materials

Article
Evaluation of Early Concrete Damage Caused by
Chloride-Induced Steel Corrosion Using a Deep Learning
Approach Based on RNN for Ultrasonic Pulse Waves
Julfikhsan Ahmad Mukhti 1 , Kevin Paolo V. Robles 1 , Keon-Ho Lee 2, * and Seong-Hoon Kee 1,3, *

1 Department of ICT Integrated Ocean Smart Cities Engineering, Dong-A University,

Busan 49304, Republic of Korea; [email protected] (J.A.M.); [email protected] (K.P.V.R.)
2 Department of Architectural Engineering, Dong-A University, Busan 49304, Republic of Korea
3 National Core Research Center for Disaster-Free and Safe Ocean Cities Construction, Dong-A University,
Busan 49304, Republic of Korea
* Correspondence: [email protected] (K.-H.L.); [email protected] (S.-H.K.)

Abstract: The objective of this study is to explore the feasibility of using ultrasonic pulse wave
measurements as an early detection method for corrosion-induced concrete damages. A series of
experiments are conducted using concrete cube specimens, at a size of 200 mm, with a reinforcing
steel bar (rebar) embedded in the center. The main variables include the water-to-cement ratio of
the concrete (0.4, 0.5, and 0.6), the diameter of the rebar (10 mm, 13 mm, 19 mm, and 22 mm), and
the corrosion level (ranging from 0% to 20% depending on rebar diameter). The impressed current
technique is used to accelerate corrosion of rebars in concrete immersed in a 3% NaCl solution.
Ultrasonic pulse waves are collected from the concrete specimens using a pair of 50 kHz P-wave
transducers in the through-transmission configuration before and after the accelerated corrosion test.
Deep learning techniques, specifically three recurrent neural network (RNN) models (long short-term
memory, gated recurrent unit, and bidirectional long short-term memory), are utilized to develop a
classification model for early detection of concrete damage due to rebar corrosion. The performance
Citation: Mukhti, J.A.; Robles, K.P.V.; of the RNN models is compared to conventional ultrasonic testing parameters, namely ultrasonic
Lee, K.-H.; Kee, S.-H. Evaluation of pulse velocity and signal consistency. The results demonstrate that the RNN method outperforms
Early Concrete Damage Caused by
the other two methods. Among the RNN methods, the bidirectional long short-term memory RNN
Chloride-Induced Steel Corrosion
model had the best performance, achieving an accuracy of 74% and a Cohen’s kappa coefficient of
Using a Deep Learning Approach
0.48. This study establishes the potentiality of utilizing deep learning of ultrasonic pulse waves with
Based on RNN for Ultrasonic Pulse
RNN models for early detection of concrete damage associated with steel corrosion.
Waves. Materials 2023, 16, 3502.
https://doi.org/10.3390/
ma16093502 Keywords: concrete deterioration; non-destructive evaluation; ultrasonic testing; deep learning;
recurrent neural network
Academic Editor: Krzysztof
Schabowicz

Received: 25 March 2023

Revised: 16 April 2023 1. Introduction
Accepted: 25 April 2023 Chloride-induced corrosion of steel is one of the most significant sources of deteriora-
Published: 1 May 2023 tion in reinforced concrete structures [1–3]. It has been known that the chloride-induced
deterioration mechanism occurs in three phases: corrosion initiation, rust propagation,
and corrosion acceleration, as illustrated in Figure 1 [4–6]. The corrosion initiation pro-
cess starts when the thin passive layer (Fe2 O3 ) on the surface of the reinforcing steel bar
Copyright: © 2023 by the authors.
(rebar) becomes unstable and depleted due to the migration of chloride ions [7,8]. The
Licensee MDPI, Basel, Switzerland.
rust then propagates and begins to form on the rebar surface. This oxidation process of
This article is an open access article
the metallic iron causes the rebar volume to increase up to about sixfold on the oxidized
distributed under the terms and
section compared to its uncorroded state [9]. This increasing volume induces internal
conditions of the Creative Commons
tensile stress on the concrete, which could result in enhanced porosity and microcracks in
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
the steel and concrete interfaces [10]. As the tensile stress increases, the microcracks start
4.0/).
to spread and open more paths for chloride ions to penetrate the concrete [11], increasing

Materials 2023, 16, 3502. https://doi.org/10.3390/ma16093502 https://www.mdpi.com/journal/materials

Materials 2022, 15, x FOR PEER REVIEW 2 of 29

Materials 2023, 16, 3502 2 of 26

concrete interfaces [10]. As the tensile stress increases, the microcracks start to spread and
open more paths for chloride ions to penetrate the concrete [11], increasing the permea-
bility of the concrete. This also accelerates the penetration rate of harmful substances and
the permeability
moisture within theof concrete
the concrete. This
and the also accelerates
corrosion the penetration
process. Without rate
sufficient of harmful
maintenance
substances and moisture within the concrete and the corrosion process. Without
activities, this could lead to major damages such as surface-breaking cracks, spalling, sufficient
and
maintenance activities, this could lead to major damages such as surface-breaking
delamination defects, which affect the integrity of the structure [12]. Moreover, beyond cracks,
spalling,
the point and delamination
of corrosion defects, which
acceleration, affect
the cost of the integrity of the
rehabilitation structure [12].
of reinforced Moreover,
concrete ele-
beyond the point of corrosion acceleration, the cost of rehabilitation of reinforced concrete
ments would exponentially increase after the formation of visible concrete damages such
elements would exponentially increase after the formation of visible concrete damages
as surface-breaking cracks and spalling [13] (see Figure 1). Therefore, it is important to
such as surface-breaking cracks and spalling [13] (see Figure 1). Therefore, it is important
evaluate the condition of reinforced concrete exposed to harsh environmental conditions
to evaluate the condition of reinforced concrete exposed to harsh environmental condi-
on corrosion of steel in concrete and, if necessary, to perform appropriate preventive
tions on corrosion of steel in concrete and, if necessary, to perform appropriate preventive
maintenance actions.
maintenance actions.
Initiation Rust Propagation Acceleration

Passivation
Depassivation Surface-breaking
Rust layer crack formation
Concrete Condition

expansion Crack
formation

Repair Cost
Concrete
spalling
Demolition &
Replacement

1 2 3 4 5 6
Patching/ Member
grouting failure
Surface coating Sealing/
Coating

Time
Figure 1. Conceptual illustration of concrete deterioration caused by corrosion of steel in concrete
Figure 1. Conceptual illustration of concrete deterioration caused by corrosion of steel in concrete
and cost
and cost for
for rehabilitation
rehabilitation of
of the
the deteriorated
deteriorated concrete
concrete with
with various
various severities
severities [14].
[14]. Deterioration
Deterioration
stages: 1 rebar de-passivation; 2 rust layer expansion; 3 crack formation; 4 surface-breaking
stages: ⓛ rebar de-passivation; ② rust layer expansion; ③ crack formation; ④ surface-breaking
crack formation; 5 concrete spalling; and 6 member failure.
crack formation; ⑤ concrete spalling; and ⑥ member failure.
It is known that microscopic changes in corrosion-induced concrete damages can be
It is known
effectively that microscopic
evaluated changes intesting
by several laboratory corrosion-induced
techniques such concrete damageselectron
as scanning can be
effectively
microscopyevaluated by several
(SEM), nuclear laboratory
magnetic resonance testing
(NMR), techniques such as scanning electron
and thermo-gravimetric/derivative
microscopy
thermo-gravimetric(SEM), nuclear
(TG/DTG) magnetic resonance
[15]. However, (NMR),
these methods andrequire
thermo-gravimetric/deriva-
an invasive procedure
tive thermo-gravimetric
to obtain testing samples (TG/DTG) [15]. However,
for laboratory inspection theseandmethods
inevitablyrequire
inducean invasive
some surfacepro-
damages
cedure to in concrete.
obtain testingFurthermore,
samples for these methodsinspection
laboratory could include and additional
inevitablyprocesses
induce some (e.g.,
surfacedamages
surface preparation, sampling,
in concrete. and repairing)
Furthermore, these when
methods applied to actual
could includestructures,
additionalwhich pro-
makes(e.g.,
cesses the methods labor-intensive
surface preparation, and high-priced.
sampling, and repairing) Therefore,
wheninvasive
applied testing
to actual methods
struc-
are reluctantly
tures, which makes usedthe formethods
condition assessment ofand
labor-intensive actual structures.Therefore, invasive test-
high-priced.
There are
ing methods arevarious non-destructive
reluctantly used for conditionevaluation (NDE) techniques
assessment that are effective for
of actual structures.
in situ evaluation
There of chloride-induced
are various non-destructivesteel corrosion
evaluation (NDE) in concrete structures.
techniques that areThe corrosion
effective for
of steel in concrete can be explained by an electrochemical process
in situ evaluation of chloride-induced steel corrosion in concrete structures. The corrosion in which both flows of
electrical currents and chemical reactions occur. NDE methods
of steel in concrete can be explained by an electrochemical process in which both flows of based on electrochemical
principlescurrents
electrical have been andwidely usedreactions
chemical to evaluate the initiation
occur. NDE methods and activity
based on of steel corrosion in
electrochemical
concrete. For example, the half-cell potential (HCP) measurement
principles have been widely used to evaluate the initiation and activity of steel corrosion is used to investigate the
electrical
in concrete. activity of steel corrosion
For example, the half-cellin concrete
potential induced
(HCP) by chloride ionsis[16].
measurement usedThe probability
to investigate
of corrosion activity of steel in concrete is determined based on
the electrical activity of steel corrosion in concrete induced by chloride ions [16]. TheHCP (or corrosion potential,
prob-
E corr ) readings in accordance with ASTM C876-15. Furthermore,
ability of corrosion activity of steel in concrete is determined based on HCP (or corrosion the rate of steel corrosion
(or corrosion
potential, Ecorr)current
readings density, icorr ) canwith
in accordance be measured
ASTM C876-15.by the polarization
Furthermore, resistance
the rate of ofsteel
steel
in concrete, R , which is directly proportional
corrosion (or pcorrosion current density, icorr) cancorr to i . There are several NDT
be measured by the polarization re- methods for
measuring the polarization resistance of steel in concrete: linear
sistance of steel in concrete, Rp, which is directly proportional to icorr. There are several polarization resistance
method [17], Tafel extrapolation method [18], electrochemical impedance spectroscopy [19],
etc. The electrical resistivity (ER) measurement has also been widely studied for evaluating
the corrosive environment (e.g., water saturation and chloride penetration) of concrete and
 NDT methods for measuring the polarization resistance of steel in concrete: linear polari-
zation resistance method [17], Tafel extrapolation method [18], electrochemical imped-
Materials 2023, 16, 3502
ance spectroscopy [19], etc. The electrical resistivity (ER) measurement has also 3been of 26

widely studied for evaluating the corrosive environment (e.g., water saturation and chlo-
ride penetration) of concrete and the corrosion rate of steel in concrete [20]. However, it
the been
has corrosion rate of steel
challenging in concrete
to evaluate [20]. However,
the damage it has
of concrete been challenging
associated with steeltocorrosion
evaluate
theelectrochemical
by damage of concrete associatedparameters
measurement with steel corrosion
such as HCP by electrochemical
(or Ecorr), Rp, icorrmeasurement
, and/or ER.
parameters
Those such ascould
parameters HCPbe Ecorr ), affected
(orhighly Rp , icorr , and/or
by variousER. environmental
Those parameters could
factors be highly
(e.g., water
affected by various environmental factors (e.g., water saturation,
saturation, humidity, and temperature) which are not really correlated with concrete de- humidity, and tempera-
ture) which(e.g.,
terioration are not really correlated
enhanced with concrete
porosity, microcracks, deterioration
and surface-breaking(e.g., enhanced
cracks). porosity,
microcracks, and surface-breaking cracks).
On the other hand, it has been demonstrated by numerous researchers that ultrasonic
pulseOnwavethemeasurements
other hand, itare haseffective
been demonstrated
for evaluating bythenumerous researchers
various types that dete-
of concrete ultra-
sonic pulse
riorations wave
such as measurements
honeycombingare effective
[21], for evaluating
delamination the various
defects [22], types of concrete
surface-breaking cracks
deteriorations
[23,24], such as
microcracks honeycombing
[25], and bottom-up [21], delamination
cracks [26]. Figuredefects [22], surface-breaking
2 illustrates the ultrasonic
cracks [23,24], microcracks [25], and bottom-up cracks
pulse wave propagation through concrete, a heterogenous and anisotropic [26]. Figure 2 illustrates the ultra-
material. A
sonic pulse wave propagation through concrete, a heterogenous
transmitting transducer, placed on one side of the concrete, generates ultrasonic and anisotropic material.
pulse
A transmitting
waves travellingtransducer,
through the placed on one
concrete, whichsideareof the concrete,
measured bygenerates
a receiving ultrasonic
transducer pulse
on
waves travelling through the concrete, which are measured by a receiving
the opposite side of the concrete. Concrete acts as a low-pass filter for ultrasonic pulse transducer on
the opposite
waves. Some side of the concrete.
low-frequency Concrete
components acts aspropagate
directly a low-pass to filter for ultrasonic
the receiver, pulse
while some
waves. Some low-frequency components directly propagate to
high-frequency components are suppressed and/or delayed by reflection and/or diffusion the receiver, while some
high-frequency components are suppressed and/or delayed by reflection
due to the heterogeneous and anisotropic features of concrete. Theoretically, the earlier and/or diffusion
due to the heterogeneous and anisotropic features of concrete. Theoretically, the earlier
part (also called the coherent part) of the ultrasonic signals is informative of the global
part (also called the coherent part) of the ultrasonic signals is informative of the global
properties of concrete, while the later part (also called the incoherent part) of the signals
properties of concrete, while the later part (also called the incoherent part) of the signals
is a result of the superposition of the diffused waves from the presence of aggregates and
is a result of the superposition of the diffused waves from the presence of aggregates and
various defects in concrete [27]. Several previous researchers have used the coherent part
various defects in concrete [27]. Several previous researchers have used the coherent part of
of the ultrasonic pulse waves to evaluate the properties of concrete [20,22,27]. For exam-
the ultrasonic pulse waves to evaluate the properties of concrete [20,22,27]. For example, it
ple, it has been demonstrated that ultrasonic pulse velocity (UPV) of concrete is a good
has been demonstrated that ultrasonic pulse velocity (UPV) of concrete is a good indicator
indicator of the overall quality of concrete with various deterioration levels. The presence
of the overall quality of concrete with various deterioration levels. The presence of defects
of defects in concrete would delay the first arrival time (time of flight) of the ultrasonic
in concrete would delay the first arrival time (time of flight) of the ultrasonic pulse waves
pulse waves through concrete [27]. UPV has been demonstrated to be sensitive to major
through concrete [27]. UPV has been demonstrated to be sensitive to major faults or open
faults or open fractures, which serve as an effective barrier to ultrasonic wave transmis-
fractures, which serve as an effective barrier to ultrasonic wave transmission. However,
sion. However,
it is known that itUPV
is known that to
is unlikely UPV is unlikely
be affected to beminor
by such affected by as
faults such minor faults
enhanced porosityas
enhanced porosity and/or micro-, ill-defined, and closed cracks
and/or micro-, ill-defined, and closed cracks in concrete, which are generated by steel in concrete, which are
generated
corrosion in byconcrete.
steel corrosion in concrete.

Concrete

Output signal
④
ⓛ ⑤ ③ ②

Transmitter Receiver

ⓛ Input signal Coarse aggregate

② Coherent part in output signal Fine aggregate
③ Incoherent part in output signal Internal cracks
④ Low-frequency component Path of low-frequency signal
⑤ High-frequency component Path of high-frequency signal

Figure 2.
Figure Illustration of
2. Illustration of the
theultrasonic
ultrasonicpulse
pulsewave
wavepropagation through
propagation concrete,
through a heterogenous
concrete, and
a heterogenous
anisotropic
and material.
anisotropic material.

Some researchers found that the signal interpretation based on the change in the
incoherent part is more effective for evaluating the early concrete damages using nonlinear
ultrasonic parameters such as coda wave interferometry (CWI), sideband peak count index
(SPC-I), and energy redistribution index [25]. Schurr et al. [28] used CWI to detect small-
Materials 2023, 16, 3502 4 of 26

scale concrete damages caused by external loadings. In their research, the change in the
phase shift of the incoherent part of ultrasonic signals is far clearer than that of the coherent
part. Research by Castellano et al. [29] observed that more cracks and voids in concrete,
caused by cyclic loadings, increased the SPC-I, the number of peaks above a threshold in
the spectral amplitude of ultrasonic signals. It was demonstrated that SPC-I is sensitive to
minor concrete damages caused by the early load steps, while UPV values remained stable.
Furthermore, Arumaikani et al. [24] noticed that the SPC-I is effective in evaluating internal
concrete damages caused by corrosion of steel in concrete that could not be observed
on the surface of the concrete. However, it has been argued that the performance of the
evaluation model based on such nonlinear ultrasonic parameters can be strongly dependent
on engineering judgments. As will be discussed in this study, the sensitivity of the nonlinear
parameters could be affected by the choice of nonlinear parameters and input signals used
for the calculation of the parameters.
Another, more systematic solution for exploiting ultrasonic pulse waves to detect early
concrete damage is the use of deep learning. Since ultrasonic pulse signals are time series,
the data sequence is an important feature, and the recurrent neural network (RNN) would
be a good method for deep learning of ultrasonic pulse waves. The RNN method has been
successfully used in several fields, most notably in medicine. Singh et al. [30] utilized the
RNN models for the classification of electrocardiogram (ECG) data for detecting arrhythmia
with an accuracy of up to 88.1%. Kim et al. [31] utilized both RNN and convolutional
neural network (CNN) methods for the classification of ECG data. In their study, the RNN
models perform slightly better than CNN. Additionally, the RNN method has also been
successfully applied for more complex signals such as human speech. Rejaibi et al. [32]
applied the RNN method for differentiating human voices in depressed and non-depressed
states with more than 70% accuracy. However, the application of RNN to monitor concrete
conditions based on ultrasonic pulse data is currently very limited.
The main objective of this research is to investigate the feasibility of ultrasonic pulse
wave measurements as an early detection method for corrosion-induced concrete damages
by using a deep learning classification model based on RNNs. For these purposes, this study
aims to perform three main tasks as follows: (1) evaluating the change of parameters in
the ultrasonic pulse signal from each method to corrosion levels, which will be undertaken
by performing accelerated corrosion to reinforced concrete specimens, (2) developing
classification models through deep learning of ultrasonic pulse waves based on three RNN
algorithms (long short-term memory, gated recurrent unit, and bidirectional long short-
term memory), and (3) developing classification models based on conventional ultrasonic
testing parameters (ultrasonic pulse wave velocity and signal coherence). This study
will demonstrate the potential of deep learning classification models based on RNN of
ultrasonic pulse waves for early detection of concrete damages, which is superior to the
classification models based on the conventional ultrasonic testing parameters.

2. Materials and Methods

2.1. Fabrication of Test Specimens
Figure 3 illustrates a reinforced concrete cube specimen with a size of 200 × 200 ×
200 mm3 used in this study. A 235 mm long reinforcing steel bar (rebar) was embedded in
the middle of the concrete cube specimen (see Figure 3b). A middle part of the rebar was
waterproofed by three layers of coating (see Figure 3c): first, two thin layers of urethane
were applied on the surface of the rebar; second, Teflon tape was rounded on the hardened
urethane layer; third, a 100 mm long PVC pipe was placed on the Teflon coating in the
middle of the rebar. Consequently, only 70 mm of the rebar was directly exposed for
accelerated corrosion. The concrete cubes were shaped using 20 mm thick wooden forms.
Concrete was cast as the rebar was horizontally situated through a punch-hole on one side
of the forms. Concrete specimens were cured in the air for 24 h after casting concrete. After
demolding, the concrete specimens were moved to and stored in a constant temperature
and humidity room in the laboratory (temperature of 20 ± 3 ◦ C and relative humidity of
 the middle of the rebar. Consequently, only 70 mm of the rebar was directly exposed for
accelerated corrosion. The concrete cubes were shaped using 20 mm thick wooden forms.
Concrete was cast as the rebar was horizontally situated through a punch-hole on one side
Materials 2023, 16, 3502 of the forms. Concrete specimens were cured in the air for 24 h after casting concrete.5After of 26
demolding, the concrete specimens were moved to and stored in a constant temperature
and humidity room in the laboratory (temperature of 20 ± 3 °C and relative humidity of
50±
50 ± 5%). The concrete cube specimens were divided divided into
into four
four groups
groups based
based on
onnominal
nominal
diameters of
diameters of rebars in concrete: 10 10 mm,
mm,1313mm,
mm,1919mm,mm,and and2222mm, mm,which
which are referred
are referredto
asas
to D10, D13,
D10, D19,
D13, D19,and D22,
and respectively.
D22, Each
respectively. group
Each group hashas
three sub-groups
three sub-groupswith three
with dif-
three
ferent design
different design compressive
compressivestrengths
strengthsofofconcrete:
concrete:1818MPa,
MPa,24 24MPa,
MPa, and
and 40 MPa. Table
Table 11
showsthe
shows themixture
mixtureproportions
proportionsofofthe the concrete
concrete used
used forfor
thethe fabrication
fabrication of the
of the concrete
concrete cu-
cubes.
bes. Rebars
Rebars in theinconcrete
the concrete specimens
specimens werewere subjected
subjected to various
to various corrosion
corrosion levels.
levels. TheThe tar-
target
get corrosion
corrosion levelslevels in study
in this this study
werewere 0%, 6%,
0%, 3%, 3%, and
6%, 12%
and for12%D10 for and
D10 D13
and specimens
D13 specimensand
0%,
and5%,0%,10%, and 20%
5%, 10%, and for
20%D19
for and
D19D22and specimens. In addition,
D22 specimens. therethere
In addition, werewere
threethree
copies of
cop-
concrete cube specimens
ies of concrete for each
cube specimens forcombination of test of
each combination variables. Consequently,
test variables. this study
Consequently, this
included a total of
study included 108 reinforced
a total concrete
of 108 reinforced cube specimens.
concrete cube specimens.

Figure3.3.Illustration
Figure Illustration ofof aareinforced
reinforcedconcrete
concrete cube
cube specimen
specimen used
used in
in this
this study:
study: (a)
(a) isometric
isometric view,
view,
(b) section AA in (a), (c) section BB in (b), and (d) detail in a dashed box in
(b) section AA in (a), (c) section BB in (b), and (d) detail in a dashed box in (b). (b).

Table 1. Summary of mixture proportions of concrete used for fabrication of concrete cube specimens
in this study.

Mix w/c Mixture Proportion (kg/m3 )

Design (%) W C S G AE
Mix 1 58.5 168 287 957 898 2.58
Mix 2 50.7 170 335 870 956 2.50
Mix 3 34.6 166 480 720 993 4.32
Note: W: water, C: cement, SV : volume of sand, AV : volume of aggregates, C: Portland cement type I, S: sand, G:
gravel, AE: high-performance air-entraining agent.
 Mix 1 58.5 168 287 957 898 2.58
Mix 2 50.7 170 335 870 956 2.50
Mix 3 34.6 166 480 720 993 4.32
Note: W: water, C: cement, SV: volume of sand, AV: volume of aggregates, C: Portland cement type
I, S: sand, G: gravel, AE: high-performance air-entraining agent.
Materials 2023, 16, 3502 6 of 26
2.2. Accelerated Corrosion Tests
The rebars in concrete cube specimens were subjected to an accelerated corrosion
2.2. Accelerated
process Corrosion
to simulate Tests
the chloride-induced steel corrosion using the impressed current tech-
nique. The target theoretical corrosion levels
The rebars in concrete cube specimens were insubjected
this studytowere calculatedcorrosion
an accelerated based onpro-the
following Faraday’s Law [33]:
cess to simulate the chloride-induced steel corrosion using the impressed current technique.
The target theoretical corrosion levels in this study 𝑊𝐼 were𝑇 calculated based on the following
Faraday’s Law [33]: 𝑀 = (1)
𝐹
W Iapp T
where 𝑀 is the theoretical mass density Mth =
of steel
F rust (kg/cm ), W is the equivalent weight
2 (1)
of steel as a ratio of the atomic iron weight to the iron valency (27,925 g), I app is the current
2 ), W is the equivalent weight
where
density Mapplied
th is the theoretical mass density
to the specimen (A/cm2),ofTsteel rust
is the (kg/cmof
duration current flows (or corrosion
of steel as a ratio of the atomic iron weight to the iron valency (27,925 g), Iapp is the current
process) (s), and F is Faraday’s constant2(~96,487 As). Before the accelerated corrosion pro-
density applied to the specimen (A/cm ), T is the duration of current flows (or corrosion
cess, each specimen was immersed in a 3% NaCl solution until fully saturated condition.
process) (s), and F is Faraday’s constant (~96,487 As). Before the accelerated corrosion
Then, the specimen was electrified on a setup shown the Figure 4. The positive pole of the
process, each specimen was immersed in a 3% NaCl solution until fully saturated condition.
power supply (Sorensen XPF 60-20D) was attached to the rebar, which effectively made
Then, the specimen was electrified on a setup shown the Figure 4. The positive pole of the
the rebar an anode. The negative pole of the power supply was connected to the stainless-
power supply (Sorensen XPF 60-20D) was attached to the rebar, which effectively made the
steel mesh (i.e., SUS 316) around the specimen, which served as a cathode. The corrosion
rebar an anode. The negative pole of the power supply was connected to the stainless-steel
of the rebars in the concrete specimens started as soon as the current was sent by the
mesh (i.e., SUS 316) around the specimen, which served as a cathode. The corrosion of
power supply. A digital multimeter (Keysight 34461A) was located between the power
the rebars in the concrete specimens started as soon as the current was sent by the power
supply and the stainless-steel mesh to monitor the current flowing through the concrete
supply. A digital multimeter (Keysight 34461A) was located between the power supply
specimens. The current data measured by the digital multimeter were stored in a desktop
and the stainless-steel mesh to monitor the current flowing through the concrete specimens.
computer
The currentthrough a LabVIEW-based
data measured monitoring
by the digital multimeter program and the
were stored theoretical
in a corrosion
desktop computer
level of rebar was automatically calculated in real time in the program.
through a LabVIEW-based monitoring program and the theoretical corrosion level of rebar The accelerated
corrosion
was test for calculated
automatically each specimen
in realcontinued
time in theuntil the target
program. theoreticalcorrosion
The accelerated corrosiontest levels
for
were reached.
each specimen continued until the target theoretical corrosion levels were reached.

Figure4.4.Test
Figure Testsetup
setupfor
forthe
theaccelerated
acceleratedcorrosion
corrosiontest
testby
bythe
theimpressed
impressedcurrent
currenttechnique.
technique.

2.3. Steel Mass Loss Ratio

In this study, the steel mass loss ratio of corroded rebars, θ (i.e., corrosion level in
this study), was defined as the ratio (in percentage) of the mass loss of the corroded rebar,
ms,loss (θ ), normalized by the mass of the solid rebar in the working area, ms,WA (0), as
follows:
ms,loss (θ )
θ= × 100 [%] (2)
ms,WA (0)
The actual amount of steel mass loss was evaluated in accordance with ASTM G1-
03 [34]. After the accelerated corrosion process finished, the specimens were broken into
two parts by using a splitting tensile test setup, and rebars in concrete were separated
from the concrete. The corroded rebars were first cleaned by using ultrasonic waves and
immersed in NaOH solution to remove the rust layer in steel bars. The sandblasting
 Here, θ is corrosion level of corroded rebars and ρs and ρw are mass densities of steel
and water, respectively. Then, the steel mass loss of corroded rebars, 𝑚 , (𝜃), was de-
termined by difference between 𝑚 , (0) and 𝑚 , (𝜃), as follows:
𝑚 , (𝜃) = 𝑚 , (0) − 𝑚 , (𝜃) (4)
Materials 2023, 16, 3502 7 of 26
Figure 6 shows the comparison of theoretical and actual steel loss of the reinforcing
steel considering all design mixes and rebar diameters. It can be observed that the theo-
retical steel loss generally overestimates the actual steel loss. In this study, the actual (or
method
measured)wassteel
usedloss
to remove the remaining
values were rust that was
used for correlating thenot removed
degree using the
of concrete mentioned
damages and
procedure, which resulted in cleaned rebars
the change in the ultrasonic pulse waves. (see Figure 5).

Figure
Figure5. Rebarcondition
5. Rebar conditionafter
aftercorrosion
corrosion
forfor
(a) (a)
D10D10
andand
D13D13 specimens,
specimens, and
and (b) (b)and
D19 D19D22
and D22
spec-
imens, respectively.
specimens, respectively.

The actual steel mass loss was determined by Archimedes’s principle of buoyancy.
First, the weight of the cleaned rebar was measured in the air, which is denoted as ms,air .
Second, the cleaned rebar was submerged in the water by 70 mm (the depth of the working
area), and the weight of the cleaned rebar was measured, which is denoted as ms,water .
Then, the mass of corroded rebars corresponding to the working area can be determined
as follows:
ms,WA (θ ) = (ms,air (θ ) − ms,water (θ ))ρs /ρw (3)
Here, θ is corrosion level of corroded rebars and ρs and ρw are mass densities of
steel and water, respectively. Then, the steel mass loss of corroded rebars, ms,loss (θ ), was
determined by difference between ms,WA (0) and ms,WA (θ ), as follows:

ms,loss (θ ) = ms,WA (0) − ms,WA (θ ) (4)

Figure 6 shows the comparison of theoretical and actual steel loss of the reinforcing
steel considering all design mixes and rebar diameters. It can be observed that the theo-
retical steel loss generally overestimates the actual steel loss. In this study, the actual (or
measured) steel loss values were used for correlating the degree of concrete damages and
the change in the ultrasonic pulse waves.

2.4. Ultrasonic Pulse Wave Measurements

Figure 7 illustrates the test setup of ultrasonic pulse wave measurements transmitted
through a reinforced concrete cube specimen. The standard test procedure according
to ASTM C 597/C597M-16 [35] was used to measure ultrasonic pulse waves through
concrete cube specimens that were subjected to different steel corrosion levels. The setup is
composed of a pulser-and-receiver (Olympus 5077PR), a digital oscilloscope (PXIe1073), a
pair of P-wave transducers (Olympus X1021), and a desktop computer for data acquisition,
display, and storage. The pulser-and-receiver droved the 50 kHz P-wave transducer by a 100
V rectangular pulse with a width of 10 µs. Transducers with a center frequency range of 50–
54 kHz are commonly used for NDE based on ultrasonic pulse wave data [36,37], including
studies for the detection of internal cracks due to corrosion [24]. The receiving transducer
placed on the opposite side of the concrete specimen measured the ultrasonic pulse waves
through the concrete. The received signal was digitized with a 10 MHz sampling rate by
the digital oscilloscope. Ultrasonic pulse wave measurements were conducted on each
specimen before and after the accelerated corrosion test. In this study, ultrasonic tests were
performed on the two test points on the surface of ultrasonic measurements (see Figures 3a
 Materials 2023, 16, 3502 8 of 26

and 7b). Sensors on test point 1 were attached to the concrete surface directly above the
Materials 2022, 15, x FOR PEER REVIEW 8 of 29
rebar, while sensors on test point 2 were located 50 mm horizontally from the center to
avoid the rebar. In this study, five measurements were repeated at each test point.

FOR PEER REVIEW 9 of 29

Figure6.6.Relationship
Figure Relationship between
between the
theactual
actual(measured)
(measured) and
andthe
thetheoretical
theoreticalmass
massloss
lossof
ofreinforcing
reinforcing
steel bars with various nominal diameters (D10, D13, D19, and D22).
steel bars with various nominal diameters (D10, D13, D19, and D22).

2.4. Ultrasonic Pulse Wave Measurements

Figure 7 illustrates the test setup of ultrasonic pulse wave measurements transmitted
through a reinforced concrete cube specimen. The standard test procedure according to
ASTM C 597/C597M-16 [35] was used to measure ultrasonic pulse waves through concrete
cube specimens that were subjected to different steel corrosion levels. The setup is com-
posed of a pulser-and-receiver (Olympus 5077PR), a digital oscilloscope (PXIe1073), a pair
of P-wave transducers (Olympus X1021), and a desktop computer for data acquisition,
display, and storage. The pulser-and-receiver droved the 50 kHz P-wave transducer by a
100 V rectangular pulse with a width of 10 µs. Transducers with a center frequency range
of 50–54 kHz are commonly used for NDE based on ultrasonic pulse wave data [36,37],
including studies for the detection of internal cracks due to corrosion [24]. The receiving
transducer placed on the opposite side of the concrete specimen measured the ultrasonic
pulse waves through the concrete. The received signal was digitized with a 10 MHz sam-
pling rate by the digital oscilloscope. Ultrasonic pulse wave measurements were con-
ducted on each specimen before and after the accelerated corrosion test. In this study,
ultrasonic tests were performed on the two test points on the surface of ultrasonic meas-
urements (see Figures 3a and 7b). Sensors on test point 1 were attached to the concrete
surface directly above the rebar, while sensors on test point 2 were located 50 mm hori-
zontally from the center to avoid the rebar. In this study, five measurements were re-
peated at each test point.

Figure 7. Test setup for ultrasonic

Figure pulse
7. Test setup wave measurements
for ultrasonic through the
pulse wave measurements concrete
through cube specimens:
the concrete cube specimens:
(a) source and receiver in theand
(a) source through-transmission configuration,
receiver in the through-transmission and (b) location
configuration, of test of
and (b) location points 1 1
test points
and 2 on a surface for
and 2 on a surface for ultrasonic measurements.ultrasonic measurements.

Figure 8a presents the typical P-wave signals measured from the concrete cube speci-
Figure 8a presents
mens, the
and typical P-wave
the enlarged signals
signals measured
are shown in Figurefrom thesignals
8b. The concrete
from cube
Figurespec-
8a,b were
imens, and the enlarged signals are shown in Figure 8b. The signals from Figure 8a,b were
processed in MATLAB using normalization by Z-scoring. The velocity of an ultrasonic
wave can be calculated by dividing the wave path by the travel time, as follows:
 Materials 2023, 16, 3502 9 of 26

processed in MATLAB using normalization by Z-scoring. The velocity of an ultrasonic

wave can be calculated by dividing the wave path by the travel time, as follows:

L
Vp = (5)
t a − td

where Vp is the wave propagation velocity, L is the distance between transducers (200 mm
in this study), ta is the initial wave arrival time, and td is the delay time computed during
probe calibration. When the two transducers were positioned opposite each other, the
time for the first arrival wave was recorded, and the delay time was calculated. It should
be noted that P-waves are potentially faster in time signals than any other refracted and
reflected waves from the boundary of concrete cube specimens. The arrival of transient
stress waves through cylinders was computed using the modified threshold approach
based on the observed ultrasonic signals. Using the conventional threshold method used
in earlier investigations, an estimated arrival time was initially obtained in this way. After
that, a precise arrival time was calculated by fitting a line to the signal data. The intersection
OR PEER REVIEW of the two P-wave travel times was then used to determine the P-wave travel time.
10 of 29 The
intersection of the fitting line and the measured zero-signal stage was used to determine
the P-wave travel time.

Figure 8. Typical timeFigure

signals of ultrasonic
8. Typical pulse
time signals waves propagating
of ultrasonic through the
pulse waves propagating D19 concrete
through cube cube
the D19 concrete
specimens under the specimens
four different
under target
the foursteel corrosion
different levels:
target steel (a) ultrasonic
corrosion waves atwaves
levels: (a) ultrasonic different
at different
target corrosion levels (θ = 0%, 5%, 10%, and 20%), and (b) enlarged ultrasonic signal around thethe first
target corrosion levels (θ = 0%, 5%, 10%, and 20%), and (b) enlarged ultrasonic signal around
arrival
first arrival of the wave shownof the
inwave
(a). shown in (a).
To analyze the nonlinear parameters of ultrasonic wave signals, this study used
To analyze thesignal
nonlinear parameters
coherence in the formofofultrasonic
magnitude wave
square signals,
coherencethis study
(MSC). The used sig- is
MSC function
nal coherence in the form of magnitude
calculated by square coherence (MSC). The MSC function is
calculated by S xy ( f )
γxy ( f ) = q (6)
Sxx ( f )Syy ( f )
𝑆 (𝑓)
𝛾 (𝑓 ) =
where γxy ( f ) is the coherence,𝑆Sxy(𝑓)𝑆 (6)Sxx ( f )
(𝑓)
( f ) is the cross-spectral density of x and y, and
and Syy ( f ) are the power spectral densities of x and y, respectively. The resultant value is
where 𝛾 (𝑓) is the acoherence,
number between𝑆 (𝑓0 )and 1.0, with
is the a value around
cross-spectral 1.0 indicating
density of x andhighy,signal 𝑆 (𝑓)
and coherence.
and 𝑆 (𝑓) are the power spectral densities of x and y, respectively. The resultant value
is a number between 0 and 1.0, with a value around 1.0 indicating high signal coherence.
As can be seen from the equation, assessing the coherence between concretes with
and without damage requires a baseline value. To obtain the coherence values in this
Materials 2023, 16, 3502 10 of 26

As can be seen from the equation, assessing the coherence between concretes with and
without damage requires a baseline value. To obtain the coherence values in this study,
signals were initially collected on specimens before and after the accelerated corrosion
process. A subset of the time-domain signals was then chosen, and each signal was
converted into power spectral density using fast Fourier transform (FFT). In this study, a
part of the tail end of the ultrasonic wave signals with a length of 0.1 ms was used for signal
consistency calculations. The coherence was computed from the converted signals using
MATLAB’s ‘mscohere’ function [38]. The time window of 4500 ns to 14,500 ns was chosen
for the average MSC analysis, which represents the tail end of the signal. The limited
length is based on studies related to coda wave interferometry, which typically uses a very
short window at 3 ms or lower [26,28]. Typical coherence curves are shown in Figure 9.
The ultrasonic signals from the different steel corrosion levels were each compared to the
signals from the solid conditions (0% vs. 3%, 0% vs. 6%, 0% vs. 12% for D10 and D13
specimens, and 0% vs. 5%, 0% vs. 10%, 0% vs. 20% for D19 and D22 specimens). The
coherence value was averaged within a certain frequency frame in this investigation so that
the outcome may be reported as a single number and analyzed with steel corrosion levels.
Two frequency ranges used for averaging MSC are shown in Figure 9. Ranges 1 and 2
Materials 2022, 15, x FOR PEER REVIEW 11 of 29
represents frequencies lower than the central frequency of the transducers and frequencies
adjacent (lower and higher) to the central frequency, respectively.

Figure 9.
Figure 9.Typical
Typicalsignal
signalcoherence
coherencecurves of of
curves ultrasonic pulse
ultrasonic waves
pulse measured
waves on the
measured onMIX1 concrete
the MIX1 spec-
concrete
imens with various corrosion levels: (a) D10 and D13 specimens, and (b) D19 and D22 specimens.
specimens with various corrosion levels: (a) D10 and D13 specimens, and (b) D19 and D22 specimens.

3. Recurrent Neural
3. Recurrent Neural Networks
Networks (RNNs)
(RNNs) of
of Ultrasonic
Ultrasonic Pulse
Pulse Waves
Waves
3.1. Architecture
3.1. Architecture
The general flowchart
The general flowchartofofRNNRNNdevelopment
development conducted
conducted in this
in this study
study is shown
is shown in
in Fig-
Figure 10. The RNN has the capability of analyzing time series data. This
ure 10. The RNN has the capability of analyzing time series data. This is because the RNNis because the
RNN
has thehas the “memory”
“memory” feature,feature, called hidden
called hidden state,
state, that that the
retains retains the information
information from thefrom
pre-
the previous time steps. A typical RNN algorithm works as shown in Figure 11 in the
vious time steps. A typical RNN algorithm works as shown in Figure 11 in the following
following steps [39]: (1) the first data point in the sequence, a0 , is delivered from the input
steps [39]: (1) the first data point in the sequence, a0, is delivered from the input layer to
layer to an RNN cell; (2) the cell adjusts the value of the a0 by a weighting function w,
an RNN cell; (2) the cell adjusts the value of the a0 by a weighting function w, resulting as
resulting as y1 ; (3) the second data in the sequence, x1 , is put into the hidden layer; (4) both
y1; (3) the second data in the sequence, x1, is put into the hidden layer; (4) both x1 and y1
x1 and y1 are put together into the layer and have both of their output readjusted by weight;
are put together into the layer and have both of their output readjusted by weight; and (5)
and (5) the process (2–4) is repeated until all of the data points within a sample have been
the process (2–4) is repeated until all of the data points within a sample have been pro-
processed in that layer and in the next hidden layers. The loss function, which is based on
cessed in that layer and in the next hidden layers. The loss function, which is based on the
the errors obtained from every calculation in each layer, is then used to update the w from
errors obtained from every calculation in each layer, is then used to update the w from
each hidden layer. The w, along with the algorithm, is updated from the last layer to the
each hidden layer. The w, along with the algorithm, is updated from the last layer to the
first. This process is called backpropagation [40] and is performed multiple times during
first.iteration
the This process is called
until the backpropagation
least error [40] and is performed multiple times during
value is obtained.
the iteration until the least error value is obtained.
Materials 2022,
Materials 2023,15,
16,x3502
FOR PEER REVIEW 11 of 26 12 of 2

Materials 2022, 15, x FOR PEER REVIEW 13 of 29

Figure10.
Figure 10.General
General flowchart
flowchart of RNN
of RNN model
model development.
development.

Figure 11.Arrangement
Figure11. Arrangementof
ofartificial
artificialneural
neuralnetwork
networkinindeep
deeplearning.
learning.
3.2. Long Short-Term Memory (LSTM)
3.2. Long Short-Term Memory (LSTM)
The primary form of a RNN might not give a satisfactory performance due to the
The primary form of a RNN might not give a satisfactory performance due to the
vanishing gradient problem caused by long iterations [41]. Alternative approaches such
vanishing gradient problem caused by long iterations [41]. Alternative approaches such
as long short-term memory (LSTM) and gated recurrent unit (GRU) have been used to
as long short-term memory (LSTM) and gated recurrent unit (GRU) have been used to
overcome the limitation of the RNN. The LSTM approach has internal gates to regulate
overcome the
information limitation
that should beofretained
the RNN. The LSTM[42].
or forgotten approach
Figure has internalthe
12a shows gates to regulate
architecture of
information that should be retained or forgotten [42].
an LSTM cell within an RNN and its working mechanism. Figure 12a shows the architecture of
an LSTM cell within an RNN and its working mechanism.
Materials 2022, 15, x FOR PEER REVIEW 1

Materials 2023, 16, 3502 12 of 26

Figure 12. Architecture of a cell in the modified RNN methods for: (a) LSTM, and (b) GRU.
Figure 12. Architecture of a cell in the modified RNN methods for: (a) LSTM, and (b) GRU.
In general, LSTM works by the following steps: (1) selecting information to be
In general,
discarded from the LSTM
current works
input,by (2)the following
selecting steps:to(1)beselecting
information kept in theinformation
cell state, to b
(3) updating
carded fromthethe oldcurrent
cell information,
input, (2)andselecting
(4) determining state features
information to be of the
keptoutput
in thecellcell stat
that should be retained. The work on the first step is conducted by the sigmoid unit that
updating the old cell information, and (4) determining state features of the output ce
acts as the forget gate. In the second step, the at−1 and xt are utilized to decide which
should be retained.
information needs to be The work through
updated on the first steplayer,
the tanh is conducted
which effectively by thebecomes
sigmoid theunit tha
as the gate.
input forget The gate.
at−1 In
andthe second
xt are step,tothe
then used at-1 andthe
go through xt are
tanhutilized to decide
layer to obtain which
the new cell inform
needs to be updated through
information candidate ct . The the tanh layer, which effectively becomes the input gate
third step updates the old information currently being kept
ain
t-1 the
andcell arect−then
xt or 1 based
used on the decision
to go throughmadetheby the
tanh forget
layer gate to(first
obtainstep) and
the decided
new cell inform
information candidate ct (second step). In the final step, the input yt−1 and xt are used by
candidate ct. The third step updates the old information currently being kept in the c
the sigmoid (output gate) to obtain a vector of values ranging from −1 to 1. These values
care
t-1 based on the decision made by the forget gate (first step) and decided inform
then multiplied by the weighting function from the output gate, resulting in the final
candidate
output. Thectoutput (second step).
of each gateIninthe
the final
LSTMstep, the input
is described yt-1 andshown
in equations xt are in
used
Tableby 2, the sig
(output fgate)
where w , w i , wto
o obtain
, and w y a vector of values ranging
are weighting functions and b f ifrom
, b , b o , and−1 to
b h 1. These values are then
are bias vectors. It is
also reported
tiplied by thethat having a bidirectional
weighting function from iteration of LSTMgate,
the output (or BiLSTM)
resulting can in
improve
the finalthe outpu
model performance significantly [43]. This bidirectional approach basically uses the same
output of each gate in the LSTM is described in equations shown in Table 2, where
principle as the conventional, forward-moving RNN, but with the process starting from the
wend
o, and wy are weighting functions and bf, bi, bo, and bh are bias vectors. It is also rep
part of the data in addition to the forward-moving RNN.
that having a bidirectional iteration of LSTM (or BiLSTM) can improve the model p
mance significantly [43]. This bidirectional approach basically uses the same princip
the conventional, forward-moving RNN, but with the process starting from the end
of the data in addition to the forward-moving RNN.

Table 2. Outputs in an LSTM and a GRU cell.

Model Gate Type Equation

Forget gate 𝑓 = 𝜎(𝑤 𝑎 , 𝑥 + 𝑏 )
Input gate 𝑖 = 𝜎(𝑤 𝑎 , 𝑥 + 𝑏 )
Output gate 𝑜 = 𝜎(𝑤 𝑎 , 𝑥 + 𝑏 )
LSTM
Hidden gate 𝑦 = 𝑡𝑎𝑛ℎ(𝑤 𝑎 , 𝑥 + 𝑏 )
𝑐 =𝑓𝑐 +𝑖 𝑦
Materials 2023, 16, 3502 13 of 26

Table 2. Outputs in an LSTM and a GRU cell.

Model Gate Type Equation

 
Forget gate f t = σ w f [ a t −1 , x t ] + b f
Input gate i t = σ ( w i [ a t − 1 , x t ] + bi )
LSTM Output gate o t = σ ( w o [ a t − 1 , x t ] + bo )

Hidden gate yt = tanh wc [ at−1 , xt ] + by
c t = f t c t −1 + i t y t
Final Output
at = ot tanh ct
Update gate z t = σ ( w z x t + v z a t − 1 + bz )
Reset gate rt = σ(wr xt + vr at−1 + br )

GRU
Hidden state yt = tanh wc [ at−1 , xt ] + by
Final output at = zt .at−1 + (1 − zt )yt

3.3. Gated Recurrent Unit (GRU)

Another widely used RNN method is the gated recurrent unit (GRU). Originally
introduced by Cho et al. [44], GRU has recently been adopted by various researchers with
a performance comparable to LSTM and, in some cases, outperforms LSTM [43,45]. For
the same set of hyperparameters, the GRU network can also be trained faster since the
architecture is simpler than LSTM: it has no cell state ct and instead uses the hidden state yt
to transfer the information to the next cell. Furthermore, GRU only has two gates which
are called the update gate and reset gate, as shown in Figure 12b. The equations in the
GRU model are shown in Table 2, where w and v are the weight for xt and at −1 , respectively.
With this simpler process, the GRU process takes less memory to compute and can be
trained faster.

3.4. Data Preparation

The collected data was preprocessed to start with zero-mean, which improves learning
efficiency of the RNN algorithms. The data was then prepared as two types: raw signal
in the form of time series and extracted features in a systematic way. Both types of data
were then downscaled into several sampling frequencies, which will be discussed further
in Section 4.2, to further improve the training efficiency as well as to reduce the signal
noises caused by a training process. The raw signal has a resolution equal to the sampling
rate of the oscilloscope, which is 10 MHz. Considering that the central frequency of the
transducers is 50 kHz, the maximum data resolution was limited to only 1 MHz to make the
training process more efficient while still allowing higher frequency bands to be analyzed.
In this study, time series data were compressed by systematic feature extraction
processes to further improve computational efficiency. For these purposes, standardized
spectral entropy and instantaneous frequency were calculated to extract both time and
frequency information from time series, following prior research works [46,47]. Spectral
entropy is based on the equations of the power spectrum and probability distribution of
signals. The power spectrum of the mth signal (or x(m)) is denoted by S(m), where m is the
index of frequency point. S(m) can be obtained by squaring the magnitude of its discrete
Fourier transform X(m). Mathematically, the fundamental equations are as follows:

M
H=− ∑ R(m)log2 R(m) (7)
m =1

S(m)
R(m) = (8)
∑iMS (i )
=1

S(m) = | X (m)|2 (9)

where H is the spectral entropy, R(m) is the probability distribution, i is frequency index in
the calculation of R(m), and M is the total frequency points. The second feature is instanta-
Materials 2023, 16, 3502 14 of 26

neous frequency, which is a measure of change in the time parameter of a nonstationary

signal associated with the average of frequencies as the signal changes. The fundamental
equation of the instantaneous frequency f inst is as follows:

∑iN=1 f i P(ti , f i )
f inst (t) = (10)
∑iN=1 P(ti , f i )
Materials 2022, 15, x FOR PEER REVIEW 16 of 29
Here, f inst is a summation of value from i = 1 to the N, where i is a step and N is the
end step of the analyzed time-frequency window within a signal. Every step corresponds to
the time step interval and its corresponding frequency measurement. P is the spectrogram
spectrogram
power power
spectrum, t isspectrum,
time, andtfisistime, and f isoffrequency
frequency the signalofinput.
the signal
Moreinput. More
details details
regarding
regarding instantaneous frequency can be found in the research by
instantaneous frequency can be found in the research by Boashash [48]. Both spectralBoashash [48]. Both
spectraland
entropy entropy and instantaneous
instantaneous frequency frequency are standardized,
are standardized, which furtherwhich further
improves theimproves
learning
the learning
efficiency for efficiency for neural
neural network network
training training
[46]. The [46]. The standardization
standardization is based
is based on Z-scoring on
with
Z-scoring
the with
following the following
equation appliedequation applied
to each data pointto in
each
thedata point in the dataset:
dataset:
𝑧 x=− µ (11)
Z= (11)
σ
where z is the new value on a data point after the standardization, x is the existing value
where z ispoint,
of a data the new value
μ is on a data
the mean point after
of dataset, andthe σ isstandardization, x is the existing
the standard deviation. Figurevalue of a
13 shows
data point, is the mean of dataset, and is the standard deviation. Figure 13 shows
typical results from feature extraction. The feature extraction reduces the number of sam-
µ σ typical
results from
ples from feature
10,000 extraction.
points in timeThe feature
series to 129extraction reduces
points, which the numberreduces
significantly of samples from
computa-
10,000
tional points
cost. in time series to 129 points, which significantly reduces computational cost.

Figure13.
Figure 13.Typical
Typicalfeature
featureextraction
extractionresults
results for:
for: (a,b)
(a,b) normalized
normalized spectral
spectral entropy,
entropy, andand
(c,d)(c,d) nor-
normal-
malized instantaneous frequency of ultrasonic pulse waves measured from the concrete
ized instantaneous frequency of ultrasonic pulse waves measured from the concrete specimens with specimens
with various
various corrosion
corrosion levels.levels. Leftright
Left and and right columns
columns represent
represent the results
the results from from the concrete
the concrete speci-
specimens
mens belonging to the MIX1 and D10, and the MIX1 and D19 groups, respectively.
belonging to the MIX1 and D10, and the MIX1 and D19 groups, respectively.

3.5. Bilinear Classification Model

A bilinear classification model was developed to divide the dataset into two classes
based on the steel corrosion level, θ, with a threshold of 3%. A part of the dataset corre-
sponding to θ < 3% was classified as Class 1, which represents the solid to initial corrosion
level. The remaining dataset (θ ≥ 3%) was classified as Class 2, which represents a medium
Materials 2023, 16, 3502 15 of 26

3.5. Bilinear Classification Model

A bilinear classification model was developed to divide the dataset into two classes
based on the steel corrosion level, θ, with a threshold of 3%. A part of the dataset corre-
sponding to θ < 3% was classified as Class 1, which represents the solid to initial corrosion
level. The remaining dataset (θ ≥ 3%) was classified as Class 2, which represents a medium
to high corrosion level. A preliminary numerical experiment revealed that the effects of
concrete mixture proportion and rebar size were not considerable on the performance of
the classification model. In this study, ultrasonic pulse data were classified only based
on steel corrosion levels (or concrete early damage condition). The more general model
would be more effective as an in situ NDE method because specific material information
is not always given or reliable in field surveys. The threshold level of the classification
model was determined based on the initial steel corrosion level that starts to cause concrete
deterioration (enhanced porosity and microcracking in concrete). In this study, it was
observed that the relative mass of concrete specimens, which is the difference between
damaged and solid concrete specimens, gradually increased as the amount of impressed
current increased in the accelerated corrosion test. Based on that observation, it can be
inferred that the expansion of rust products causes some internal concrete defects, which
resulted in changes in the porosity of concrete. Surface-breaking cracks typically were
observed at corrosion levels between 4% and 5%. Therefore, it is reasonable to say that a
corrosion level of 3% can be selected as a threshold to detect early damage caused by steel
corrosion in concrete. Based on the classification, there are a total of 222 data in Class 1 and
164 data in Class 2. RNN models were trained using 80% of data, randomly selected, from
each class. The remaining 20% of the data for that class were used to test the trained model.

3.6. Performance Evaluation

The accuracy and Cohen’s kappa (later referred as kappa) were used to compare the
classification performance of each method. The accuracy is a ratio between the true positive
(TP) and true negative (TN) classifications given by the method to the total number of
predictions [49]. Therefore, high numbers of false-positive (FP) and false-negative (FN)
classifications will reduce network accuracy. However, accuracy alone is not enough to
demonstrate a method’s performance. Researchers have incorporated kappa to evaluate
prediction models [50,51], in addition to the accuracy. The kappa is a chance-corrected
method for assessing agreement among raters [52]. The equations to calculate these param-
eters for this study are as follows:

TP + TN
Accuracy = , and (12)
TP + TN + FP + FN

Accuracy + pe
Cohen0 s Kappa = . (13)
1 − pe
where pe is the rate of agreement between the prediction and actual class value by chance,
which is calculated using following equation,

(( TN + FN ) × ( TN + FP)) + (( FP + TP) × ( FN + TP))

pe = . (14)
( TP + TN + FP + FN )

with Kappa ranges from −1 to 1, with the value below 0.4 regarded as low agreement, 0.41
to 0.60 regarded as moderate agreement, 0.61 to 0.80 as substantial agreement, and 0.81 to
0.99 as near-perfect agreement [53].

4. Results and Discussion

4.1. Bilinear Classification Models Based on Conventional Ultrasonic Testing Parameters
Bilinear classification models were developed based on the two conventional ultrasonic
testing parameters, which will be compared with deep learning classification models in
Materials 2023, 16, 3502 16 of 26

Section 4.3. In this study, two conventional ultrasonic testing parameters (relative P-wave
velocity and signal consistency) were reduced from the ultrasonic pulse waves collected
from the concrete specimens with different steel corrosion levels, θ. The two reduced
ultrasonic testing parameters were divided into two classes according to the steel corrosion
level, with a threshold of 3%, consistent with the definition described in Section 3.5 for the
deep learning classification models.
Figure 14 shows the distribution of the relative P-wave velocity (i.e., the P-wave
velocity of damaged concrete after the accelerated corrosion process normalized by the
P-wave velocity of solid concrete) with actual steel corrosion levels. Overall, the effect
of the steel corrosion levels cannot be clearly seen because of a scattering of the relative
Materials 2022, 15, x FOR PEER REVIEW 18 of 29
P-wave velocity with steel corrosion levels from 0% to 20%. In this study, a linear equation
was used to relate the relative P-wave velocity and the steel corrosion levels as follows:

Vr,P = 0.0007θ + 1.0006, with R2 = 0.0008 (15)

𝑉 , = 0.0007𝜃 + 1.0006, 𝑤𝑖𝑡ℎ 𝑅 = 0.0008 (15)
whereVVr,P
where r,P is the relative P-wave velocity ofof concrete, and θ𝜃 isisthe
concrete, and theactual
actual(measured)
(measured)steel
steel
2
mass
massloss
loss(or
(orcorrosion
corrosionlevel).
level).The
Thebest-fit
best-fitline
linelooks
looksalmost
almostflat flatwith
withparticularly lowRR2,,
particularlylow
which
whichreveals
revealsthatthatthe
therelative
relativeP-wave
P-wavevelocity
velocity might
might not
not be
bean aneffective
effectiveparameter
parameter for
for
evaluating concrete damages caused by corrosion of steel in concrete.
evaluating concrete damages caused by corrosion of steel in concrete.

Figure14.
Figure 14.Classification
Classificationbased
basedon
onthe
therelative
relativeP-wave
P-wavevelocity
velocitydata.
data.

Figure
Figure 15a,b
15a,b show
show the
the distribution
distribution of
of average
average magnitude
magnitude square
square coherence
coherence (MSC)
(MSC)
values
valuesfor
forthe
thetwo
twodifferent
differentfrequency
frequencyranges:
ranges: Range
Range 11 and
and 2,2, respectively.
respectively.Average
AverageMSCMSC
shows a descending trend with respect to the steel mass loss. It can be seen
shows a descending trend with respect to the steel mass loss. It can be seen that average that average
MSC
MSCisisaabetter
betterindicator
indicatorfor
forclassifying
classifyingearly
earlyconcrete
concretedamage
damagecausedcausedby bysteel
steelcorrosion
corrosion
2
than VVr,P
than Using linear
r,P. Using regression, the obtained R
linear regression, R were 0.18 and 0.17
2 0.17 for
for Range
Range 11 and
and
Range
Range2,2,respectively.
respectively.The
Theregression
regression equations
equations forfor
Range
Range1 and
1 andRange
Range2 were obtained
2 were as
obtained
follows:
as follows:
MSC1 = −0.0166θ + 0.4523 for Range 1 (16)
𝑀𝑆𝐶1 = −0.0166𝜃 + 0.4523 for Range 1 (16)

𝑀𝑆𝐶2==−−0.0133𝜃
MSC2 0.0133θ + +
0.3983 forfor
0.3983 Range 2 2
Range (17)
(17)
Table 3 summarizes the threshold values for Vr,P, and average MSCs was determined
by the linear regression equations (Equations (15)–(17)), corresponding to the steel corro-
sion level of 3%. The confusion matrix based on the classification models for the two con-
ventional ultrasonic testing parameters is summarized in Table 4. Accuracy and Cohen’s
kappa values were calculated using Equations (12)–(14). As expected, the performance of
the classification model based on relative P-wave velocity, Vr,p, was not satisfactory, with
accuracy of 53% and kappa of 0.07. Those values are as low as those predicted by chance
alone. Conversely, the classification models based on MSC resulted in far more improved
accuracies greater than 70%. The improved accuracy demonstrates that ultrasonic pulse
waves can be used for early detection of concrete damages caused by steel corrosion. It
can be inferred that incoherence parts of ultrasonic pulse waves are more sensitive to the
Materials2023,
Materials 2022,16,
15,3502
x FOR PEER REVIEW 19of
17 of26
29

Figure15.
Figure 15.Classification
Classificationbased
basedononthe
the average
average MSC
MSC method
method using
using twotwo different
different frequency
frequency ranges:
ranges: (a)
(a) Range 1, and (b) Range 2 (see Figure
Range 1, and (b) Range 2 (see Figure 9). 9).

Table 3. Parameter
Table thresholds
3 summarizes the for Vr,p and values
threshold MSC methods.
for Vr,P , and average MSCs was determined
by the linear regression equations (Equations (15)–(17)), corresponding to the steel cor-
Class Steel Corrosion Loss Vr,p MSC1 MSC2
rosion level of 3%. The confusion matrix based on the classification models for the two
Class 1 (N) <3% Vr,p < 1.0027 MSC > 0.40 MSC > 0.36
conventional ultrasonic testing parameters is summarized in Table 4. Accuracy and Cohen’s
Class 2 (P) ≥3% values were calculated
kappa Vr,p ≥ 1.0027 MSC ≤ 0.40
using Equations (12)–(14). MSC
As expected, the ≤ 0.36
performance of
the classification model based on relative P-wave velocity, Vr,p , was not satisfactory, with
Table 4. Confusion
accuracy of 53% andmatrix
kappafor of
classification
0.07. Those based on Vare
values r,p and MSC methods.
as low as those predicted by chance
alone. Conversely, the classification models based
Predicted on MSC resulted in far more improved
accuracies greater than 70%. The improved accuracy demonstrates that ultrasonic pulse
Vr,p MSC1 MSC2
waves can be used for early detection of concrete damages caused by steel corrosion. It
P N P N P N
can be inferred that incoherence parts of ultrasonic pulse waves are more sensitive to the
P 95 (45.2%) 91 (43.4%)
internal concrete damages119 (62.9%)
by progression of45 (23.8%)
steel 123 (65.1%)
corrosion than 41 (21.7%)
coherence parts. However,
Actual
N 8 (3.8%) 16 (7.6%) 10 (5.3%) 15 (7.9%) 10
it should be noted that the performance of the MSC-based models, like other (5.3%) 15 (7.9%)
nonlinear
Accuracy 52.8%parameters, is strongly affected
ultrasonic 71.0%by engineering + judgment 73.0% on the selection of
Cohen’s kappa 7%
input signals 20.0% For example, two different
and signal processing processes. 22.9%
frequency ranges
(Ranges 1 and 2) in this study resulted in different performances (Table 3). Therefore, more
systematic approaches
4.2. Deep Learning are needed
Classification Model to optimize the capabilities of ultrasonic pulse wave
measurements
A series offor conditionexperiments
numerical assessment of wasconcrete.
performed to investigate the variation of the
performance of deep learning classification models, with varying inputs and hyperparam-
Table 3. Parameter thresholds for Vr,p and MSC methods.
eters for RNN models. As such, an optimal set of parameters was determined, which re-
sulted in the best performance of the deep learning classification model. Table 5 summa-
Steel Corrosion
Class Vr,p MSC1 MSC2
Lossand ranges of each
rizes the critical parameters parameter considered in this study.
Class 1 (N) <3% Vr,p < 1.0027 MSC > 0.40 MSC > 0.36
Class 2 (P) ≥3% Vr,p ≥ 1.0027 MSC ≤ 0.40 MSC ≤ 0.36
Materials 2023, 16, 3502 18 of 26

Table 4. Confusion matrix for classification based on V r,p and MSC methods.

Predicted
Vr,p MSC1 MSC2
P N P N P N
P 95 (45.2%) 91 (43.4%) 119 (62.9%) 45 (23.8%) 123 (65.1%) 41 (21.7%)
Actual
N 8 (3.8%) 16 (7.6%) 10 (5.3%) 15 (7.9%) 10 (5.3%) 15 (7.9%)
Accuracy 52.8% 71.0% 73.0%
Cohen’s kappa 7% 20.0% 22.9%

4.2. Deep Learning Classification Model

A series of numerical experiments was performed to investigate the variation of the
performance of deep learning classification models, with varying inputs and hyperparame-
ters for RNN models. As such, an optimal set of parameters was determined, which resulted
in the best performance of the deep learning classification model. Table 5 summarizes the
critical parameters and ranges of each parameter considered in this study.

Table 5. Summary of critical parameters as variables for training deep learning models of ultrasonic
pulse waves in this study.

Parameter Range
Length of time signal 1 ms, 2 ms, and 5 ms
Sampling frequency of time signal 100 kHz, 250 kHz, 500 kHz, and 1000 kHz
Time series (TS), instantaneous frequency
Type of input (IF), spectral entropy (SE), and
combination of IF and SE (IFSE)
Network LSTM, BiLSTM, and GRU
Batch size 1, 2, 4, 8, and 16
Hyperparameter
Epoch 100, 200, 300, 400, and 500

4.2.1. Effects of Input Type

To observe the effect of input type on the model performance, a series of trainings
using different types of input has been conducted. The input consists of (1) time series (TS),
which is the raw signal in the time domain, (2) spectral entropy (SE), and (3) a combination
of instantaneous frequency and spectral entropy (IFSE). Figure 16 shows the variations of
accuracy of the deep learning classification models with these inputs based on the three
RNN models (i.e., LSTM, BiLSTM, and GRU). While not shown in this article, it is found
that the use of IF as an input in the RNN models resulted in similar trends to those based
on SE. The top, middle, and bottom rows of the figure are results for LSTM, BiLSTM, and
GRU, respectively. The left, middle, and right columns of the figure are the results using
time series data with lengths of 1 ms, 2 ms, and 5 ms, respectively. Training of the RNN
models was performed with fixed hyperparameters, epoch of 100 and batch size of 16.
Hyperparameters will be discussed in more detail in Section 4.2.3. In addition, the Adam
optimizer and a network size of 100 were used as default values in the deep learning
toolbox in MATLAB.
Overall, the use of SE and IFSE resulted in comparable accuracy levels. It was also
observed that the use of IFSE for fixed networks and length of time resulted in slightly
better accuracy values compared to the models based on SE. The best accuracy was 69%,
which was obtained based on SE of ultrasonic pulse waves with a length of 5 ms and a
sampling frequency of 1000 kHz, as a base input of the deep learning model. The accuracy
based on IFSE was 68% for the same test setup (see Figure 16f). It can be noticed that the
accuracy of the deep learning classification models based on an input of time series (TS)
was less than that of using spectral entropy (SE) or the combination of spectral entropy and
Materials 2023, 16, 3502 19 of 26

instantaneous frequency (IFSE). The best accuracy based on TS was only 56% for the LSTM
model using an input of TS with a length of 1 ms and a sampling frequency of 1000 kHz.
Materials 2022, 15, x FOR PEER REVIEW 21 of 29
For the same set of parameters, the use of SE and IFSE resulted in more improved accuracy
of 63% and 68%, respectively.

Accuracy

Accuracy
Accuracy

Accuracy

Accuracy
Accuracy

Accuracy

Accuracy
Accuracy

Figure
Figure 16.
16. Variation
Variation of
of accuracy
accuracy of
of the
the deep
deep learning
learning classification
classification models
models based
based on
on three
three input
input data
data
types (time series, TS; spectral entropy, SE; and combination of instantaneous frequency and spec-
types (time series, TS; spectral entropy, SE; and combination of instantaneous frequency and spectral
tral entropy, IFSE). Top, middle, and bottom rows represent the results based on the three RNN
entropy, IFSE). Top, middle, and bottom rows represent the results based on the three RNN models:
models: LSTM, BiLSTM, and GRU, respectively. Left, middle, and right columns represent the re-
LSTM, BiLSTM, and GRU, respectively. Left, middle, and right columns represent the results based
sults based on three different lengths of ultrasonic pulse waves: 1 ms, 2 ms, and 5 ms, respectively.
on three different lengths of ultrasonic pulse waves: 1 ms, 2 ms, and 5 ms, respectively.
4.2.2. Effects of Input Data
Care is needed when using the time series data as an input of RNN models of ultrasonic
pulse waves17
Figures inand 18 show
concrete. the variations
Ultrasonic pulseofwave
the accuracy of the deep
data collected in thislearning
studyclassifica-
involved
tion models with increasing sampling frequency of ultrasonic pulse
relatively high variabilities in the amplitude and first arrival times. The variability waves used for theof
calculation of input
time series data coulddata
be for RNNby
caused models.
variousThese figures
sources represent the
of experimental results basedsuch
uncertainties, on SEas
and IFSE extracted
inconsistent couplingfrom ultrasonic
conditions, pulse
surface wave data
roughness with different
of concrete lengths (1 ms,
and heterogeneous 2 ms,
features
and 5 ms). Training
of concrete, which areofnot thereally
models was to
related performed withoffixed
the severities hyperparameters,
concrete damages. RNN epoch
modelsof
100
are and batch
known to size of 16. The Adam
be especially optimizer
sensitive and network
to the sequence sizedata
of the of 100 were used
points. as default
Consequently,
values in toolbox
experimental in MATLAB.
uncertainties Overall, it
in ultrasonic was noticed
waves could havethatmore
the best accuracy
impact on theofperformance
each model
was obtained at a sampling frequency lower than 1.0 MHz. This
of the RNN-based classification models [54]. Moreover, RNN models trained using is reasonable since the
use
timeofseries
50 kHz
datatransducers in this study resulted
have high computational in an effective
costs. Therefore, it is notbandwidth
recommended with toa fre-
use
quency range
time series dataofas100ankHz
input tofor
1000 kHz. The deep
RNN-based best accuracy
learning in wasthisobserved
study. at 69% by using
SE of ultrasonic pulse waves with a length of 5 ms and a sampling frequency of 500 kHz
as an input of BiLSTM (see Figure 17c).
 Materials 2023, 16, 3502 20 of 26

4.2.2. Effects of Input Data

Figures 17 and 18 show the variations of the accuracy of the deep learning classifica-
tion models with increasing sampling frequency of ultrasonic pulse waves used for the
calculation of input data for RNN models. These figures represent the results based on SE
and IFSE extracted from ultrasonic pulse wave data with different lengths (1 ms, 2 ms, and
5 ms). Training of the models was performed with fixed hyperparameters, epoch of 100 and
batch size of 16. The Adam optimizer and network size of 100 were used as default values
in toolbox in MATLAB. Overall, it was noticed that the best accuracy of each model was
obtained at a sampling frequency lower than 1.0 MHz. This is reasonable since the use of
50 kHz transducers in this study resulted in an effective bandwidth with a frequency range
terials 2022, 15, x FOR PEER REVIEW of 100 kHz to 1000 kHz. The best accuracy was observed at 69% by using SE of ultrasonic 22 o
pulse waves with a length of 5 ms and a sampling frequency of 500 kHz as an input of
BiLSTM (see Figure 17c).

Using SE of me series with a length of 1 ms Using SE of me series with a length of 2 ms Using SE of me series with a length of 5 ms
(a) 1 (b) 1 (c) 1
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 LSTM 0.2 LSTM
0.2 LSTM
0.1 BiLSTM BiLSTM 0.1 BiLSTM
0.1
GRU GRU GRU
0 0 0
0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000
Frequency [kHz] Frequency [kHz] Frequency [kHz]
Using IFSE of me series with a length of 1 ms Using IFSE of me series with a length of 2 ms Using IFSE of me series with a length of 5 ms
(d) 1 (e) 1 (f) 1
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 LSTM 0.2 LSTM
0.2 LSTM
0.1 BiLSTM BiLSTM 0.1 BiLSTM
0.1
GRU GRU GRU
0 0 0
0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000
Frequency [kHz] Frequency [kHz] Frequency [kHz]

Figure 17. Variation

Figure of accuracy
17. Variation ofofthe
of accuracy thedeep
deeplearning classification
learning classification models
models based
based withwith various sa
various
pling sampling
frequencies for calculation
frequencies of inputs.
for calculation Top
of inputs. Topand
and bottom
bottom rowsrows represent
represent the results
the results based on
based on
SE andtheIFSE,
SE andrespectively. Left,
IFSE, respectively. middle,
Left, middle,and rightcolumns
and right columns represent
represent the based
the results results
on based on the th
the three
different lengths of ultrasonic pulse waves: 1 ms, 2 ms, and 5 ms, respectively.
different lengths of ultrasonic pulse waves: 1 ms, 2 ms, and 5 ms, respectively.
In addition to sampling frequency, the variation of signal length was also considered.
The networks
In addition to were trained with
sampling 5 ms (fullthe
frequency, length), 2 ms, and
variation 1 ms. Itlength
of signal is noticed
wasthat
alsoeach
consider
network behaves differently to the input data properties. The LSTM network performed
The networks were trained with 5 ms (full length), 2 ms, and 1 ms. It is noticed that ea
better using shorter sample lengths, with the best accuracy of 68%. The accuracy of LSTM
network
modelsbehaves differently
increases from 54% toto65%
theasinput dataofproperties.
the length The from
signals decreases LSTM network
5 ms to 2 ms. Itperform
bettercan
using
be inferred that the LSTM is more sensitive to the coherent part of the signal,accuracy
shorter sample lengths, with the best accuracy of 68%. The which is of LST
models increases
mostly located from 54% part
in the early to 65% astime
of the the series.
lengthOn ofthe
signals decreases
other hand, from 5ofms
the accuracy the to 2 ms
model based on GRU remained stagnant without regard for the length of
can be inferred that the LSTM is more sensitive to the coherent part of the signal, whichtime series, with
the best accuracy in a range of 60% to 64%. The most noticeable improvement was found
mostly located in the early part of the time series. On the other hand, the accuracy of t
with the use of BiLSTM: the longer time series resulted in greater accuracy. The accuracy of
model thebased
BiLSTM onclassification
GRU remained modelsstagnant
increased without
from 62% regard
to 69% asfor
thethe length
length of time
of times series, w
signals
the best
usedaccuracy in a of
for calculation range of 60% tofrom
IFSE increased 64%. The
1 ms to 5most
ms. noticeable improvement was fou
with the use of BiLSTM: the longer time series resulted in greater accuracy. The accura
of the BiLSTM classification models increased from 62% to 69% as the length of tim
signals used for calculation of IFSE increased from 1 ms to 5 ms.
Materials 2022, 15, x FOR PEER REVIEW 23 of 29
Materials 2023, 16, 3502 21 of 26

SE as an input of LSTM SE as an input of BiLSTM SE as an input of GRU

(a) 1 (b) 1 (c) 1
100 kHz 500 kHz 5000 kHz 100 kHz 500 kHz 5000 kHz 100 kHz 500 kHz 5000 kHz
0.9 250 kHz 1,000 kHz 10,000 kHz 0.9 250 kHz 1,000 kHz 10,000 kHz 0.9 250 kHz 1,000 kHz 10,000 kHz
0.8 0.8 0.8
0.7 0.7 0.7
Accuracy

Accuracy

Accuracy
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
1 ms 2 ms 5 ms 1 ms 2 ms 5 ms 1 ms 2 ms 5 ms
Length of me series Length of me series Length of me series
IFSE as an input of LSTM IFSE as an input of BiLSTM IFSE as an input of GRU
(d) 1 (e) 1 (f) 1
100 kHz 500 kHz 5000 kHz 100 kHz 500 kHz 5000 kHz 100 kHz 500 kHz 5000 kHz
0.9 250 kHz 1,000 kHz 10,000 kHz 0.9 250 kHz 1,000 kHz 10,000 kHz 0.9 250 kHz 1,000 kHz 10,000 kHz
0.8 0.8 0.8
0.7 0.7 0.7
Accuracy

Accuracy

Accuracy
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
1 ms 2 ms 5 ms 1 ms 2 ms 5 ms 1 ms 2 ms 5 ms
Length of me series Length of me series Length of me series

Figure 18. Variation

Figure 18. Variationofofaccuracy
accuracy of thedeep
of the deeplearning
learning classification
classification models
models basedbased
on threeondifferent
three different
lengths of time
lengths series
of time forforcalculation
series of inputs.
calculation of inputs.Top
Topandand bottom
bottom rowsrows represent
represent the results
the results based onbased on
the SE
theand IFSE,
SE and IFSE,respectively. Left,
respectively. Left, middle,
middle, and and
right right
columnscolumns represent
represent the resultsthe results
based on the based
three on the
threedifferent RNN
different RNNnetworks:
networks: LSTM, BiLSTM,
LSTM, and GRU,
BiLSTM, and respectively.
GRU, respectively.
4.2.3. Effects of Hyperparameter
4.2.3. Effects of Hyperparameter
This section discussed the effect of the hyperparameter setup, particularly related
toThis
the section
number discussed
of epochs and thebatch
effectsizes,
of the onhyperparameter
the accuracy of thesetup, particularly
deep learning related to
models.
Deep learning
the number models,
of epochs including
and batch RNNs,
sizes, oncanthebe run with virtually
accuracy of thelimitless combinations
deep learning of Deep
models.
hyperparameters. Models that are undertrained often have the characteristics
learning models, including RNNs, can be run with virtually limitless combinations of hy- of unbalanced
true predictions between each class, which translates to low accuracy. On the other hand,
perparameters. Models that are undertrained often have the characteristics of unbalanced
models that are overtrained often cannot recognize the dataset outside of the training
truedataset,
predictions leadingbetween each class,
to overfitting. In an which
RNN, an translates to low
epoch refers to aaccuracy.
single passOn the other
through the hand,
models
entirethat are overtrained
dataset during training. often cannot
During each recognize
epoch, thethe dataset
model’s outside are
parameters of the training da-
updated
taset,based
leading
on the toerrors
overfitting. In an RNN,
made in predicting the an epoch
output refers
for each to a single
example in the pass through
dataset. The batchthe entire
size refers to the number of sequences that are processed simultaneously
dataset during training. During each epoch, the model’s parameters are updated based on by the network
during made
the errors training.inThe batch sizethe
predicting is adjusted
outputtofor balance
each the trade-off
example inbetween computational
the dataset. The batch size
efficiency (training time) and the ability to estimate the true gradient of the loss function.
refers to the number of sequences that are processed simultaneously by the network dur-
Figure 19 shows the variation of accuracy of the deep learning classification model
ing training. The batch
with five different size(100,
epochs is adjusted to balance
200, 300, 400, and 500)the andtrade-off between
four different batch computational
sizes (2,
efficiency (training time) and the ability to estimate the true
4, 8, and 16). All results in the figure were calculated from BiLSTM based on SE or gradient of the loss
IFSEfunction.
asFigure
an input. 19 shows the variation
SE and IFSE of accuracy
were calculated based onofthe thetime
deep learning
series classification
with a length of 5 ms. model
with five different epochs (100, 200, 300, 400, and 500) and four different batchkHz)
The results from three different sampling frequencies (250 kHz, 500 kHz, and 1000 sizes (2, 4,
are shown in each figure. Overall, it was observed that the IFSE input resulted in slightly
8, and 16). All results in the figure were calculated from BiLSTM based on SE or IFSE as
higher accuracy than the models trained only with SE. For most results from the use of
an input. SE and IFSE were calculated based on the time series with a length of 5 ms. The
SE and IFSE, accuracy of the RNN model changes only little with the various epochs and
results
batchfromsizesthree different
considered in thissampling frequencies
study. However, (250 kHz,
it was clearly 500 kHz,
noticeable that theand 1000 of
accuracy kHz) are
shownthe classification model could be enhanced by tunning the hyperparameters. For example,slightly
in each figure. Overall, it was observed that the IFSE input resulted in
higher
the accuracy than the
use of SE results models
in best accuracytrained
of 71%only with
for the SE. Forsetup
parameter mostofresults
samplingfrom the use of SE
frequency
of 1000 kHz with epoch of 200 and batch size of 8. Furthermore,
and IFSE, accuracy of the RNN model changes only little with the various epochs the use of IFSE results in and
slightly higher best accuracy of 74% for the parameter setup of sampling frequency of 250
batch sizes considered in this study. However, it was clearly noticeable that the accuracy
kHz with epochs of 500 and batch size of 16. The accuracy of 74% on that setup was the
of the classification
best-attained model
accuracy could
in our beThis
tests. enhanced
reveals thatby tunning the hyperparameters.
the deep learning classification model For exam-
ple, the use of SE results in best accuracy of 71% for the parameter setup of sampling
frequency of 1000 kHz with epoch of 200 and batch size of 8. Furthermore, the use of IFSE
results in slightly higher best accuracy of 74% for the parameter setup of sampling fre-
quency of 250 kHz with epochs of 500 and batch size of 16. The accuracy of 74% on that
setup was the best-attained accuracy in our tests. This reveals that the deep learning clas-
Materials 2023, 16, 3502 22 of 26

Materials 2022, 15, x FOR PEER REVIEW 24 of 29

based on BiLSTM can improve the accuracy of the classification up to 150% more than the
estimation by chance alone.

SE as an input of BiLSTM with a batch size of 2 SE as an input of BiLSTM with a batch size of 4 SE as an input of BiLSTM with a batch size of 8 SE as an input of BiLSTM with a batch size of 16
(a) 1 (b) 1 (c) 1 (d) 1

0.9 0.9 0.9 0.9

0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3

fs = 200 kHz fs = 200 kHz fs = 200 kHz fs = 200 kHz
0.2 0.2 0.2 0.2
fs = 500 kHz fs = 500 kHz fs = 500 kHz fs = 500 kHz
0.1 0.1 0.1 0.1
fs = 1,000 kHz fs = 1,000 kHz fs = 1,000 kHz fs = 1,000 kHz
0 0 0 0
50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550
Epoch Epoch Epoch Epoch
(e) IFSE as an input of BiLSTM with a batch size of 2 IFSE as an input of BiLSTM with a batch size of 4 IFSE as an input of BiLSTM with a batch size of 8 IFSE as an input of BiLSTM with a batch size of 16
1 (f) 1 (g) 1 (h) 1

0.9 0.9 0.9 0.9

0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3

fs = 200 kHz fs = 200 kHz fs = 200 kHz fs = 200 kHz
0.2 0.2 0.2 0.2
fs = 500 kHz fs = 500 kHz fs = 500 kHz fs = 500 kHz
0.1 0.1 0.1 0.1
fs = 1,000 kHz fs = 1,000 kHz fs = 1,000 kHz fs = 1,000 kHz
0 0 0 0
50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550 50 100 150 200 250 300 350 400 450 500 550
Epoch Epoch Epoch Epoch

Figure 19.
Figure 19. Variations
Variationsofofaccuracy
accuracywith epochs
with epochsin in
a range
a rangeof 100 to 500
of 100 to with various
500 with batchbatch
various sizes and
sizes
two different types of input for BiLSTM. The upper and lower rows indicate the results based on
and two different types of input for BiLSTM. The upper and lower rows indicate the results based
spectral entropy (SE) and combination of instantaneous frequency and spectral entropy (IFSE), re-
on spectral entropy (SE) and combination of instantaneous frequency and spectral entropy (IFSE),
spectively. The first to fourth columns from the left side indicate the results determined by using
respectively.
batch sizes ofThe
2, 4,first to fourth
8, and columns from
16, respectively. the leftfrom
The results sidethree
indicate the results
different determined
sampling by using
frequencies (250
batch
kHz, 500 kHz, and 1000 kHz) are presented as red, blue, and green solid circles in each figure. (250
sizes of 2, 4, 8, and 16, respectively. The results from three different sampling frequencies
kHz, 500 kHz, and 1000 kHz) are presented as red, blue, and green solid circles in each figure.
4.3. Performance Comparison of Methods
4.3. Performance Comparison of Methods
Figure 20 compares the best performance of the deep learning classification models based
Figure 20 compares the best performance of the deep learning classification models
on the three RNN models (LSTM, GRU, and BiLSTM) with tuned hyperparameters. Further-
based on the three RNN models (LSTM, GRU, and BiLSTM) with tuned hyperparameters.
more, the performance of the classification models based on the two conventional ultrasonic
Materials 2022, 15, x FOR PEER REVIEW 25 of 29
Furthermore, the performance of the classification models based on the two conventional
parameters (relative P-wave velocity and signal coherence) is also shown in the figure. The
ultrasonic parameters (relative P-wave velocity and signal coherence) is also shown in the
classification thresholds for the conventional methods are based on the obtained regression
figure. The classification thresholds for the conventional methods are based on the obtained
lines with 𝜃 at 3%, which returns 𝑉 , % of 1.00027 and coherence of 0.36.
regression lines with θ at 3%, which returns Vr,P 3% of 1.00027 and coherence of 0.36.

(a) 1 (b) 1
0.9 0.9
0.8 0.8
0.7
Cohen's kappa

0.7
Accuracy

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
UPV MSC LSTM GRU BiLSTM UPV MSC LSTM GRU BiLSTM
Type of classifica on model Type of classifica on model
Figure 20.
20. Comparison
Comparisonof of
performance of classification
performance models
of classification basedbased
models on allon
tested methods:
all tested (a) ac-
methods:
curacy, and (b)
(a) accuracy, andCohen’s kappa.
(b) Cohen’s kappa.

It was demonstrated that the performance of the deep learning classification models
was far more improved than
than those
those based
based on
on the
the relative
relative P-wave velocity,VVr,p
P-wave velocity, r,p. Among the

tested parameters,
parameters, the
the BiLSTM
BiLSTMmodel
modelwith
withfine-tuned
fine-tunedhyperparameter
hyperparameterhas hasthethebest
bestoverall
over-
performance,
all with
performance, anan
with accuracy ofof
accuracy 74%
74%and
andkappa
kappaofof0.48.
0.48.The
TheGRU
GRU model
model has slightly
less performance, with the top performing model demonstrating an accuracy of 71% and
kappa of 0.40. The least performing RNN model was the LSTM with an accuracy of 67%
and kappa of 0.32. The UPV method was placed in a distant last place, with an accuracy
of 53% and kappa of 0.07. Therefore, Vr,p is not an effective parameter for detecting early
Materials 2023, 16, 3502 23 of 26

less performance, with the top performing model demonstrating an accuracy of 71% and
kappa of 0.40. The least performing RNN model was the LSTM with an accuracy of 67%
and kappa of 0.32. The UPV method was placed in a distant last place, with an accuracy
of 53% and kappa of 0.07. Therefore, Vr,p is not an effective parameter for detecting early
concrete damages caused by steel corrosion in the rust propagation period. Previous
researchers observed that UPV of concrete was sensitive to the presence of surface-breaking
cracks and subsurface cracks and voids [55]. However, Vr,p of concrete did not show a
clear correlation with the corrosion levels of steel in concrete in this study, even for the
concrete specimens with surface-breaking cracks (i.e., D19 and D22 specimens). It was
observed in this study that the surface-breaking cracks first appeared on the surface of
concrete at a corrosion level ranging from 4% to 5%. The surface-breaking cracks were
tightly closed at the early stage of corrosion. Furthermore, corrosion products could fill the
microcracks and enhanced porosity of damaged concrete. Therefore, it can be inferred that
early concrete damages caused by steel corrosion could not affect the coherence parts of
ultrasonic pulse waves. Even so, the signal-coherence-based classification model resulted
in far more improved accuracy of 73% compared to the UPV model. This result reveals
that incoherent parts of ultrasonic pulse waves are informative of the minor concrete
damages associated with steel corrosion. However, the balance of true positive predictions
between the classes from the signal coherence method was relatively low, which resulted in
a kappa of 0.23. The overall low performance of the signal coherence method is attributed
to the significant changes in frequency response caused by the formation of cracks in
concrete. The variation of signal coherence remains a very low value and becomes stable
regardless of the corrosion level progress. Furthermore, as discussed in this study, the
distribution of signal coherence, as a nonlinear ultrasonic parameter, could be dependent
on several signal processing parameters such as the length of signals, the range of time,
and frequency windows, as discussed in Section 4.1. Therefore, special cares are needed to
find an optimal set of parameters that result in the best performance of the model based on
signal coherence. Therefore, it was demonstrated that the deep learning approach based on
RNN for ultrasonic pulse waves is a potential method for evaluating early concrete damage
caused by steel corrosion in the rust propagation stage.

5. Conclusions
This study investigated the feasibility of ultrasonic pulse wave measurements for early
detection of concrete damages caused by corroded steel in concrete using a deep learning
approach based on RNN. A series of experimental studies was performed in the laboratory
to collect ultrasonic pulse waves through reinforced concrete cube specimens where a
reinforcing steel with various corrosion levels was embedded. A bilinear model, classifying
the reinforced concrete cubes into solid and damaged concrete according to the threshold of
3% steel corrosion, was developed based on deep learning of ultrasonic pulse waves using
RNN. The performance of the deep learning classification model based on three different
RNN models (LSTM, GRU, and BiLSTM) were compared. Furthermore, the performance
of the deep learning classification models were compared with the classification models
based on two conventional ultrasonic testing parameters (relative P-wave velocity and
signal consistency). Summarized below are four important findings in this study:
1. The performance of deep learning classification models was affected by various
parameters: length of time signal, sampling frequency of time signal, type of input,
networks, and hyperparameters (batch size and epoch). The use of an extracted
feature (i.e., IF and/or SE) as an input of RNN-based deep learning models resulted in
better performance and far more improved computational efficiency than using time
series. It was observed that time series with a length of 5 ms and a sampling frequency
of 500 MHz was appropriate as an input of the feature extraction processes. However,
it was difficult to reach general conclusions on the effects of various input and training
parameters because different sets of parameters affected the performance results for
Materials 2023, 16, 3502 24 of 26

the three RNN algorithms in a different way. A fine-tuning of hyperparameters further

improved the performance of deep learning classification models.
2. Deep learning classification models, in the form of RNNs, were effective for the evalu-
ation of early concrete damage caused by steel corrosion in concrete with acceptable
accuracy. The best performance was obtained using the bidirectional long short-term
memory (BiLSTM) model based on a combination of instantaneous frequency and
spectral entropy (IFSE) of ultrasonic pulse waves with a length of 5 ms and a sampling
frequency of 250 kHz. Based on this method, the obtained accuracy and Cohen’s
kappa were 74% and 0.48, respectively.
3. It was demonstrated that the relative P-wave velocity was not practical for detect-
ing early concrete damages associated with steel corrosion in the rust propagation
phase. The classification model based on ultrasonic pulse velocity (UPV) resulted in a
relatively low accuracy of 53% with a kappa of 0.07.
4. The classification model based on signal coherence achieved more improved accuracy
of 73% than that based on UPV. However, the method resulted in a relatively low
kappa of 0.23, which is attributed to unbalanced true positives between classes. In
addition, there are many parameters in this method, such as the analyzed signal
time series and frequency window, that are not standardized and require engineering
judgment. Therefore, special care is needed to optimize the model that results in the
best performance of the model.
5. The ultrasonic pulse wave data used in this study were collected from the limited
number of concrete specimens fabricated in the laboratory, with corroded steel artifi-
cially accelerated by the impressed current technique, in specific saturation conditions
(fully saturated conditions). Furthermore, signal measurements in this research were
performed in the direct measurement configuration, which could limit this study’s
practicality for actual structures. Therefore, more systematic studies that consider
various experimental conditions in field applications are needed to reach more general
conclusions.

Author Contributions: Conceptualization, K.P.V.R., J.A.M. and S.-H.K.; methodology, K.P.V.R. and
J.A.M.; software, J.A.M. and S.-H.K.; validation, J.A.M.; formal analysis, J.A.M. and S.-H.K.; investi-
gation, J.A.M.; resources, S.-H.K.; data curation, J.A.M.; writing—original draft preparation, J.A.M.;
writing—review and editing, S.-H.K. and K.-H.L.; visualization, J.A.M.; supervision, S.-H.K.; project
administration, S.-H.K. and K.-H.L.; funding acquisition, S.-H.K. All authors have read and agreed to
the published version of the manuscript.
Funding: This work was supported by the Korea Institute of Marine Science and Technology Promo-
tion (KIMST) grant funded by the Ministry of Oceans and Fisheries for the project titled ‘Develop-
ment of smart maintenance monitoring techniques to prepare for disaster and deterioration of port
infra structures’.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or
in the decision to publish the results.

References
1. Kashani, M.M.; Lowes, L.N.; Crewe, A.J.; Alexander, N.A. Finite Element Investigation of the Influence of Corrosion Pattern on
Inelastic Buckling and Cyclic Response of Corroded Reinforcing Bars. Eng. Struct. 2014, 75, 113–125. [CrossRef]
2. Kashani, M.M.; Crewe, A.J.; Alexander, N.A. Nonlinear Cyclic Response of Corrosion-Damaged Reinforcing Bars with the Effect
of Buckling. Constr. Build. Mater. 2013, 41, 388–400. [CrossRef]
3. Ma, Y.; Zhang, J.; Wang, L.; Liu, Y. Probabilistic Prediction with Bayesian Updating for Strength Degradation of RC Bridge Beams.
Struct. Saf. 2013, 44, 102–109. [CrossRef]
Materials 2023, 16, 3502 25 of 26

4. Alexander, M.; Beushausen, H. Durability, Service Life Prediction, and Modelling for Reinforced Concrete Structures—Review
and Critique. Cem. Concr. Res. 2019, 122, 17–29. [CrossRef]
5. Bezuidenhout, S.R.; van Zijl, G.P.A.G. Corrosion Propagation in Cracked Reinforced Concrete, toward Determining Residual
Service Life. Struct. Concr. 2019, 20, 2183–2193. [CrossRef]
6. Ramezanianpour, A.A.; Pilvar, A.; Mahdikhani, M.; Moodi, F. Practical Evaluation of Relationship between Concrete Resistivity,
Water Penetration, Rapid Chloride Penetration and Compressive Strength. Constr. Build. Mater. 2011, 25, 2472–2479. [CrossRef]
7. Ormellese, M.; Berra, M.; Bolzoni, F.; Pastore, T. Corrosion Inhibitors for Chlorides Induced Corrosion in Reinforced Concrete
Structures. Cem. Concr. Res. 2006, 36, 536–547. [CrossRef]
8. Tuutti, K. Corrosion of Steel in Concrete Report No. 4; Swedish Cement and Concrete Research Institute: Stockholm, Sweden, 1982;
Available online: http://www.cbi.se/viewNavMenu.do?menuID=317&oid=857 (accessed on 15 February 2023).
9. Aveldaño, R.R.; Ortega, N.F. Characterization of Concrete Cracking Due to Corrosion of Reinforcements in Different Environments.
Constr. Build. Mater. 2011, 25, 630–637. [CrossRef]
10. Angst, U.M. Challenges and Opportunities in Corrosion of Steel in Concrete. Mater. Struct. 2018, 51, 4. [CrossRef]
11. Sohail, M.G.; Kahraman, R.; Alnuaimi, N.A.; Gencturk, B.; Alnahhal, W.; Dawood, M.; Belarbi, A. Electrochemical Behavior of
Mild and Corrosion Resistant Concrete Reinforcing Steels. Constr. Build. Mater. 2020, 232, 117205. [CrossRef]
12. Hou, B.; Li, X.; Ma, X.; Du, C.; Zhang, D.; Zheng, M.; Xu, W.; Lu, D.; Ma, F. The Cost of Corrosion in China. NPJ Mater. Degrad.
2017, 1, 4. [CrossRef]
13. Goyal, A.; Pouya, H.S.; Ganjian, E.; Claisse, P. A Review of Corrosion and Protection of Steel in Concrete. Arab. J. Sci. Eng. 2018,
43, 5035–5055. [CrossRef]
14. Kim, J.K.; Yee, J.J.; Kee, S.H. Electrochemical Deposition Treatment (Edt) as a Comprehensive Rehabilitation Method for Corrosion-
Induced Deterioration in Concrete with Various Severity Levels. Sensors 2021, 21, 6287. [CrossRef]
15. Li, Y.; Zhou, Y.; Wang, R.; Li, Y.; Wu, X.; Si, Z. Experimental Investigation on the Properties of the Interface between RCC Layers
Subjected to Early-Age Frost Damage. Cem. Concr. Compos. 2022, 134, 104745. [CrossRef]
16. American Standard of Testing and Materials International. ASTM C876-15; Standard Test Method for Standard Test Method for
Corrosion Potentials of Uncoated Reinforcing Steel in Corrosion Potentials of Uncoated Reinforcing Steel in Concrete. ASTM:
West Conshohocken, PA, USA, 2015. Available online: https://www.astm.org/c0876-15.html (accessed on 15 February 2023).
17. Chen, L.; Su, R.K.L. Corrosion Rate Measurement by Using Polarization Resistance Method for Microcell and Macrocell Corrosion:
Theoretical Analysis and Experimental Work with Simulated Concrete Pore Solution. Constr. Build. Mater. 2021, 267, 121003.
[CrossRef]
18. Feng, H.; Cui, L.; Zhang, M. Steel Corrosion Behavior Measurement Based on Electrochemical Approach. Int. J. Electrochem. Sci.
2016, 11, 4658–4666. [CrossRef]
19. Ribeiro, D.V.; Abrantes, J.C.C. Application of Electrochemical Impedance Spectroscopy (EIS) to Monitor the Corrosion of
Reinforced Concrete: A New Approach. Constr. Build. Mater. 2016, 111, 98–104. [CrossRef]
20. Hornbostel, K.; Larsen, C.K.; Geiker, M.R. Relationship between Concrete Resistivity and Corrosion Rate—A Literature Review.
Cem. Concr. Compos. 2013, 39, 60–72. [CrossRef]
21. Lorenzi, A.; Caetano, L.F.; Chies, J.A.; Da Silva Filho, L.C.P. Investigation of the Potential for Evaluation of Concrete Flaws Using
Nondestructive Testing Methods. Int. Sch. Res. Not. 2014, 2014, 543090. [CrossRef]
22. Antonaci, P.; Bruno, C.L.E.; Bocca, P.G.; Scalerandi, M.; Gliozzi, A.S. Nonlinear Ultrasonic Evaluation of Load Effects on
Discontinuities in Concrete. Cem. Concr. Res. 2010, 40, 340–346. [CrossRef]
23. Pahlavan, L.; Zhang, F.; Blacquière, G.; Yang, Y.; Hordijk, D. Interaction of Ultrasonic Waves with Partially-Closed Cracks in
Concrete Structures. Constr. Build. Mater. 2018, 167, 899–906. [CrossRef]
24. Arumaikani, T.; Sasmal, S.; Kundu, T. Detection of Initiation of Corrosion Induced Damage in Concrete Structures Using Nonlinear
Ultrasonic Techniques. J. Acoust. Soc. Am. 2022, 151, 1341–1352. [CrossRef] [PubMed]
25. Basu, S.; Thirumalaiselvi, A.; Sasmal, S.; Kundu, T. Nonlinear Ultrasonics-Based Technique for Monitoring Damage Progression
in Reinforced Concrete Structures. Ultrasonics 2021, 115, 106472. [CrossRef] [PubMed]
26. Wang, X.; Niederleithinger, E. Coda Wave Interferometry Used to Detect Loads and Cracks in a Concrete Structure under Field
Conditions. In Proceedings of the 9th European Workshop on Structural Health Monitoring, Manchester, UK, 10–13 July 2018;
Volume 807, pp. 1–8.
27. Ahn, E.; Shin, M.; Popovics, J.S.; Weaver, R.L. Effectiveness of Diffuse Ultrasound for Evaluation of Micro-Cracking Damage in
Concrete. Cem. Concr. Res. 2019, 124, 105862. [CrossRef]
28. Schurr, D.P.; Kim, J.Y.; Sabra, K.G.; Jacobs, L.J. Damage Detection in Concrete Using Coda Wave Interferometry. NDT E Int. 2011,
44, 728–735. [CrossRef]
29. Castellano, A.; Fraddosio, A.; Piccioni, M.D.; Kundu, T. Linear and Nonlinear Ultrasonic Techniques for Monitoring Stress-Induced
Damages in Concrete. J. Nondestruct. Eval. Diagn. Progn. Eng. Syst. 2021, 4, 041001. [CrossRef]
30. Singh, S.; Pandey, S.K.; Pawar, U.; Janghel, R.R. Classification of ECG Arrhythmia Using Recurrent Neural Networks. Procedia
Comput. Sci. 2018, 132, 1290–1297. [CrossRef]
31. Kim, H.; Phan, T.Q.; Hong, W.; Chun, S.Y. Physiology-Based Augmented Deep Neural Network Frameworks For ECG Biometrics
With Short ECG Pulses Considering Varying Heart Rates. Pattern Recognit. Lett. 2022, 156, 1–6. [CrossRef]
Materials 2023, 16, 3502 26 of 26

32. Rejaibi, E.; Komaty, A.; Meriaudeau, F.; Agrebi, S.; Othmani, A. MFCC-Based Recurrent Neural Network for Automatic Clinical
Depression Recognition and Assessment from Speech. Biomed. Signal Process. Control 2022, 71, 103107. [CrossRef]
33. Hong, S.; Zheng, F.; Shi, G.; Li, J.; Luo, X.; Xing, F.; Tang, L.; Dong, B. Determination of Impressed Current Efficiency during
Accelerated Corrosion of Reinforcement. Cem. Concr. Compos. 2020, 108, 103536. [CrossRef]
34. Designation: G1-03; ASTM Standard Practice for Preparing, Cleaning, and Evaluating Corrosion Test Specimens. ASTM Special
Technical Publication: West Conshohocken, PA, USA, 2017.
35. ASTM C597-09; Standard Test Method for Pulse Velocity Through Concrete. ASTM: West Conshohocken, PA, USA, 2009; Volume 4.
36. Petro, J.T.; Kim, J. Detection of Delamination in Concrete Using Ultrasonic Pulse Velocity Test. Constr. Build. Mater. 2012, 26,
574–582. [CrossRef]
37. Arbaoui, A.; Aribi, C.; Boumaiza, M.; Mohamadi, S.; Ahmed, F.A. CNN-Based Concrete Cracks Detection Using Multiresolution
Analysis. In Proceedings of the 2022 7th International Conference on Image and Signal Processing and their Applications (ISPA
2022), Mostaganem, Algeria, 8–9 May 2022. [CrossRef]
38. MathWorks Magnitude-Squared Coherence—MATLAB Mscohere. Available online: https://www.mathworks.com/help/signal/
ref/mscohere.html (accessed on 4 January 2022).
39. Tsantekidis, A.; Passalis, N.; Tefas, A. Chapter 5—Recurrent Neural Networks; Iosifidis, A., Tefas, A., Eds.; Academic Press:
Cambridge, MA, USA, 2022; pp. 101–115. ISBN 978-0-323-85787-1.
40. Alzubaidi, L.; Zhang, J.; Humaidi, A.J.; Al-Dujaili, A.; Duan, Y.; Al-Shamma, O.; Santamaría, J.; Fadhel, M.A.; Al-Amidie, M.;
Farhan, L. Review of Deep Learning: Concepts, CNN Architectures, Challenges, Applications, Future Directions. J. Big Data 2021,
8, 53. [CrossRef] [PubMed]
41. Zhang, Q.; Barri, K.; Babanajad, S.K.; Alavi, A.H. Real-Time Detection of Cracks on Concrete Bridge Decks Using Deep Learning
in the Frequency Domain. Engineering 2021, 7, 1786–1796. [CrossRef]
42. Hochreiter, S.; Schmidhuber, J. Long Short-Term Memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef] [PubMed]
43. Li, F.; Liu, M.; Zhao, Y.; Kong, L.; Dong, L.; Liu, X.; Hui, M. Feature Extraction and Classification of Heart Sound Using 1D
Convolutional Neural Networks. EURASIP J. Adv. Signal Process. 2019, 2019, 59. [CrossRef]
44. Cho, K.; van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations
Using RNN Encoder-Decoder for Statistical Machine Translation. arXiv 2014, arXiv:1406.1078.
45. Choe, D.E.; Kim, H.C.; Kim, M.H. Sequence-Based Modeling of Deep Learning with LSTM and GRU Networks for Structural
Damage Detection of Floating Offshore Wind Turbine Blades. Renew. Energy 2021, 174, 218–235. [CrossRef]
46. Kłosowski, G.; Rymarczyk, T.; Wójcik, D.; Skowron, S.; Cieplak, T.; Adamkiewicz, P. The Use of Time-Frequency Moments as
Inputs of Lstm Network for Ecg Signal Classification. Electronics 2020, 9, 1452. [CrossRef]
47. Gaona, A.J.; Arini, P.D. Deep Recurrent Learning for Heart Sounds Segmentation Based on Instantaneous Frequency Features.
Elektron 2020, 4, 52–57. [CrossRef]
48. Boashash, B. Estimating and Interpreting The Instantaneous Frequency of a Signal—Part 1: Fundamentals. Proc. IEEE 1992, 80,
520–538. [CrossRef]
49. Suto, J.; Oniga, S.; Lung, C.; Orha, I. Comparison of Offline and Real-Time Human Activity Recognition Results Using Machine
Learning Techniques. Neural Comput. Appl. 2020, 32, 15673–15686. [CrossRef]
50. Nguyen, H.; Hoang, N. Automation in Construction Computer Vision-Based Classification of Concrete Spall Severity Using
Metaheuristic-Optimized Extreme Gradient Boosting Machine and Deep Convolutional Neural Network. Autom. Constr. 2022,
140, 104371. [CrossRef]
51. Khadse, V.; Mahalle, P.N.; Biraris, S.V. An Empirical Comparison of Supervised Machine Learning Algorithms for Internet of
Things Data. In Proceedings of the 2018 Fourth International Conference on Computing Communication Control and Automation
(ICCUBEA), Pune, India, 16–18 August 2018; pp. 2–7. [CrossRef]
52. Thawkar, S. A Hybrid Model Using Teaching–Learning-Based Optimization and Salp Swarm Algorithm for Feature Selection and
Classification in Digital Mammography. J. Ambient Intell. Humaniz. Comput. 2021, 12, 8793–8808. [CrossRef]
53. Ho, G.Y.; Leonhard, M.; Volk, G.F.; Foerster, G.; Pototschnig, C.; Klinge, K.; Granitzka, T.; Zienau, A.K.; Schneider-Stickler, B.
Inter-Rater Reliability of Seven Neurolaryngologists in Laryngeal EMG Signal Interpretation. Eur. Arch. Oto-Rhino-Laryngol. 2019,
276, 2849–2856. [CrossRef] [PubMed]
54. Nossier, S.A.; Wall, J.; Moniri, M.; Glackin, C.; Cannings, N. A Comparative Study of Time and Frequency Domain Approaches
to Deep Learning Based Speech Enhancement. In Proceedings of the 2020 International Joint Conference on Neural Networks
(IJCNN), Glasgow, UK, 19–24 July 2020. [CrossRef]
55. Kee, S.H.; Zhu, J. Using Piezoelectric Sensors for Ultrasonic Pulse Velocity Measurements in Concrete. Smart Mater. Struct. 2013,
22, 115016. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.