contains more amorphous phases, less porosity and more silicate and sodium hydroxide), and super-plasticizer were used.
pores in the mesopores range than concrete with OPC [11]. The For making conventional concrete, 53 grade cement, river sand,
previous research methods for determining the crack width and coarse aggregate and water were used. The chemical composition
depth used ultrasound, which is inconvenient to operate since it of the aluminosilicate materials and the physical properties of
requires a coupling agent [12]. Fibre optics approaches employed the concrete materials used for making the geopolymer concrete
to determine the crack width and location create complexity in are presented in Tables 1 and 2.
measurement [13]. Image-based crack detection methods have
been developed in the past 25 years and have been applied to test 3.1.2. BFRP, GFRP, steel bars and stirrups
civil structures such as pavements, bridges, and water retaining Steel bars of 12 mm, 10 mm and 8 mm diameter were used as
structures. Crack detection techniques used in research include reinforcement bars and stirrups, both, in the geopolymer concrete
image processing methods such as Canny, Fast Fourier Transform, and the conventional concrete. The properties of BFRP, GFRP and
Fast Haar Transform, and Sobel implemented by MATLAB [14], steel bars are presented in Table 3.
percolation-based approach [4,15,16], stereovision-based
approach [2], computational approach [17], image processing 3.1.3. Mix proportions and mechanical properties
methods such as morphological approach and integrated algorithm The mix proportion for M30 grade concrete used in this work
[18], and Principal Component Analysis (PCA) based algorithm was chosen from the previous work [25] as per IS 10262 [26].
[19]. Another crack detection technique used in concrete bridges The volume of binder, fine aggregate, coarse aggregate and liquid
is Spatially Tuned Robust Multifeature (STRUM) classifier-based used were 380 kg/m3, 660 kg/m3, 1189 kg/m3 and 171 kg/m3,
algorithm that uses on-site robotic scanning that delivers 90% respectively. In order to produce M30 grade geopolymer concrete,
accuracy [20]. FA and GGBS were taken in the ratio of 80:20. The ratio of liquid to
Wang and Huang [21] used Otsu threshold segmentation algo- binder was chosen as 0.45. In geopolymer concrete, river sand was
rithm integrated with a modified Sobel operator for crack detection fully replaced by M sand and one percent of naphthalene based

Table 1
Chemical composition of fly ash and GGBS.

Materials SiO2 CaO MgO Al2O3 Fe2O3 K2O SO4 Na2O LOI
Fly ash (%) 63.32 2.49 0.29 26.76 5.55 0.0002 0.36 0.0004 0.97
GGBS (%) 35.05 34.64 6.34 12.5 0.3 0.6 0.38 0.9 0.26

2
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

Table 2
Physical properties of the concrete materials.

Description Fly Ash GGBS Cement River Sand M Sand Coarse Aggregate
Specific gravity 2.13 2.85 3.13 2.66 2.72 2.73
Fineness modulus – – – 3.04 2.90 –
Water absorption (%) – – – 1.33 1.52 0.64

Table 3 The mechanical properties of the geopolymer and conventional

Properties of FRP bars and Steel bars. concretes are tabulated in Table 4 and were validated using the
Levenberg–Marquardt training algorithm using MATLAB software
Properties BFRP GFRP Steel
[28].
Elastic Modulus (GPa) 94 54 200
Tensile Strength (MPa) 513 515 495
Poisson’s ratio 0.23 0.24 0.27 3.2. Methods

3.2.1. Beam details and preparation

super plasticizer was used to achieve the workability of the Nine concrete beams having width, depth, and length of
geopolymer concrete. Geopolymer concrete is mainly made using 100 mm, 160 mm, and 1700 mm, respectively, were cast under
class F fly ash obtained from North Chennai Thermal power plant four point static bending test with an effective span of 1500 mm.
station, Chennai. Commercially available GGBS was used as part The beam specimen details are illustrated in Fig. 1. Three beams
replacement material for fly ash in this study. The mixture of each were cast in BFRP and GFRP bars in geopolymer concrete,
sodium silicate and sodium hydroxide solution was used as Alkali and steel bars in conventional concrete. Two 12 mm diameter bars
Activated Solution (AAS). The 8M NaOH flakes were mixed with at the bottom and two 10 mm diameter bars at the top, with 8 mm
distilled water a day before its use, to lower down the ambient diameter bars as stirrups were used in all the nine beams. The BFRP
temperature and then was mixed with Na2SiO3 solution to produce and GFRP stirrups were prepared by connecting the vertical and
alkaline activator solution to accelerate the reactivity of the solu- horizontal bars (8 mm diameter) using Anabond resin and the joint
tion. The sodium silicate solution with sodium hydroxide ratio by was connected with FRP mats using epoxy resin. The beams were
mass of 2.5 was used. In the mixing process, fine and coarse aggre- placed and loaded by varying the ratio of the shear span ‘a’ to
gates were first mixed in Saturated Surface Dry (SSD) condition in a the effective depth ‘d’ from 3.6 to 4.3. The beams were denoted
mixer machine. Then, the binder contents were mixed with the as follows: BRGC-3.6, BRGC-3.9, BRGC-4.3 for the BFRP bars,
aggregates. The mixing was continued for about 3 min. The already GRGC-3.6, GRGC-3.9, GRGC-4.3 for the GFRP bars, and SRCC-3.6,
prepared AAS was poured in the mixer machine and the mixing SRCC-3.9, SRCC-4.3 for the steel bars.
was continued for about 4 min. Finally, the super plasticizer was After mixing the aggregates in saturated surface dry condition
added to the mixtures to get the required workability until the followed by the addition of binders, alkali activated solution, and
concrete appeared homogenous and had the desired consistency super plasticizer to the mixer machine, the mix was poured in
[27]. three layers, and vibrated in beam moulds. The specimens were
demoulded after 24 h and kept under ambient curing for geopoly-
mer concrete and under water curing for conventional concrete for
28 days.
Table 4
Mechanical properties of geopolymer and conventional concretes.
3.2.2. Test setup and instrumentation
Mechanical properties Geopolymer concrete Conventional concrete
The details of the loading arrangements with varying ratios of
Compressive strength (MPa) 40.35 38.95
shear span to effective depth are shown in Fig. 2. Details of the test
Split tensile strength (MPa) 3.32 3.17
Flexural strength (MPa) 4.69 4.46 setup and instrumentation are shown in Fig. 3.
Modulus of elasticity (GPa) 19.10 22.19 Nine beams were tested under four point static bending test by
varying the ‘a/d’ ratio as 3.6, 3.9 and 4.3. The load was applied on

Load cell

2L - 8 mm Ø bars
Steel beam
at 100 mm c/c 2 Nos. 10 mm Ø

160 2 Nos. 12 mm Ø

LVDT 1 LVDT 2 LVDT 3 100

Cross section
L/3 L/3 L/3
Span, L = 1500 mm

Longitudinal section

Fig. 1. Reinforcement and Load setup details.

3
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

Shear Span P Spacing between loads P Shear Span

a = 500 = 500 a = 500

Beam Specimen with 'a/d' = 3.6

160

Pellets

LVDT 1 LVDT 2 LVDT 3

L/3 L/3 L/3

Span, L = 1500 mm

Shear Span Spacing between loads Shear Span

P P
a = 550 = 400 a = 550

Beam specimen with 'a/d' = 3.9

160

Pellets

LVDT 1 LVDT 2 LVDT 3

L/3 L/3 L/3

Span, L = 1500 mm

Spacing
between
Shear Span loads Shear Span
a = 600 P P a = 600
= 300

Beam specimen with 'a/d' = 4.3

160

Pellets

LVDT 1 LVDT 2 LVDT 3

L/3 L/3 L/3

Span, L = 1500 mm

Fig. 2. Details of loading arrangements with ‘a/d’ ratio of 3.6, 3.9, and 4.3.

the specimens and the cracks for the corresponding load intervals developed in the outer region of the Constant Bending Moment
at the time of testing were marked. The crack pattern until failure (CBM) zone (i.e. in the shear zone). But no shear cracks were devel-
was also noted. oped in SRCC-3.6. It was also evident that the mode of failure for
SRCC comprises of both flexure and compression with little shear
failure as the shear span to effective depth ratio was increased.
4. Crack analysis The crack patterns were similar in all the beams at initial load
intervals. But in BRGC and GRGC beams, it was found that inclined
From the experimental tests [29], the failure mode and crack cracks developed from flexural cracks. As compared to BRGC-3.6
pattern for all the nine beams subjected to static load by varying and BRGC-4.3, more inclined cracks developed in the shear zone
the shear span to the effective depth ratio were observed and are in BRGC-3.9. The same pattern was also observed in GRGC-3.9
shown in Fig. 4. when compared to GRGC-3.6 and GRGC-4.3. When the shear span
From Fig. 4 it was observed that when the ratio of the shear to the effective ratio was 4.3, the beam deflection re-cambered to
span to the effective depth in SRCC was increased, new cracks 20 mm at the ultimate load level after releasing the load for both
4
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

ficient shear reinforcement. The modes of failure and shear

strength at ultimate load levels for the steel and geopolymer
beams are shown in Table 5.
Table 5 indicates that the ultimate load-carrying capacity and
the shear capacity at ultimate load level in the steel and FRP rods
decreased as the ratio of the shear span to the effective depth
increased. Further, in both the FRP rods, the failure pattern chan-
ged from shear to flexure on increasing the ‘a/d’ ratio from 3.6 to
4.3 while no such change was seen in SRCC.
From the Table 6, it was observed that the number of cracks for
steel increased as the ratio of the shear span to the effective depth
increased to 3.9, and then decreased with further increase in the ‘a/
d’ ratio to 4.3. The same trend was observed for both the FRP bars.
In the case of steel and glass rods, the crack propagation decreased
Fig. 3. Details of test setup and Instrumentation. with increase in the shear span to the effective depth ratio while
for the basalt reinforced geopolymer concrete beams, the crack
propagation reached a high value with increase in this ratio. The
the beams. But for BRGC-3.6, BRGC-3.9, GRGC-3.6, and GRGC-3.9, spacing of crack decreased when the numbers of cracks and loads
sudden failure was observed after attaining 95% of the ultimate increased. The spacing of cracks was constant during the loading
load level. Thus it can be concluded that sudden shear and flexure and unloading of the ultimate load level for SRCC-4.3 and GRGC-
failure at premature has occurred. The sudden shear failure 4.3. From table 6, it was observed that the crack spacing decreased
observed in GRGC and BRGC beams can be attributed to the insuf- suddenly when the load level for SRCC-3.9 increased. When the

SRCC-3.6

SRCC-3.9

SRCC-4.3

(a). Crack pattern and failure mode of Steel reinforced conventional concrete.

BRGC-3.6

BRGC-3.9

BRGC-4.3

(b). Crack pattern and failure mode of Basalt reinforced geopolymer concrete.

GRGC-3.6

GRGC-3.9

GRGC-4.3

(c). Crack pattern and failure mode of Glass reinforced geopolymer concrete.

Fig. 4. (a). Crack pattern and failure mode of Steel reinforced conventional concrete. (b). Crack pattern and failure mode of Basalt reinforced geopolymer concrete. (c). Crack
pattern and failure mode of Glass reinforced geopolymer concrete.

5
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

Table 5
Modes of failure and shear strength at ultimate load levels.

Sl.No Specimen id Ultimate Load (kN) Shear Strength at Ultimate load (MPa) Failure mode
1 SRCC-3.6 49.80 3.56 Flexure & compression
2 SRCC-3.9 47.95 3.43 Flexure & compression
3 SRCC-4.3 40.20 2.87 Flexure & compression
4 BRGC-3.6 33.45 2.39 Shear compression
5 BRGC-3.9 32.55 2.33 Shear compression
6 BRGC-4.3 32.05 2.29 Flexure
7 GRGC-3.6 32.40 2.31 Shear compression
8 GRGC-3.9 26.65 1.90 Shear compression
9 GRGC-4.3 26.20 1.87 Flexure

Table 6
Behaviour of cracks under static load with varying ‘a/d’ ratio.

Specimen Id Total Number Total Number of Crack Crack Crack spacing – Crack spacing – Crack Width –First Crack Width –
of Cracks –First Cracks –Ultimate Propagation – Propagation – First (CBZ) Ultimate (CBZ) Ultimate
First Ultimate
(Nos.) (Nos.) (mm) (mm) (mm) (mm) (mm) (mm)
SRCC-3.6 5 15 48 116 115 71 0.093 0.4
SRCC-3.9 3 17 70 122 296 126 Nil 0.86
SRCC-4.3 3 11 71 101 97 97 0.355 1.32
BRGC-3.6 9 16 105 130 128 89 0.195 0.56
BRGC-3.9 4 18 85 159 164 75 Nil 1.21
BRGC-4.3 5 16 90 141 123 45 0.064 2.12
GRGC-3.6 6 13 135 153 197 73 0.235 0.78
GRGC-3.9 9 22 95 120 105 57 0.418 1.3
GRGC-4.3 3 11 70 117 135 135 Nil 3.8

ratio of the shear span to the effective depth was increased, the is known. Based on the experimental results, most of the data are
average crack width also increased. known and few data are unknown. Hence, only one machine learn-
The architecture for crack detection and pattern recognition ing was considered from the unsupervised learning. The above
based on image processing and machine learning algorithm in mentioned machine learning applications were used to study var-
python script is depicted in Fig. 5. ious patterns and determine the accuracy of crack detection and
Machine Learning method is broadly classified into supervised pattern recognition.
and unsupervised learning. Supervised learning includes Logistic
Regression, Naïve Bayes, Stochastic Gradient Descent, K-Nearest 4.1. Crack image gathering
Neighbors, Decision Tree, Random Forest, Support Vector Classifier
and Deep Learning Neural Network while unsupervised learning The images from the structure were captured using a camera
comprises of K-Means Clustering, Gaussian Mixture Model and and were then subjected to the image pre-processing process and
Spectral Clustering. Adaboost is used to combine multiple classi- further analysis.
fiers. In the present work Stochastic Gradient Descent, K-Nearest
Neighbors, Decision Tree and Support Vector Classifier were 4.2. Image Pre-processing
selected from the supervised learning and Gaussian Mixture Model
was chosen from the unsupervised learning. Adaboost was used to Image pre-processing comprises of gray scale image conversion,
combine multiple classifiers. The reason for selecting four machine image resizing, canny edge detection, and feature extraction.
learning algorithms from the supervised learning is that the train-
ing set of data contains observations whose category membership 4.2.1. Gray scale image conversion
A colour image contains lots of data, most of which are not
required for crack detection. Colour images contain RGB (Red,
Crack Detection Green, and Blue) colours with different intensity labels, which
Initialization Machine Learning
Techniques
increase the image size as well as consume more time for process-
Experimental Image Pre-processing Pattern ing. Hence conversion of the colour image to gray scale image is
Study Collections necessary to select relevant information, discard unnecessary
Gray Scale Image Training Testing
information and reduce the processing time. In this work, grayscale
Patterns Patterns image conversion was carried out on the input image to convert it
Crack Image
Image Resizing
from colour to a grayscale image using python package ‘skimage.-
Gathering
Different Classifiers color’. The obtained grayscale image contained a range of black to
white shades. Fig. 6(a) depicts the original colour image of the
Canny Edge
Detector crack with its intensity and Fig. 6(b) depicts the corresponding gray
Accuracy
Precision
scale image with its intensity. The grayscale image thus obtained
Feature Recall was used for image resizing.
Extraction Confusion

4.2.2. Image resizing

Crack Detection Pattern Recognition
The grayscale image was resized to the standard size using
python package ‘skimage.transform’ to avoid unnecessary content
Fig. 5. Architecture for crack detection and pattern recognition. in the image and make it ready to identify the cracks. The gray
6
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

(a)Original colour image and its intensity (b)Gray scale image and its intensity.
Fig. 6. Image and intensity details.

(a)Gray scale image (b)Resized gray scale image

Fig. 7. Image scale details.

scale image and the resized gray scale image are shown in Fig. 7(a) 4.3. Machine learning
and (b). This resized gray scale image was used for further
processing. Machine learning consists of pattern collection and training,
and testing and classification. In this work the machine learning
methods were implemented using python package ‘sklearn’.
4.2.3. Canny edge detector
Canny Edge detector is very popular method to separate the
background information from the images and to identify the edges 4.3.1. Pattern collection and training
as cracks. Canny Edge detector method was implemented using In this stage, the patterns were collected by employing pre-
python package ‘skimage.feature’. processing techniques and the collected patterns were used to
train the classifiers and evaluate the models. The failure patterns
of few images is shown in Fig. 10.
4.2.4. Work flow of canny edge detector
The Canny edge detector uses the following steps: 4.3.2. Training and testing patterns using different classifiers
To train the patterns and predict the failure pattern in the
i. The image was smoothened using Gaussian filter to elimi- images, Support vector machine (SVM), Decision tree classifier,
nate noise. Gaussian NB classifier, SGD classifier, K- neighbor classifier and
ii. The intensity gradients of the image were determined. Adaboost classifiers were used. Three classes of pattern failures
iii. The false positives in edge detection were eliminated by car- were considered: flexure, shear, and compression failure, which
rying out Non-maximum suppressing. were treated as target values and subjected to processing by each
iv. The potential boundaries were determined by employing machine learning classifier.
double threshold. Of the pattern images, 50% were trained and the remaining 50%
v. Edge tracking was carried out using hysteresis in which edge were tested. In addition, the performance of each classifier was
detection was done by eliminating all weak edges and evaluated using confusion matrix and accuracy score. The evalua-
unconnected to strong edges. tion results showed that SVM exhibited the highest accuracy when
compared to other classifiers.
Fig. 8(a) shows the resized gray scale image before applying the
canny edge detector function. Fig. 8(b) shows the noise reduced
4.3.3. Support vector Machine
using canny edge detector function with r = 0.6 and Fig. 8(c) shows
SVM is a machine learning algorithm used for classifying the
the noise further reduced using canny edge detector function with
different types of cracks and training the machine to identify the
r = 1.6. It is clear that the noise is reduced fully when the canny
correct type of crack pattern failure.
edge detector function with parameter r = 1.6 was applied. Hence,
Input parameters used for SVM are as follows:
r plays a key role in the noise reduction in images while using
canny edge detector.
i. Kernel: Linear
ii. Gamma: 0.0001
4.2.5. Percentile method iii. Test size: 0.4 (40%)
Percentile method was used to eliminate outliers in the images.
It was implemented using python package ‘numpy’. In the present In general, though SVM is considered as a classification method,
work, the height and width percentile were considered to detect it can also be used for regression and classification problems. SVM
the right crack pattern collection from the image. The resulting offers the advantage of simplicity in handling multiple categorical
patterns are called feature extractions and are shown in Fig. 9. and continuous variables. In SVM, a hyperplane is constructed in
7
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

(a). Resized gray scale image before applying canny edge detector function.

(b). Noise reduced using =0.6 in canny edge detector function.

(c). Noise reduced using =1.6 in canny edge detector function.

Fig. 8. (a). Resized gray scale image before applying canny edge detector function. (b). Noise reduced using r = 0.6 in canny edge detector function. (c). Noise reduced using
r = 1.6 in canny edge detector function.

Fig. 9. Feature Extraction of images.

Fig. 10. Failure pattern of images.

multidimensional space for separating various classes. An optimal Support vectors are the data points that are closest to the
hyperplane is generated iteratively and is utilized for minimizing hyperplane. These points define the separating line effectively by
the errors. A Maximum Marginal Hyperplane (MMH) is located calculating the margins and are more relevant to the construction
to perform the division of the dataset as classes in the best manner. of the classifier. Hyperplane is a decision plane which separates a
Fig. 11 is a representation of SVM showing the support vectors, set of objects having different class memberships. Margin is the
margins, and hyperplanes. gap between the two lines on the closest class points, which is cal-

8
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

5.2. Evaluating the model

The accuracy of the classifier or model for predicting the type of

failures in the crack was estimated. Accuracy was computed by
comparing the actual test set values and predicted values using
confusion matrix. The accuracies of the different classifiers for flex-
ure, shear and compression pattern are shown in Table 7.
From Table 7, it was observed that the accuracy, precision, and
recall values for all the pattern classes are the best for the support
vector classifier when compared with the other classifiers. The
confusion matrix is a Table which describes the performance of
the classifiers using the test data. The confusion matrix for SVM,
Decision tree classifier, Gaussian NB classifier, SGD classifier, K-
neighbor classifier and Adaboost classifier is depicted in Fig. 14
as normalized values.
Fig. 11. Representation of SVM. Fig. 14 indicates that the confusion matrix for SVM shows very
good performance in all the three classes (compression, flexure,
and shear). The performance and accuracy of SVM is the best
among all the classifiers.
culated as the perpendicular distance from the line to the support Decision tree classifier can cause a big variation and instability
vectors or the closest points. If the margin between the classes is even with small data variations. Calculations in Decision tree clas-
larger it is considered good while a smaller margin is considered sifier get more complex than in other algorithms with the model
bad. training time also being longer. The problem with Gaussian NB
classifier is in its implicit assumption that all attributes are mutu-
4.3.4. SVM workflow ally independent, which is impossible in real life. Owing to fre-
The main objective is to segregate the given dataset in the best quent updates, steps taken by SGD classifier to reach the minima
possible way. The distance between the nearest points is known as are most noisy, often leading to falling gradient towards other
the margin. The objective is to select a hyperplane with the maxi- directions. The noisy steps make the loss function minima conver-
mum possible margin between the support vectors in the given gence longer. The K- neighbor classifier, on the other hand, does
dataset. Fig. 12 explains the work flow of SVM when it searches not perform well when datasets are large. It also performs poorly
for the MMH. with high dimensions, since it is difficult for the classifier to calcu-
It is seen in Fig. 12 that the generated hyperplanes segregate the late each dimension’s distance and hence works poorly with more
classes in the best possible way. Four hyperplanes are seen: black, noise in the dataset. Adaboost too is sensitive to outliers and noise.
blue, orange and purple. The blue and orange hyperplanes have If the data is imbalanced, the classifying accuracy is affected.
higher classification error while the black and purple hyperplanes There have been a few studies on crack analysis using SVM. Pra-
separate the three classes correctly. sanna et al. [30] proposed a crack detection technique for concrete
bridges. They compared the results of SVM, Random Forest, and
Adaboost classifiers and found that Adaboost was the most accu-
5. Results and discussions rate with an accuracy of 90.8%. However, the other two classifiers
also performed well. Li et al. [31] studied cracks in bridges using
5.1. Failure pattern identification greedy-SVM for feature selection. They obtained a crack analysis
accuracy of 93.6%. Kim et al. [32] proposed a speeded-up robust
The beam images were taken from the experimentally tested features model for crack detection in concrete. Classification was
beams for training and testing of the models. A total of 232 pat- done by linear SVM. The model outperformed the other models
terns were taken for evaluating the model. Of the 232 patterns, by yielding an accuracy of 98%. Another model was proposed by
139 patterns were used for training, and the remaining 93 patterns Liang et al. [33] for concrete crack analysis that used the OTSU
were utilized for testing (Fig. 13). algorithm for segmentation and SVM for classification. The model
gave an accuracy of 100% in the test set identification. Zhang
et al. [34] in their proposal, modeled an automatic crack detection
and classification method for subway tunnel safety monitoring. For
feature extraction, a distance histogram-based shape descriptor
was used while different classifiers were used for classification.
The Extreme Learning Machine classifier worked best with an
accuracy of 91.6%, while the SVM also performed similarly well
(accuracy 91.3%). Xie et al. [35] studied voids within RCC structures
that impact the structure’s safety using GPR images. Their auto-
mated model was based on a novel SVM algorithm. Accuracy of
up to 97.7% for void identification was obtained using this model.
From the present study it is evident that the SVM is generally
more accurate when compared to other classifiers. SVM can easily
manage complex non-linear data points and over-fitting issues are
lesser than that in other classifiers. SVM also shows better perfor-
mance in high dimension spaces and in situations where the num-
ber of dimensions exceeds the number of samples.
Experimental analysis and machine learning model for predict-
Fig. 12. Work flow of SVM. ing the crack detection and pattern recognition for the reinforced
9
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

(a). Patterns for model evaluation of beam with Steel bars (SRCC-3.9).

(b). Patterns for model evaluation of beam with Basalt bars (BRGC-3.9).

(c). Patterns for model evaluation of beam with Glass bars (GRGC-3.9 ).
Fig. 13. (a). Patterns for model evaluation of beam with Steel bars (SRCC-3.9). (b). Patterns for model evaluation of beam with Basalt bars (BRGC-3.9). (c). Patterns for model
evaluation of beam with Glass bars (GRGC-3.9).

10
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

Table 7
The accuracies of different classifiers for flexure, shear, and compression pattern.

Classifiers/Patterns Flexure pattern Shear pattern Compression pattern

Accuracy (%) Precision Recall Accuracy (%) Precision Recall Accuracy (%) Precision Recall
Support vector 100 1.00 1.00 100 1.00 1.00 100 1.00 1.00
Decision tree 97 1.00 0.94 97 0.93 1.00 88 0.79 1.00
Gaussian NB 99 0.99 1.00 90 1.00 0.82 100 1.00 1.00
SGD 98 1.00 0.96 90 0.82 1.00 100 1.00 1.00
K- neighbor 98 1.00 0.96 100 1.00 1.00 88 0.79 1.00
Adaboost 99 0.99 1.00 96 0.93 1.00 100 1.00 1.00

(a) SVM (b) Decision tree classifier

(c) Gaussian NB classifier (d) SGD classifier

(e) K- neighbor classifier (f) Adaboost classifier

Fig. 14. Confusion matrix for different classifiers.

concrete without cement and steel reinforcement is the novelty of lished to date for the geopolymer concrete reinforced with BFRP/
the present research work. The cement concrete and steel rein- GFRP bars. Machine Learning technique is the only option to pre-
forcement were replaced by geopolymer concrete and BFRP/ GFRP dict the crack detection for the geopolymer concrete and BFRP/
bars, respectively and were used as longitudinal reinforcements for GFRP bars combination. Coefficients of thermal expansion for con-
casting and testing. For reinforced concrete, lots of codes of prac- ventional cement concrete and steel bars are almost the same. In
tice/ standards are available for the shear, flexure, deflection and contrast, the coefficient of thermal expansion of fly ash geopolymer
crack width analysis. On the other hand, no code has been pub- concrete is in the range of 5  10 6/°C to 13  10 6/°C. The specific

11
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

properties of geopolymer concrete depend on the source/ raw online in Wiley Inter Science (www.interscience.wiley.com).
DOI:10.1002/tee.20244.
materials used for the specimen preparation. Pavalan V. and
[2] B. Shan, S. Zheng, O.u. Jinping, A stereovision-based crack width detection
Sivagamasundari R have conducted experimental study to deter- approach for concrete surface assessment, KSCE J. Civ. Eng. 20 (2) (2016) 803–
mine the thermal expansion coefficient of various FRP bars. Based 812.
on the studies, the coefficient of longitudinal thermal expansion of [3] H. Nakamura, R. Sato. K. Kawamura, A. Miyamoto. Proposal of a crack pattern
extraction method from digital images using an interactive genetic algorithm.
BFRP and GFRP bars are around 2  10 6/°C and 9  10 6/°C, Japan Society of Civil Engineers, 60 (742) 115–131, September 2003 (in
respectively [36]. Owing to the variation in the coefficient of ther- Japanese).
mal expansion, machine learning technique was adopted for the [4] T. Yamaguchi, K. Suzuki, P. Hartono, S. Hashimoto, Percolation approach to
image-based crack detection, in: Proceedings of the 7th international
crack identification and damage pattern recognition of geopolymer conference on Quality Control by Artificial Vision, 2005, pp. 291–296.
concrete reinforced with BFRP/ GFRP bars. [5] Amer Hassan, Mohammed Arif and M. Shariq, Effect of Curing Condition on the
Mechanical Properties of Fly Ash based Geopolymer Concrete, SN Applied
Sciences, Springer Nature, 1:1694, 2019. https://doi.org/10.1007/s42452-019-
6. Conclusions 1774-8.
[6] Amer Hassan, Mohammed Arif, M. Shariq, Use of geopolymer concrete for a
cleaner and sustainable environment - A review of mechanical properties and
If pre-processing techniques and pattern classifier techniques microstructure, J. Clean. Prod. Elsevier, 223, (2019) 704-728. https://doi.org/
are robust, then high accuracy can be obtained for pattern recogni- 10.1016/j.jclepro.2019.03.051.
tion very easily. In this research, pre-processing techniques such as [7] Amer Hassan, Mohammed Arif, M. Shariq, Mechanical behaviour and
microstructural investigation of geopolymer concrete after exposure to
Gray-scale image conversion, Image resizing, Canny edge detector
elevated temperatures, Arabian J. Sci. Eng., Springer, 45, (2020), 3843–3861.
and Percentile method were used and validated for the extracted https://doi.org/10.1007/s13369-019-04269-9.
crack patterns. [8] A. Hassan, M. Arif, M. Shariq, A review of properties and behaviour of
reinforced geopolymer concrete structural elements - A clean technology
Different classifiers namely, SVM, Decision tree classifier, Gaus-
option for sustainable development, J. Clean. Prod. 245 (2020), https://doi.org/
sian NB classifier, SGD classifier, K-neighbour classifier, and Ada- 10.1016/j.jclepro.2019.118762 118762.
boost classifier were used to segregate the compression, flexure, [9] A. Hassan, M. Arif, M. Shariq, Age-dependent compressive strength and elastic
and shear classes of patterns from the given crack patterns. Among modulus of fly ash-based geopolymer concrete, structural concrete, J. Fib.
(2020), https://doi.org/10.1002/suco.202000372.
these classifiers, SVM recorded the best performance and accuracy [10] A. Hassan, M. Arif, M. Shariq, Structural performance of ambient cured
for prediction of crack failures across all three classes, whereas reinforced geopolymer concrete beams with steel fibres, structural concrete, J.
other classifiers performed well only for one or two classes. Fib. (2020), https://doi.org/10.1002/suco.202000191.
[11] Influence of Microstructure of Geopolymer Concrete on its Mechanical
The present research work recommends that the geopolymer Properties — A Review 35 (2020), https://doi.org/10.1007/978-981-13-7480-
reinforced concrete prepared with fly ash, GGBS, manufactured 7_10.
sand and GFRP/ BFRP bars can be used as alternative/ sustainable [12] D. McCann, M. Forde, Review of NDT methods in the assessment of concrete
and masonry structures, NDT & E Int. 34 (2) (2001) 71–84, https://doi.org/
materials for cement, fine aggregates and steel bars, respectively. 10.1016/S0963-8695(00)00032-3.
Flexural members of concrete structures prepared with GFRP/ BFRP [13] W.T. Zhang, J.Y. Dai, H. Xu, B.C. Sun, Y.L. Du. Distributed fiber optic crack sensor
bars and a geopolymer concrete combination is supportive to avoid for concrete structures.” Proc. of SPIE, Sandiego, USA, 6830 (2007) 68300F,
DOI: 10.1117/12.757864.
reinforcement corrosion and reduce air pollution. In the present
[14] Ikhlas Abdel-Qader, Osama Abudayyeh, Michael E. Kelly, Analysis of edge
study, various machine learning methods were adopted to classify detection techniques for crack identification in bridges, J. Comput. Civ. Eng.
several types of cracks in geopolymer concrete beams. The crack Am. Soc. Civ. Eng. 17 (3) (2003) 255–263.
[15] T. Yamaguchi, S. Hashimoto, Automated Crack Detection for Concrete Surface
analysis proved that staged image processing techniques and
Image Using Percolation Model and Edge Information, 1-4244-0136-4/06
machine learning classifiers used in this research work are greatly (2006) IEEE.
advantageous in the detection of tensile, shear, and compression [16] T. Yamaguchi, K. Suzuki, P. Hartono, S. Hashimoto. An efficient crack detection
cracks in reinforced concrete beams. method using percolation-based image processing, Proc. of ICIEA2008, 3
(1875-1880) (2008).
[17] John Canny, A Computational Approach to Edge Detection, IEEE Transactions
CRediT authorship contribution statement on Pattern Analysis and Machine Intelligence, PAMI-8 (6) NOVEMBER 1986.
[18] H. Wang Pand Huang, Comparison Analysis on Present Image-based Crack
Detection Methods in Concrete Structures, 3rdInternational Congress on Image
N. Aravind: Validation, Investigation. S. Nagajothi: Writing - and Signal Processing, 2010.
original draft, Visualization, Methodology, Conceptualization, [19] I. Abdel-Qader, S. Pashaie-Rad, O. Abudayyeh, S. Yehia, PCA-Based algorithm
for unsupervised bridge crack detection, Adv. Eng. Softw. 37 (2006).
Investigation, Resources, Validation, Software, Data curation. [20] Prateek Prasanna, Kristin J. Dana, Nenad Gucunski, Basily B. Basily, Hung M. La,
S. Elavenil: Formal analysis, Supervision, Investigation, Validation, Ronny Salim Lim, Hooman Parvardeh, IEEE Trans. Automat. Sci. Eng. 13 (2)
Writing-review & editing. (2016) 591–599.
[21] Yun Wang, Ju Yong Zhang, Jing Xin Liu, Yin Zhang, Zhi Ping Chen, Chun Guang
Li, Kai He, Rui Bin Yan, Research on Crack Detection Algorithm of the Concrete
Declaration of Competing Interest Bridge Based on Image Processing, 8thInternational Congress of Information
and Communication Technology, ICICT 2019, Procedia Computer Science 154
(2019) 610–616.
The authors declare that they have no known competing finan- [22] Cao Vu Dunga, Le DucAnh, Autonomous concrete crack detection using deep
cial interests or personal relationships that could have appeared fully convolutional neural network, Automat. Constr. 99 (2019) 52–58.
to influence the work reported in this paper. [23] Yusuke Fujita, Yoshihiro Mitani, Yoshihiko Hamamoto, A Method for Crack
Detection on a Concrete Structure, 18thInternational Conference on Pattern
Recognition (ICPR’06) 0-7695-2521-0/06, 2006.
Acknowledgements [24] S. Nagajothi, S. Elavenil, Parametric studies on the workability and
compressive strength properties of geopolymer concrete, J. Mech. Behav.
Mater. 20180019 (2018) 1–11.
Authors would like to acknowledge the Management and Dean- [25] S. Nagajothi, S. Elavenil, GGBFS & M-Sand impact on workability and strength
School of Civil Engineering, Vellore Institute of Technology, Chen- properties of fly ash based geopolymer concrete, Indian J. Eng. Mater. Sci. 27
nai, India for providing the necessary support to carry out this (2020) 67–76.
[26] IS 10262, Indian Standard recommended guidelines for concrete mix design.
research.
Bureau of Indian Standards, New Delhi 2009.
[27] S. Nagajothi, S. Elavenil, Experimental investigations on compressive, impact
References and prediction of stress-strain of fly ash-geopolymer and portland cement
concrete, J. Polym. Eng. 40 (7) (2020) 583–590.
[28] S. Nagajothi, S. Elavenil, Influence of aluminosilicate for the prediction of
[1] Tomoyuki Yamaguchi, Shingo Nakamura, RyoSaegusa and Shuji Hashimoto,
mechanical properties of geopolymer concrete – Artificial Neural Network,
Image-Based Crack Detection for Real Concrete Surfaces, Transactions on
Silicon 12 (2020) 1011–1021.
Electrical and Electronic Engineering, IEEJ Trans 2008; 3: 128–135, Published

12
N Aravind, S Nagajothi and S Elavenil Construction and Building Materials 297 (2021) 123785

[29] S. Nagajothi, S. Elavenil, Shear Prediction of geopolymer concrete beams using [33] S. Liang, X. Jianchun, Z. Xun, An algorithm for concrete crack extraction and
Basalt/Glass FRP bars, J. Adv. Concr. Technol. 19 (2021) 216–225. identification based on machine vision, IEEE Access 6 (2018) 28993–29002.
[30] P. Prasanna, K.J. Dana, N. Gucunski, B.B. Basily, H.M. La, R.S. Lim, H. Parvardeh. [34] W. Zhang, Z. Zhang, D. Qi, Y. Liu. Automatic crack detection and classification
Automated crack detection on concrete bridges, IEEE Trans. Automat. Sci. Eng. method for subway tunnel safety monitoring. Sensors, 14(10) (2014) 19307-
13(2) (2014) 591-599. 19328.
[31] G. Li, X. Zhao, K. Du, F. Ru, Y. Zhang, Recognition and evaluation of bridge [35] X. Xie, P. Li, H. Qin, L. Liu, D.C. Nobes, GPR identification of voids inside
cracks with modified active contour model and greedy search-based support concrete based on the support vector machine algorithm, J. Geophys. Eng. 10
vector machine, Automat. Constr. 78 (2017) 51–61. (3) (2013) 034002.
[32] H. Kim, E. Ahn, M. Shin, S.H. Sim, Crack and non-crack classification from [36] V. Pavalan, R. Sivagamasundari, Thermal expansion coefficient of basalt fibre
concrete surface images using machine learning, Struct. Health Monit. 18 (3) reinforced polymer bars, Int. J. Res. Eng. Appl. Manag. 5 (1) (2019) 414–418,
(2019) 725–738. https://doi.org/10.18231/2454-9150.2019.0333.

13