Measurement of Reinforcement Corrosion in Concrete Adopting

Ultrasonic Tests and Artificial Neural Network

Yidong Xua, Ruoyu Jinb,*

a
Ningbo Institute of Technology, Zhejiang University, Ningbo 315100, China;
b
School of Environment and Technology, University of Brighton, Cockcroft Building 616, Lewes Road,
Brighton, U.K., BN2 4GJ

Abstract

Limited research has been performed in testing and measuring the reinforcement corrosion

levels using non-destructive tests. This research applied ultrasonic-based non-destructive test

and artificial neural network to the diagnosis and prediction of rebar’s non-uniform

corrosion-induced damage within reinforced concrete members. Ultrasonic velocities were

tested by applying ultrasonic to reinforced concrete prisms before and after the rebar

corrosion. Input parameters including concrete strength, ultrasonic velocity, and the specimen

dimension-related variable were used for the prediction of reinforcement corrosion level

adopting artificial neural network models. Using totally 50 experimental observations, Radial

Basis Function-based model was found with higher accuracy in predicting corrosion levels

compared to Back Propagation-based model. This study leads to future research in

high-accuracy non-destructive measurement of reinforcement corrosion in concrete.

Keywords: Reinforcement corrosion; reinforced concrete; non-destructive test; artificial

neural network

1. Introduction

Concrete structural damage due to reinforcement corrosion is not uncommon and could be

significant [1]. For example, chloride-induced durability problems shorten the service life of

concrete structures [2]. The durability of concrete materials under chloride penetration
requires particular attention [3]. However, reinforcement corrosion is not easily noticed. How

to identify and measure the rebar corrosion within reinforced concrete structure remains a key

issue in the concrete durability research. The commonly adopted non-destructive test

approaches for the measurement include half-cell potential method [4] and concrete

resistivity method (e.g, Wenner Four-Electrode Method [5]). These existing methods have

their own advantages of being easy to operate for on-site tests with simple detection facility.

However, both methods can only be used for the probability analysis of rebar corrosion

within concrete but are unable to quantify the corrosion level or the corrosion-induced

damage, especially for the non-uniform corrosion-induced damage within reinforced concrete.

A more effective approach to test the rebar corrosion level would be to separate the rebar

from concrete and measure the shape and depth of rebar etch-pits. Nevertheless, the

separation of rebar from concrete would cause further structural damage and hence being not

feasible. Therefore, it cannot be widely applied. There is hence an urgent need of developing

a more applicable and practical measurement technology for the non-destructive test of rebar

corrosion within concrete structures.

Ultrasonic measurement can be used to test concrete mechanical properties and internal

defects through the elastic wave signals sent to concrete structures. The ultrasonic test

implemented in concrete structure is based on the artificial method to generate elastic wave

signals, which contains the information of propagation time (or speed), amplitude and

frequency. The concrete structure is elastic-viscoplastic, with variations of acoustic

impedance among its internal interfaces. The ultrasonic waves have strong reflection,

scattering, diffraction, absorption, and waveform distortion when passing through the

concrete structure. By extracting the information carried out in the signals and performing

inversion analysis, the data of mechanical properties and defect distribution of materials and

structures can be obtained.

 Huygens-Fresnel Principle [6] can be applied in supporting the ultrasonic measurement

by comparing the defect size and the wave length, as the wave path would be different

depending on the defect size within concrete structures. The predication of concrete strength

by integrating ultrasonic wave and artificial neural network (ANN) has been performed in

some existing studies such as Trtnik et al. [7] and Kewalramani and Gupta [8]. However,

there has not been sufficient research on applying ultrasonic wave in measuring the rebar

corrosion within concrete structures. ANN, as one of the data mining methods that have been

widely applied in estimating concrete mechanical properties (e.g., strength) in existing studies

such as Chithra et al. [9], Omran et al. [10], Sadowski et al. [11], Wang et al. [12], Wang et al.

[13], has not been sufficiently used in the prediction of corrosion level of reinforcement in

concrete.

Non-destructive tests have been applied in the test and evaluation of properties of

construction materials, such as bond between concrete layers [11], compressive strength [12], and

other types of mechanical properties [15]. However, limited studies have been performed in

applying non-destructive tests to measure reinforcement corrosion level in concrete members in a

quantitative approach. Yeih and Huang [16] adopted ultrasonic testing for the non-destructive

detection of concrete reinforcement corrosion, and found certain correlation between the

reduction of ultrasonic amplitude and the electrochemical parameters of corrosion. Continued

from the research of Yeih and Huang [16], this study aims to apply the ultrasonic test in

measuring the corrosion level of non-uniform surfaces of rebar, and to evaluate the accuracy of

ultrasonic test-based measurement on reinforcement corrosion level. Radial Basis Function

(RBF)-based and Back Propagation (BP)-based models are applied and to be compared of

their accuracies in the prediction. The study serves as the initial exploration of non-destructive

tests of rebar corrosion level in non-uniform surfaces, leading to further research in establishing
the non-destructive test procedure in measuring steel corrosion within reinforced concrete

structures.

2. Experimental program

2.1. Linking reinforcement corrosion level to ultrasonic velocity

The corrosion process of reinforced concrete was defined by Zhao and Jin [17] in terms

of three stages, namely: 1) the first stage of free expansion of rust, in which the steel

corrosion products fill the rebar/concrete interface voids; 2) tensile stress within concrete

cover, in which the concrete cover starts cracking; and 3) cracking within concrete cover,

when the cracking continues to be filled with the rust products. Empirical formulas have been

established to estimate the rebar corrosion level in concrete structures. Zhao and Jin [17, 18]

further proposed empirical formulas to estimate the corrosion depth of reinforcement. Zhao et

al [19] used stereoscopic microscope to observe the seriously corroded layer and found that

the corrosion depth was not more than 500 μm. Following Huygens-Fresnel Principle [3],

when the corrosion layer size is smaller than the wavelength, the ultrasound would diffract,

resulting in a longer receiver signal. In addition, since the acoustic impedance of the air is far

lower than that of concrete, sound energy would decrease and the sound duration would

increase resulting from the reflecting and scattering of pulse wave at the corrosion surface.

Therefore, it is feasible to detect the corrosion level of reinforcement in concrete according to

the ultrasonic velocity. The effect of test specimen size described in Fig. 1 in ultrasonic

velocity was studied by Zhang and Qiu [20].

 (a) Before corrosion (b) After corrosion
Fig.1. Ultrasonic propagation path in reinforced concrete

As illustrated in Fig.1, R denotes the rebar radius, δ represents the depth of steel

corrosion layer, and L is the distance between the emission transducer and receiving

transducer within the ultrasonic generator. It was previously found by Zhang and Qiu [20]

that under the same ultrasonic velocity in concrete, the effect of rebar in ultrasonic velocity

would be reduced as the ratio of rebar to the test distance (denoted as 2R/L) decreases. When

this ratios is lower then 1/12, the effect of rebar size could be ignored according to Zhang and

Qiu [20].

2.2. Experimental set-up

Four different types of reinforced concrete specimens, namely C20, C25, C30, and C35, were

prepared by Gutetong Company from Ningbo China. The mix design details of these four

types of specimens are provided in Table 1.

Table 1. Mix design details of four different concrete specimens in this study
Granulated
Portland Coarse Water
Specimen Fly ash blast furnace Sand Water Slump
Cement aggregate reducer
Type (kg/m3) slag (kg/m3) (kg/m3) (mm)
(kg/m3) (kg/m3) (kg/m3)
(kg/m3)
C20 182 112 26 757 1045 195 3.52 150
C25 202 87 47 762 1052 185 4.04 150
C30 220 83 58 736 1059 180 4.50 150
C35 245 78 66 695 1086 175 5.06 150
 Totally 50 groups of concrete specimens were cast, among them 44 were prisms sized at

100mm×100mm×400mm, and six of them being prisms sized at 150mm×150mm×550mm,

each prism was embedded with a hot-rolled round steel rebar with the diameter of 12mm.

Cured at the standard laboratory condition for 28 days, specimens were placed in the chloride

salt and stray current condition to reach different levels of rebar corrosion. Specifically, two

types of corrosion tests were adopted in this study, namely chlorine salt wet-dry cycle and

stray current method. In the chlorine salt wet-dry cycle approach, specimens after 28 days’

standard curing were soaked and placed in 5% sodium chloride solution for 12 hours, and

then oven dried in the temperature of 50±1 °C for another 12 hours. This was counted as one

wet-dry cycle. The same cycles were repeated to corrode reinforcement in the chloride salt

environment; in the stray current condition, concrete specimens after the same 28 days’

standard curing were buried in the 5%-sodium-chloride-wetted sand. Rebar was then used as

the anode with copper bars as cathodes. They were electrified with a DC power supply to

create rebar corrosion in the stray current conditions.

The NM-4A type of non-metallic instrument was adopted in this study to test the

ultrasonic velocity within the concrete specimen. To prevent the ultrasonic velocity being

affecting by the variation of specimens’ moisture contents, each specimen was stored indoor

under constant room temperature and dried for three days. Each specimen’s longitudinal

surface was set up with nine test areas. The average ultrasonic velocity was calculated from

the nine test areas. Test details, including specimen sections and the nine test zones are

described in Fig. 2. The test procedure is illustrated in Fig.3.

Fig.2. Reinforced concrete specimen for ultrasonic tests with details of test specimens (Size

units /mm)
 (a) Concrete curing (b) Half-cell test (c) Ultrasonic test for prisms sized at
100mm×100mm×400mm

(d) Ultrasonic test for prisms sized (e) NM-4A type of non-metallic (f) Specimens after testing
at 150mm×150mm×550mm ultrasonic instrument

Fig.3. Schematic illustration of ultrasonic testing

Before the ultrasonic test had been performed in concrete specimen, the effect of

concrete moisture content on ultrasonic velocity was examined. Mix design targeting on C30

ordinary concrete was adopted to cast two standard prisms, with the dimension at

100mm×100mm×400mm. Cured at the standard condition on Day 28, the reinforced prism

was connected to electrical circuits to achieve different degrees of corrosion severity. Later,

following the procedure described by Liu et al. [21], the corroded specimens were immersed
in water for three days to reach full saturation. The saturated specimens were then removed

from water, surface dried, and weighted. The specimens were then put in ovens to be

gradually dried with lower moisture content. After each drying period, the specimen was

removed from the oven, weighted, and retested of the ultrasonic velocity until it reached its

constant weight when fully internally dried. The test setup is illustrated in Fig.4.

(a) Specimens being dried (b) Schematic illustration of ultrasonic testing

Fig.4. Test setup to studying the influence of moisture content on ultrasonic velocity

Data obtained from the test setup described in Fig.4 would then allow the correlation

analysis between moisture content and the average ultrasonic velocity.

2.3. Application of artificial neural network to predicting rebar corrosion

Artificial neural network (ANN) was applied in this research to analyzing the rebar

corrosion, based on the three following features of ANN [22]: 1) it has strong

self-adaptability and is able to capture the corrosion-caused damage in concrete members; 2)

it is able to store the damage-related information in the learning stage and compare the

predicted data with the measured outcomes; 3) it is robust with its non-linear mapping

capacity to establish the relationship between the rust damage related value and the

non-destructive parameters. In this study, ANN was modeled using the NeuroSolutions [23]

as the tool.
 The Back Propagation (BP) and Radial Basis Function (RBF) were the two ANN models

that were adopted in this research to link the non-destructive parameters to corrosion levels of

rebar in concrete prisms.

BP has been widely applied in the field of civil engineering [24-27]. The topology

structure of BP-based network can be described in Fig.5.

Fig.5. topology structure of BP neural network

As displayed in Fig.5, the data samples start from the input layers, passing the hidden

layer until they reach the output layer. In Fig.4, ij and ij represent the connection

weightings when data is moving forward from the input layer to hidden layer, and from the

hidden layer to the output layer respectively. During the learning and training process, the

error would be passed backwards through hidden layer to input layer if the actual output

value does not match the expected value. Then the error is spreading to all cells in the input

layer illustrated in Fig.5. The error signal serves as a basis for modifying the weighing of
each cell. The weightings will be adjusted through a new cycle by passing the data forward

from input to output layers. This kind of cycle can be reprocessed until the error is reduced to

a tolerable level, or the learning process has completed certain specified number of cycles.

Detailed process of running BP-based models can be found in Buscema [28] and

Hecht-Nielsen [29].

RBF’s topology structure can be described in Fig.6. Differing from BP, the input layer in

RBF only passes information, with the connection weighting at 1.0. The radial basis function

is incorporated in the hidden layer. Linear weightings are adjusted in the output layer. The

steps to perform RBF-based network can be found in Billings et al. [30] and Chen et al. [31].

Fig.6. topology structure of RBF neural network

Although RBF has not been widely applied in predicting steel corrosion levels, certain

relevant existing studies can be found, such as Chou et al. [32] and Pai et al. [33].

The main difference between BP and RBF lies in that BP model utilizes the S-type

hyperbolic function in its hidden layer. In comparison, RBF is a non-linear network, with its

hidden layer utilizing Gaussian function. Both BP and RBF were adopted in the prediction of

rebar corrosion levels in this research. Both types of ANN models contain the three layers
(i.e., input, hidden, and output layers). Applied in the predication of rebar corrosion levels,

three nodes were included in the input layer for both models, namely compressive strength of

concrete specimens (X1), average ultrasonic velocity (X2), and the ratio (i.e., 2R/L) of rebar

diameter to test distance illustrated in Fig.1. Four treatment nodes were set in the hidden layer.

Rebar corrosion level was set as the single node in the output layer. The

Levenberg-Marquardt algorithm was adopted in the data training. The maximum training

time was set at 1,000. After the networks were established, data sample could be used for the

training. The data sample was divided into training sample and the cross-validation sample.

When the training was complete, the error was reduced and eventually approached 0, and

corresponding weightings were stored automatically.

3. Results and discussion

3.1. Laboratory observations

Under the laboratory condition, the interfaces between concrete and corroded reinforcements

were observed as presented in Fig.7.

Fig.7. Defects of interfaces between corroded reinforcement and concrete

The development of steel corrosion led to defects in the interfaces between reinforcement

and concrete as shown in Fig.7. To further identify the defect size of corrosion layers formed

in reinforced concrete prisms, the element distribution of the corrosion area was analyzed
adopting scanning electron microscopy (SEM) and electron probe. Adopting C30 specimen

as the showcase, the analysis is illustrated in Fig.8.

200
Ca 60
60
175
Fe
50
Fe Ca

Element content(%)
50 150

Element content(%)
Element content(%)

125 40
40

100 Fe
30
30
75 Ca
20
20
50
10
10 25

0 0
0
0 100 200 300 400 500 600 0 50 100 150 200 250 300 350 0 20 40 60 80 100 120 140
Distance(m) Distance(m) Distance(m)

Fig.8. Linear SEM/EDS analysis of chemical elements in corrosion layer

The black-colored line denoted with Fe in Fig.8 represents the rebar, and the red-colored line

denoted with Ca is concrete. Both lines underwent sudden sharp drops of their element

content as it can be seen in Fig.8, indicating the occurrence of corrosion. The length of the

overlapping zones of the two lines shows the corrosion depth. It is found from Fig.8 that the

corrosion depth was between 80μm~200μm, below the maximum depth of seriously corroded

rebar at 500μm according to Zhao et al. [19].

3.2. Effects of moisture contents in ultrasonic velocity

The experimental data of water content corresponding to the ultrasonic velocity is described

in Fig.9. Various moisture contents ranging from 0 to 2.5% in concrete were included.
 4.1

4.0 Specimen 1#
Specimen 2#
Acoustic velocity (km/s)

3.9

3.8

3.7

3.6

3.5
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Water content (%)
Fig.9. Relationship between ultrasonic velocity and moisture content

Specimens 1 and 2 in Fig.9 represent two different corrosion levels of reinforcement. It

can be found from Fig.9 that under both corrosion levels, ultrasonic velocity would not be

significantly affected when the moisture content in concrete is not higher than 1.5%. As the

moisture content increases from 1.5% in concrete specimens, the ultrasonic velocity would

grow more significantly to be 8% higher compared to that when concrete is fully dried. The

effect of moisture content in ultrasonic velocity could be due to the heterogeneous feature of

concrete structure. Voids and cracking in corroded concrete members are filled with air or

moisture. The moisture content in concrete structure would affect the air percentage in

concrete, resulting in the change of ultrasonic velocity. As ultrasonic wave travels

significantly faster in liquids than in the air, it would be significantly increased when concrete

cracking was filled with more moisture. The experimental tests indicated that the moisture

content lower than 1.5% could be considered as dry condition when the moisture is not

continuously or evenly distributed within concrete internal voids. As a result, the ultrasonic

velocity would not be significantly changed. This initial study of moisture content’s effect in

ultrasonic velocity provides the threshold value that in conducting the ultrasonic tests within

concrete structures, concrete should be kept dry with the moisture content below 1.5%.
 3.3. Ultrasonic test results

Ultrasonic velocities were tested within different types of specimens under different

corrosion levels before and after corrosion. Four typical types of recorded acoustic velocities

along the nine locations within the same specimen are showcased in Fig.10.

4.2 4.2

4.0 4.0

Acoustic velocity (km/s)

3.8 3.8
Acoustic velocity (km/s)

3.6 3.6

3.4 3.4

Before corrosion Before corrosion

3.2 3.2
After corrosion After corrosion

3.0 3.0
0 2 4 6 8 10 0 2 4 6 8 10
Measuring location Measuring location

4.2 4.2

4.0 4.0
Acoustic velocity (km/s)

Acoustic velocity (km/s)

3.8 3.8

3.6 3.6

3.4 3.4

3.2 Before corrosion 3.2 Before corrosion

After corrosion After corrosion
3.0 3.0
0 2 4 6 8 10 0 2 4 6 8 10
Measuring location Measuring location

Fig.10. Comparison of ultrasonic velocities before and after corrosion

Fig.10 displays four examples representing different corrosion levels of reinforcement.

All the four examples in Fig.10 demonstrate that ultrasonic velocity was significantly higher

in concrete before rebar corrosion. The velocity tested at the nine different areas of the same

specimen was generally stable with little variation before corrosion. After the rebar corrosion,

the ultrasonic velocity was significantly reduced, and the nine tested velocities within the

same concrete specimen had a larger variation. The variation of ultrasonic velocity could be

explained by the micro-structure of the concrete specimen. The number and size of defects

(e.g., pores and micro-cracks) within concrete structures are relatively small before rebar
corrosion, and the ultrasonic wave would mainly be transmitted with high velocity in a stable

way. Following the rebar corrosion, the corrosion products increase creating defects at the

concrete-reinforcement interfaces as shown in Fig.11. The final corrosion force leads to

concrete cracking and diffraction of ultrasonic wave. The reflection and scattering of

ultrasonic waves would then result in significant reductions of velocity according to Wang et

al. [34].

Fig. 11. Interface cracking of reinforced concrete due to corrosion expansion

3.4. Results of applying ANN to predicting rebar corrosion rates

Totally 50 observations were used as the data sample in applying BP and RBF models.

Among them 45 observations were adopted as the training set, and the rest five were used as

the cross-validation set. Fig.12 displays an example of data training.

a) Learning curve b) Comparing the desired and actual output corrosion

rates
Fig.12. Learning curves and predicting the corrosion mass loss ratio of using RBF model
The two ANN models were compared of their performance. Both models were able to

complete the data training and predict the rebar corrosion level based on the three input

parameters (i.e., X1, X2, and X3), where X1 stands for the design compressive strength pf

concrete, X2 is the average ultrasonic velocity within each observation sample, and X3 means

the ratio of rebar diameter to the test distance (denoted as 2R/L) as shown in Fig.1. The

dependent variable (i.e., corrosion level) was measured by the reinforcement mass loss ratio

as shown in Equ.(1).

m o  m1
S  100%
mo
Equ.(1)

where S denotes the mass loss ratio of reinforcement, m0 is the mass of the originally

non-corroded specimen, and m1 represents the mass of the corroded specimen after de-rusting.

Upon the end of each ultrasonic test in this research, corroded reinforcements were washed

using de-rusting solution to remove corrosion products after the corrosion process. The

appearances of rebar before and after de-rusting are shown in Fig.13.

(a) Reinforcement before de-rusting

(b) Reinforcement after de-rusting

Fig.13. Images of reinforcement before and after the corrosion

The removal solution applied in the de-rusting of rebar was prepared by mixing 3% of

hexamethylene tetramine with 97% of diluted hydrochloric acid by weight. The relative errors

of prediction from the cross-validation sample are listed in Table 2.

 Table 2. Corrosion mass loss ratio predicted by neural networks
Data
X X X Y Y Relative
Y （RBF） Relative
set 1 2 3
（Observed） （BP） error /% error /%
1 C25 3.6220 0.12 5.29 5.36 1.32 5.79 9.45
2 C30 3.8723 0.12 2.58 1.82 29.46 2.58 0.00
3 C35 3.9920 0.12 1.03 1.05 1.94 1.07 3.88
4 C20 3.4557 0.12 7.94 8.30 4.53 8.97 12.97
5 C30 4.1580 0.08 3.89 2.69 30.85 4.14 6.43

It can be found from Table 3 that the RBF model had a superior performance in terms of

generally lower relative errors. Several significantly large relative errors were found in the

validation sample adopting BP, such as data set 2 and 5 related to C30. The comparison

between BP and RBF showed that RBF had the lower standard error. Compared to BP, RBF

could better detect the non-linear relationship between the input and output layers with an

overall higher prediction accuracy, due to its own advantages in network structure, training

algorithms, and usage of network resources, specifically:

1) BP network applied weightings from its input layer to the hidden layer. It is not easy to

determine the number of hidden layer units in BP network. The network structure will not

be changed once it is confirmed in the training process. In comparison, the RBF network

utilizes the direct connection between its input and hidden layers, but weightings are

applied between its hidden layer and the output layer. RBF is capable of adjusting the

number of hidden layer units during the training.

2) theoretically, both BP and RBF models are capable of providing the non-linear prediction

approach. However, they apply different functions with varied performance of

approximation. BP network, with the S-shaped function as the transformation function,

uses the approximation of the nonlinear mapping. The RBF network is a network derived

from the regular theory. It uses the locally exponentially decaying Gaussian function to

locally approximate the nonlinear mapping and has the best approximation Nature

according to Zhang [35].

3) the determination of BP neural network weights and thresholds is directly determined by

the sum of the squared deviations of each output node. The trained network can only be a

compromise between different tasks, and the best effect cannot be achieved for one task.

In comparison, the distribution of hidden layer units in RBF network can be decided

according to the capacity, category and distribution of training samples. Based on the

distribution of hidden layer units, the mapping relationship between input and output can

be achieved by adjusting the mapping between the hidden layer’s and output layer’s units.

The impact of different tasks is relatively small, and the network resources can be fully

utilized.

It should be noted that the neural network models established in this study of

non-destructive test in reinforced concrete specimens serve as initial references in predicting

steel corrosion levels. Continuous data training will be performed as the sample size increases.

The future work will focus on building a more comprehensive prediction system with more

ANN models or data mining methods, as well as further improving the prediction

performance.

4. Conclusion

This study aimed to adopt the ultrasonic test as the non-destructive approach in

measuring the corrosion level of non-uniform surfaces of rebar within the reinforced concrete

structure. Then the artificial neural network was applied in training the experimental data for

the prediction of rebar corrosion level. Based on the experimental investigation and the neural

network modeling, the following findings can be summarized:

(1) The rebar-concrete interface was observed using scanning electron microscopy

(SEM) and electron probe. The corrosion depth in the reinforced prism was identified

between 80μm~200μm;
 (2) The effect of water content in concrete specimens’ ultrasonic velocity was studied

before conducting the corrosion tests. It was identified that moisture content below 1.5% in

concrete would not cause significant increases of ultrasonic velocity. Moisture content over

1.5% would cause more significant increase (i.e., up to 8%) of ultrasonic velocity;

(3) Rebar corrosion would cause significant reductions in ultrasonic velocity in concrete

specimens. Compared to the ultrasonic velocity before rebar corrosion, the ultrasonic velocity

tested at different zones of the same concrete specimen also tended to vary more significantly.

The larger variation of ultrasonic velocity after rebar corrosion was due to the

micro-structural perspective in the concrete structure;

(4) two types of artificial neural network models (i.e., BP and RBF) were applied in

predicting the rebar corrosion level based on 50 experimental observations. RBF was found

outperforming BP in predicting the rebar corrosion level due to the smaller average relative

error. The superior performance of RBF was described in comparison to BP in terms of

network structure, training algorithm, and network resource usage.

The current study is limited in its data sample size and the types of neural network

models. Future research would recruit a larger data sample and more neural models to further

explore the application of artificial neural network and other data mining methods in

predicting steel rebar’s non-uniform corrosion-induced damage in reinforced concrete

members. More future research needs to be performed to extend the current study, which only

adopted a single rebar in the test specimen. The current study adopted the average velocity of

nine test areas within the same specimen to estimate the overall corrosion level of

reinforcement. But the current study can lead to future research if different test areas of the

same specimen are recorded and input for data analytics. Overall, this research serves as the

initial movement towards applying the ultrasonic test as one non-destructive method to

estimate reinforcement corrosion level within concrete.

5. Acknowledgement

The authors wish to acknowledge the financial support of National Natural Science

Foundation of China (Grant No. 51778577), Zhejiang Provincial Natural Science Foundation

(Grant No. LY15E080025 and Grant No. LY16E020014), and the Ningbo Natural Science

Foundation (Grant No. 201601HJ-B01068).

6. References

[1] G.L. Feng, L.Y. Li, B.Kim B, Q.F..Liu, Multiphase modelling of ionic transport in
cementitious materials with surface charges, Comput. Mater. Sci. 111(2016) 339–349.

[2] Q.F. Liu, G.L. Feng, J. Xia, J.Yang, L.Y.Li, Ionic transport features in concrete
composites containing various shaped aggregates: a numerical study, Comp. Struc. 183(2018)
371-380.

[3] Z.Hu, L.X. Mao, J. Xia, J.B.Liu, J. Gao, J. Yang, Q.F. Liu, Five-phase modelling for
effective diffusion coefficient of chlorides in recycled concrete. Magazine of Concrete
Research.2017.

[4] M. Raupach, Half-cell potential measurements–Potential mapping on reinforced concrete

structures.

[5]ASTM G57-06(2012), Standard Test Method for Field Measurement of Soil Resistivity

Using the Wenner Four-Electrode Method, ASTM International, West Conshohocken, PA,

2012.

[6] Huygens-Fresnel Principle ] Born, M; Wolf, E. Principles of Optics 7th (expanded).

Cambridge University Press. (1999).

[7] G. Trtnik, F. Kavčič, G.Turk, Prediction of concrete strength using ultrasonic pulse

velocity and artificial neural networks, Ultrasonics. 49(1)(2009) 53-60.

[8] M.A. Kewalramani, R.Gupta, Concrete compressive strength prediction using ultrasonic

pulse velocity through artificial neural networks, Auto. Constr., 15(3)(2006) 374-379.

[9] S. Chithra, S.R.R. Senthil Kumar, K. Chinnaraju, F. Alfin Ashmita, A comparative study

on the compressive strength prediction models for High Performance Concrete containing

nano silica and copper slag using regression analysis and Artificial Neural Networks, Constr.

Build. Mater. 114 (2016) 528-536.

[10] B.A. Omran, Q. Chen, R. Jin, Comparison of different data mining techniques for

predicting compressive strength of environmentally friendly concrete, J. Comput. Civil.

Eng. 30 (6) (2016), DOI: 10.1061/(ASCE)CP.1943-5487.0000596.

[11] Ł. Sadowski, J. Hoła, S. Czarnecki, Non-destructive neural identification of the bond

between concrete layers in existing elements, Constr. Build. Mater. 127 (2016) 49-58.

[12] B. Wang, T. Man, H. Jin. Prediction of expansion behavior of self-stressing concrete by

artificial neural networks and fuzzy inference systems, Constr. Build. Mater. 84 (2015)

184-191.

[13] Y. Wang, Y. Shao, M.D.Matovic, J.K.Whalen, Recycling combustion ash for sustainable

cement production: A critical review with data-mining and time-series predictive models, Constr.

Build. Mater. 123 (2016) 673-689.

[14]M. Velay-Lizancos, J.L. Perez-Ordoñez, I. Martinez-Lage, P. Vazquez-Burgo, Analytical and

genetic programming model of compressive strength of eco concretes by NDT according to

curing temperature, Constr. Build. Mater. 144 (2017) 195-206.

[15] A, Feio, J.S. Machado, In-situ assessment of timber structural members: Combining

information from visual strength grading and NDT/SDT methods – A review, Constr. Build.

Mater. 101 (2016) 1157-1165.

[16] W. Yeih, R. Huang, Detection of the corrosion damage in reinforced concrete members by

ultrasonic testing, Cem. Con. Res. 28(7) (1998) 1071-1083.

[17] Y. Zhao, W. Jin, Chapter 6 - Rust Distribution in Corrosion-Induced Cracking Concrete,

In Steel Corrosion-Induced Concrete Cracking, Butterworth-Heinemann, (2016) 93-109

[18] Y. Zhao, W. Jin, Chapter 9 - Development of Corrosion Products-Filled Paste at the

Steel–Concrete Interface, In Steel Corrosion-Induced Concrete Cracking,

Butterworth-Heinemann, (2016) 147-158,

[19] Y. Zhao, B. Hu, J. Yu, Non-uniform distribution of rust layer around steel bar in

concrete, Corro. Sci. 53(12) (2011) 4300-4308.

[20] Z. Zhang, P. Qiu, Ultrasonic application in concrete property evaluation, 2006, Chemical

Industry Publisher, Beijing (in Chinese)

[21] B. D. Liu, W.J. Lv, L. Li, P. F. Li, Effect of moisture content on static compressive

elasticity modulus of concrete, Constr. Build. Mater. 69 (2014) 133-142.

[22] B. Lyes Bal, B.B. François, Artificial neural network for predicting drying shrinkage of
concrete, Constr. Build. Mater. 38 (2013) 248-254.

[23] NeuroSolutions. NeuroSolutions tutorials. 2015. Available via <

http://www.neurosolutions.com/neurosolutions/help/index.html?NeuroSolutionsTutorials.htm

l >, accessed on Sep 12th, 2016.

[24] D. Zohra, Z. Mehdi Sbartaï, C. Arnaud, G. Fouad, Artificial neural network model for
steel–concrete bond prediction, Eng. Struc. 31 (8) (2009)1724-1733,

[25] M. Uysal, H. Tanyildizi, Estimation of compressive strength of self compacting concrete

containing polypropylene fiber and mineral additives exposed to high temperature using
artificial neural network, Constr. Build. Mater., 27 (1) (2012) 404-414,

[26] M. Uysal, H. Tanyildizi, Predicting the core compressive strength of self-compacting

concrete (SCC) mixtures with mineral additives using artificial neural network, Constr. Build.
Mater. 25 (11) (2011) 4105-4111.

[27] L. Sadowski, Non-destructive investigation of corrosion current density in steel reinforced

concrete by artificial neural networks, Archives of Civil and Mechanical Engineering.

13(1)(2013) 104-111.
[28] M. Buscema, Back propagation neural networks, Substance use & misuse, 33(2)(1998)

233-270.

[29] R. Hecht-Nielsen, Theory of the backpropagation neural network, Neural Networks,

1(Supplement-1) (1998) 445-448.

[30] S.A. Billings, G.L. Zheng, Radial basis function network configuration using genetic

algorithms, Neural Networks. 8(6)(1995) 877-890.

[31] S. Chen, C.F. Cowan, P.M. Grant, P.M., Orthogonal least squares learning algorithm

for radial basis function networks, IEEE Transactions on neural networks, 2(2)(1991)

302-309.

[32] J.S. Chou, N.T. Ngo, W.K. Chong, The use of artificial intelligence combiners for
modeling steel pitting risk and corrosion rate, Engineering Applications of Artificial
Intelligence. 65 (2017) 471-483.

[33] P. S. Pai, M.T. Mathew, M.M. Stack, L.A. Rocha, Some thoughts on neural network
modelling of microabrasion–corrosion processes, Tribology International, 41(7) (2008)
672-681.

[34] K. Wang, M. Chen, L. Xie, Research on the evolution of bending fatigue and damage of

reinforced concrete beams under stray current corrosion, Tiedao Xuebao. 34(11)(2012) 88-93.

(in Chinese)

[35] D. Zhang, The application of neural network in the smart optimized control system,

Computer Engineering and Design.14(4)(2008) 8-10. (in Chinese)