CNN deep learning-based monitoring of stress and damage using of
electromechanical impedance responses of CSA sensor
JEONG-TAE KIM, QUOC-BAO TA, NGOC-LAN PHAM
Department of Ocean Engineering
Pukyong National University
Busan, KOREA
Abstract: A multi-task 2D CNN model is designed for integrated monitoring stress and damage in concrete specimens utilizing the raw
impedance signatures of capsule-like smart aggregated (CSA). The fundamental theory of CSA-based EMI method is presented to describe
how the sensor responds to compressive loads. Next, compression tests on a CSA-embedded concrete cylinder are conducted to record the
stress-damage EMI responses of CSA sensor under applied stresses. The multi-task 2D CNN model learned the impedance signals for
predicting the concrete stress and damage is constructed. Consequently, the generalization and robustness of the developed model are tested
against noise and untrained data.
Keywords: impedance-based, PZT sensor, capsule-like smart aggregate, convolutional neural network, stress estimation, damage
identification.
Received: April 11, 2024. Revised: September 17, 2024. Accepted: November 19, 2024. Published: December 31, 2024.
1. Introduction
Concrete structures play a vital role in civil infrastructure
due to their adaptability and cost-effectiveness. After a
prolonged operation period, key components in concrete
structures experience degradation and damage from
continuous stress. In recent studies, the CNN-based regression
algorithms have been integrated with EMI techniques for
stress estimation of concrete structures. Nguyen et al.
employed a 1D CNN algorithm to learn damage-sensitive
EMI features for monitoring damage in prestressed concrete
girders [1]. Ta et al. [2,3] developed an impedance-based 1D
CNN regression model for stress monitoring in concrete using
raw EMI data from SA and CSA sensors. The stress in the
investigated concrete structure could be automatically
estimated with high accuracy, even under noise effects and
missing data. The mentioned works demonstrated the
accuracy of the deep learning method in stress estimation in
concrete structures even in the presence of noise and missing
data.
CNN-based classification algorithms have also been
integrated with EMI techniques for damage identification in
concrete structures. An impedance-based 1D CNN deep
learning approach was proposed to detect bolt loosening in
steel structures using raw EMI data [4]. Another study by
Nguyen et al. employed the 1D CNN model to detect damage
in PZT transducers [5]. Yan et al. [6] proposed a 1D CNN
integrated with EMI data to evaluate the early-age hydration
of cement mortar, outperforming traditional machine learning
methods in quantifying EMI response changes.
Despite previous research efforts, existing CNN deep
learning models could handle stress estimation and damage
detection tasks separately using either CNN-based regression
or CNN-based classification. The status of the concrete
structures under compression has not been fully explored. To
address these gaps, this paper introduces a multi-task 2D CNN
model that integrates regression and classification, enabling
the simultaneous monitoring of concrete stress and damage
using CSA-based EMI responses.
This study presents a multi-task 2D CNN model for
integrated monitoring stress and damage in concrete
specimens utilizing the raw impedance signatures of capsule-
like smart aggregated (CSA). The fundamental theory of
CSA-based EMI method is presented to describe how the
sensor responds to compressive loads. Next, compression tests
on a CSA-embedded concrete cylinder are conducted to
record the stress-damage EMI responses of the CSA sensor
under applied stresses. The multi-task 2D CNN model learned
the impedance signals for predicting the concrete stress and
damage is constructed. Consequently, the generalization and
robustness of the developed model are tested against noise and
untrained data.
2. CSA-based EMI
Measurement Technique
The CSA sensor prototype for the EMI measurement
technique is shown in Fig. 1. The CSA sensor is fabricated by
attaching a PZT patch onto an aluminium interface, which is
covered by a hollow aluminium capsule. The dimensions of
the CSA prototype are L × W × H = 25 × 25 × 11 mm [7]. The
aluminium interface plate considers the CSA capsule's wall as
fixed ends protected and is allowed to vibrate freely. The
thickness of 2 mm of vibrating plate is chosen to pre-
determine the sensitive frequency band of the CSA sensor [8].
Fig. 1. Prototype of capsule-like smart aggregate (CSA)
Figure 2 shows a model of CSA-based impedance
monitoring for concrete structures. When stress is applied, the
CSA sensor embedded in the concrete structure experiences
compressive stress (σN) along with the vertical direction (i.e.,
z-direction). At the same time, the other CSA's surfaces (i.e.,
y-direction and x-direction) are subjected to tension stress (σT)
due to Poisson's effect (see Fig. 2a). As a result, the vibrating
plate undergoes expansion under the tensile stress in both the
Aluminum interface
(thickness 2.0 mm)
H =11
mm
Aluminum cover plate
(thickness 2.0 mm)
2.0 mm
y
z
x
PZT
Wall
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
222
Volume 4, 2024
y-direction and x-direction. The CSA's deformation affects the
structural state of the vibrating plate and impedance responses
of the PZT attached to it (see Fig. 2b and 2c).
a) CSA in structure under compression
b) Section B-B c) Changes in EMI responses
Fig. 2. Behavior of EMI responses of CSA embedded in concrete structure
under applied stress in z-direction
When the applied stress (e.g., σN + ΔσN) increases and
reaches the yield condition of the concrete material, the local
damage may occur. When the damage occurs, the tensile stress
on the vibrating plate is released rapidly, leading to abrupt
changes in the EMI responses [7,9], as shown in Fig. 2c.
3. Experimental Test
3.1 Fabrication of CSA-Embedded Concrete
Cylinder
Figure 3 presents the fabrication procedure for a CSA-
embedded concrete cylinder. A CSA sensor was strategically
placed in a cylinder mold measuring 100 × 200 mm. The CSA
was positioned 140 mm from the bottom of the cylinder mold.
To secure the CSA sensor within the mold, plastic wires and
a steel bar (2 mm in diameter and 150 mm in length) were
utilized. After 28 days of the curing process, the concrete
cylinder embedded with CSA sensor was used for the
impedance test.
Fig. 3. Fabrication of CSA-embedded concrete cylinder
3.2 Experimental Setup
The experimental setup for compression test of concrete
cylinder is presented in Fig. 4. The MTS servo-hydraulic
materials testing system (version 793) was employed for the
compression test. The compression force of system was real-
time monitored by a load cell with a capacity of up to 500 kN.
An impedance analyzer (HIOKI 3532) was utilized to capture
stress-damage EMI signals from the CSA sensor, while a
KYOWA EDX-100A measured the ambient temperature.
a) Compression test setup b) Compressive loading scenario
Fig. 4. Experimental setup for compression test of concrete cylinder
EMI responses were measured in frequency range of 15
kHz to 26 kHz using 224 intervals. The notable peaks in EMI
responses in this range could be used to assess the sensitivity
of the embedded CSA sensor to compressive loading. The
recorded EMI responses, along with corresponding structural
attributes (i.e., stress levels and concrete damage levels), were
compiled to create a stress-damage EMI dataset for the 2D
CNN deep regression and classification model. The measured
temperature ranged between 22°C and 23°C. Due to the minor
variation of 1°C, the temperature effect on the EMI responses
was considered negligible.
Figure 4b shows six loading scenarios (S0 = 0 MPa to S5 =
12.68 MPa) introduced to the CSA-embedded cylinder. The
stress was applied in constant increments of 2.54 MPa within
2.5 minutes, maintaining a consistent loading rate of 0.0113
MPa/s. Following each increment, the stress increment was
paused for 4.5 minutes to obtain EMI responses.
3.3 Stress-Damage EMI Signatures of CSA-
Embedded Concrete Cylinder
Figure 5 plots the EMI responses collected in six stress
levels (i.e., S0 to S5). As the applied stress increased, both the
frequency and magnitude of the resonant peak exhibited a
downward trend. The variation in peak frequency and peak
magnitude were potentially caused by the high compressive
stress on the CSA sensors during concrete strength
development. The corresponding visual observation of the test
specimen under applied stresses S0-S5 is shown in Fig. 6. At
loading level S3, initial crack imitation was observed. As
loading progressed to level S4, crack propagation and concrete
spalling were noted. Ultimately, concrete damage continued
to develop, leading to failure at loading level S5.
Fig. 5. Impedance responses of CSA sensor under S0-S5
Applied stress
Tensile
stress, σT
BB
PZT
Local
damage
Concrete
structure
σN
y
z
x
B-B
PZT
y
x
z
Tensile
stress, σT CSA s
deformation
Fresh concrete
60
140
Steel bar f 2, l = 150
CSA sensor
Electric wire
Concrete cylinder
f 100 200
Wet blanket
Sensor installation Concrete pouring and curing
Impedance analyser
(HIOKI-3532)
Temperature analyser
(KYOWA EDX-100A)
Load cell
(capacity: 500 kN)
CSA-embedded
concrete cylinder
Stress
(MPa)
Force increasing (2.5 min)
0S0
S1
S2
S3
Impedance
measuring
(4.5 min)
Unloading
(5 min)
S4
S5
-5.08
-7.61
-10.15
-2.54
-12.68
Time (minutes)
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
223
Volume 4, 2024
Fig. 6. Obsever concrete damage of test specimen under S0-S5
4. Evaluation of Multi-task
2D CNN-based Deep Regression and
Classification Model
4.1 Design of Multi-task 2D CNN Model
Figure 7 illustrates the architecture of a multi-task 2D
CNN deep learning model using raw EMI responses of CSA
sensor. The model employs regression learning for stress
estimation and classification learning for damage
identification in concrete structures. The parameters and
hyperparameters of the model are selected based on previous
studies [10,11] and practical guidelines [12].
Fig. 7. Architecture of 2D CNN deep regression and classification model
The model includes three convolutional (Conv) layers,
three ReLU layers, two Maxpool layers, a global average
pooling (GAP) layer, two fully connected (Fc) layers, and
separate Regression and Classification output layers. The
multi-task 2D CNN model generates two outputs: "Stress
estimation" handled by regression learning, and "Damage
identification" handled by classification learning.
4.2 Stress and Damage Monitoring for Noise-
Contaminated Stress-Damage EMI Data
1) Data Preparation
The measured EMI data from the CSA-embedded concrete
cylinder and corresponding assigned labels (i.e., "stress" and
"damage level") is listed Table 1. The stress level S0 (0 MPa)
was excluded from the model configuration due to the
uncertainty in experimental measurements. For regression
learning, compression forces ranging from 2.53 MPa to 12.68
MPa (interval of 2.53 MPa) were labelled with five stress levels
(S1 to S5). For classification learning, the damage severity of test
specimen (i.e., "No damage," "Crack initiation,” "Crack
propagation and spalling," and "Failure") was labelled with four
levels "DL0," "DL1," "DL2," and "DL3," respectively.
In the compressive test of concrete specimen, the EMI
signals were measured with four ensembles for each applied
stress level, resulting in a total of 20 signals across five stress
levels. Gaussian noise was employed to enrich the databank
and to investigate the generalization and robustness of the
multi-task 2D CNN model on noise contamination. The
training dataset was generated by adding six Gaussian noise
levels (0-5%) with four iterations to the first two ensembles at
each stress level. It resulted in a total of 240 EMI signals
generated for five applied stress levels. The third ensemble
from each level was used to construct the validation set, which
consisted of five EMI signals in total. Similar for the testing
dataset, noise levels ranging from 1% to 5% (in 1%
increments) were applied to the last ensemble, generating ten
new EMI signals per stress level. This resulted in 250
additional EMI signals across all noise levels. Combined with
the five original EMI signals, the testing dataset comprised a
total of 255 signals. The validation set takes the third ensemble
of data collected at each stress level, resulting in five signals
in total.
TABLE I. ASSIGNED LABELS OF STRESS-DAMAGE EMI DATA FOR
MULTI-TASK 2D CNN MODEL
Stress
level
Observed concrete damage
Assigned label
Stress
(MPa)
Damage
level
S1
No damage
2.53
DL0
S2
No damage
5.07
DL0
S3
Crack initiation
7.61
DL1
S4
Crack propagation and
spalling
10.15
DL2
S5
Failure
12.68
DL3
The training set is visualized in Fig. 8. Each EMI signal
was labelled with its corresponding stress level and damage
status. With 225 data points for each EMI signal, a total of
10,800 data points were obtained for each stress level. For five
stress levels, the total number of data points was 54,000.
Examples of noise-contaminated stress-damage EMI signals
are shown in Fig.9.
Fig. 8. Visualization of training set of multi-task 2D CNN model
a) 4% noise b) 5% noise
Fig. 9. Visualization of noise-contaminated EMI signals
2) Training Results
Figure 10 plots the loss values of the 2D CNN model over
100 training epochs. Overall, both the training and validation
losses showed fluctuations and generally followed a
decreasing trend as training progressed. The 2D CNN model,
consisting of 6,143 training parameters, required 35.1 seconds
to complete the training process.
S0-2: No damage S3: Crack initiation S4: Crack propagation
& spalling S5: Failure
Stress estimation
Output
Regression
Damage identification
Classification
Fc1, 1
GAP
Input
EMI signals
Hidden layers
Autonomous stress & damage feature extraction
Fc2, 4
Stress level 1
Stress level 2
Stress level n
Damage level n
Damage level 2
Damage level 1
...
...
Damage
level ith
3x3 Conv, 18,/1
ReLU
3x3 Conv,18,/1
1x2 Maxpool, /2
ReLU
3x3 Conv,18,/1
1x2 Maxpool, /2
ReLU
5.07 - DL0
7.61 - DL1
10.15 - DL2
12.68 - DL3
2.53 - DL0
Stress
(MPa)
Damage
status
S1
S2
S3
S4
S5
Stress
level
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
224
Volume 4, 2024
Figure 11 shows two representative results of investigating
the effects of noises on the stress estimation and damage
identification of the multi-task 2D CNN model. For stress
estimation, the accuracy of the model decreased as the levels
of noise increased. The relationship between the RMSE index
and noise levels is illustrated in Figure 12a. For damage
identification, the model maintained performance with no
false predictions at noise levels up to 4%. Misclassifications
began to occur at a noise level of 5%, leading to an increase in
the false discovery rate to 2.5%, as shown in Figure 12b.
Fig. 10. Loss values of 2D CNN model after 100 epochs
a) 4% noise
b) 5% noise
Fig. 11. Stress prediction and damage identification by 2D CNN model
a) RMSE b) False discovery rate
Fig. 12. Stress estimation and damage identification performance of 2D
CNN model
4.3 Stress and Damage Monitoring for
Untrained Stress-Damage EMI Data
3) Data Preparation
To assess the performance of the multi-task 2D CNN
model under missing data conditions, the EMI signals
corresponding to stress level S2 were excluded from the
training and validation sets. This led to the removal of 48
signals from the training set and one signal from the validation
set. As a result, the training and validation sets consisted of 192
signals and four signals, respectively (see Fig. 13). The testing
set had a total of 255 signals for five applied stress levels.
Fig. 13. Visualization of partial untrained training set for 2D CNN model
4) Training Results
Figure 14 plots the training process of the multi-task 2D CNN
model using the designed dataset. It is observed that the
training loss gradually converged, while validation loss
fluctuated during the whole training process (100th epoch).
Fig. 14. Loss values of 2D CNN model with partial untrained databank
Figure 15 shows presentative results of stress estimation and
damage identification with partial untrained databank. For
stress estimation, the predicted stress levels showed
consistency with actual stress, achieving an RMSE value of
0.57. The prediction error for the untrained stress level S2 was
within 30% at a noise level of 5%. The relationship between
RMSE and noise levels is shown in Fig. 16a, where RMSE
values increased as noise levels increased. For damage
identification, some misclassifications occurred for the
partial untrained damage level DL0. The overall prediction
accuracy was 84% at a noise level of 5%. The false discovery
rate for damage identification across six noise levels is
summarized in Fig. 16b.
50
50
Excluded S2
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
225
Volume 4, 2024
Fig. 15. Stress estimation and damage identification of 2D CNN model with
partial untrained databank
a) RMSE b) False discovery rate
Fig. 16. Stress estimation and damage identification performance of 2D
CNN model with partial untrained databank
5) Discussion on Damage Identification
Results
Figure 17 shows the probability assessment of damage
identification results of damage level "DL0" with partial
untrained databank. A portion of "DL0" data, corresponding
to stress level S2, was excluded from the training and
validation phases. The x-axis in the figure represents the four
damage levels ("DL0"–"DL3"), while the left and right y-axes
denote the predicted value and standard distribution,
respectively.
The shaded area indicates the range within one standard
deviation (σ) from the mean (μ), encompassing 68.8% of the
predicted values around the central tendency. For all three
noise levels (i.e., 0%, 3%, and 5%), the shaded region
indicated damage levels "DL0" and "DL1". The mean value
was balanced between "DL0" and "DL1" at 0% and 3% noise
levels, and it shifted toward "DL0" at noise level 5%.
a) Noise 0%
b) Noise 3% c) Noise 5%
Fig. 17. Probability assessment of damage identification results with partial
untrained databank
5. Concluding Remarks
In this study, the multi-task 2D CNN model was developed
for integrated monitoring stress and damage in concrete
specimens utilizing the raw impedance signatures of capsule-
like smart aggregated (CSA). The fundamental theory of
CSA-based EMI method was presented to describe how the
sensor responds to compressive loads. The compression tests
on a CSA-embedded concrete cylinder were conducted to
record the stress-damage EMI responses of CSA sensor under
applied stresses. The multi-task 2D CNN model learned the
impedance signals for predicting the concrete stress and
damage was constructed. The generalization and robustness of
the developed model were validated against noise and
untrained data.
Acknowledgment
This work was supported by the Basic Science Research
Program through the National Research Foundation of Korea.
References
[1] Pham, Q.-Q., Ta, Q.-B., and Kim, J.-T., “Capsule-like Smart Aggregate
with Pre-Determined Frequency Range for Impedance-Based Stress
Monitoring,” Sensors, vol. 23, no. 1, pp. 434, 2022.
[2] Nguyen, T.-T., Ta, Q.-B., Ho, D.-D., Kim, J.-T., and Huynh, T.-C., “A
method for automated bolt-loosening monitoring and assessment using
impedance technique and deep learning,” Developments in the Built
Environment, vol.14, pp. 100122, 2023.
[3] Nguyen, T.-T., Ho, D.-D., and Huynh, T.-C., “Electromechanical
impedance-based prestress force prediction method using resonant
frequency shifts and finite element modelling,” Developments in the
Built Environment, vol. 12, pp. 100089, 2022.
[4] Ta, Q.-B., Pham, Q.-Q., Pham, N.-L., Huynh, T.-C., and Kim, J.-T.,
“Smart Aggregate-Based Concrete Stress Monitoring via 1D CNN
Deep Learning of Raw Impedance Signals,” Structural Control and
Health Monitoring, vol. 2024, no. 1, pp. 25. 2024
[5] Ta, Q.-B., Pham, Q.-Q., Pham, N.-L., and Kim, J.-T., “Integrating the
Capsule-like Smart Aggregate-Based EMI Technique with Deep
Learning for Stress Assessment in Concrete,” Sensors, vol. 24, no. 14,
2024.
[6] Nguyen, T.-T., Kim, J.T., Ta, Q.B., Ho, D.D., Phan, T.T.V., and
Huynh, T.C., “Deep learning-based functional assessment of
piezoelectric-based smart interface under various degradations,”
SMART STRUCTURES AND SYSTEMS, vol. 28, no. 1, pp. 69-87,
2021.
[7] Yan, Q., Liao, X., Zhang, C., Zhang, Y., Luo, S., and Zhang, D.,
“Intelligent monitoring and assessment on early-age hydration and
setting of cement mortar through an EMI-integrated neural network,”
Measurement, vol. 203, pp. 111984, 2022.
[8] Pham, Q.-Q., Ta, Q.-B., Park, J.-H., and Kim, J.-T., “Raspberry Pi
Platform Wireless Sensor Node for Low-Frequency Impedance
Responses of PZT Interface,” Sensors, vol. 22, no. 24, 2022.
5% noise
Untrained S2
50
5% noise
0% noise
Normal probability
density function
(PDF)
μ (mean)
μ + σ(std)
μ σ
68.8%
3% noise
5% noise
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
226
Volume 4, 2024
[9] Kocherla, A., and Subramaniam, K.V.L. (2020), “Embedded smart
PZT-based sensor for internal damage detection in concrete under
applied compression,” Measurement, Vol. 163.
[10] Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P., “Gradient-based
learning applied to document recognition,” Proceedings of the IEEE,
vol. 86, no. 11, pp. 2278-2324, 1998.
[11] Pham, N. L., Ta, Q. B., and Kim, J. T., “CNN-Based Damage
Identification of Submerged Structure-Foundation System Using
Vibration Data,” Applied Sciences, vol. 14, no. 17, pp. 7508, 2024.
[12] Zhang, Y., and Wallace, B., “A sensitivity analysis of (and
practitioners' guide to) convolutional neural networks for sentence
classification,” arXiv preprint arXiv:.03820, 2015
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2024.4.24
Jeong-Tae Kim, Quoc-Bao Ta,
Ngoc-Lan Pham
E-ISSN: 2732-9984
227
Volume 4, 2024