Engineering Structures 226 (2021) 111228

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Bolt loosening detection using impedance based non-destructive method and

probabilistic neural network technique with minimal training data
Wongi S. Na
Seoul National University of Science & Technology, Department of Civil Engineering, 232 Gongeung-ro, Nowon-gu, Seoul, 01811, Republic of Korea

ARTICLE INFO ABSTRACT

Keywords: For centuries, bolted fastening has been widely used in aerospace, civil and mechanical engineering areas to
Bolt loosening achieve joining of parts. However, studies have shown that vibration occurs virtually in all dynamic systems such
Bolted structure as in machines and structures which seems to be the main cause of bolt loosening. As this can significantly
Electromechanical impedance reduce the load bearing capacities of a system, it is important to monitor the state of bolts to ensure safety. In this
Piezoelectric transducer
study, piezoelectric transducer based method known as the electromechanical impedance (EMI) technique with
Artificial neural networks
probabilistic neural networks (PNN) was used to identify torque loss of bolts on three bolted structure specimens.
The training data from the first specimen was used to predict torque loss of different specimens to evaluate the
possibility of the proposed idea. Results show over 90% accuracy with the PNN algorithm designed for this work
bringing one step close for the piezoelectric based non-destructive testing technique to be applied to real
structures.

1. Introduction proposed a new method for monitoring the preload torque in a com­
posite bolted connection by designing a unique washer. Chen et al [6]
For centuries, bolted fastening has been widely used in aerospace, created a smart washer encircled by an FBG sensor to measure the pre-
civil and mechanical engineering areas to achieve joining of parts. load on bolted connections. Ho et al [7] created a smart anchor plate
While it has many advantages such as the ease of removing parts for with FBG sensor to monitor the load level of the rock bolt in the pro­
repair and being non-sensitive to surface preparations, studies have posed study. Argatov and Sevostianov [8] proposed a method using
shown that vibrations during the service period can cause the bolts to electrical constriction resistance of the contact interface between two
loosen. As vibration occurs virtually in all dynamic systems such as in conductive members clamped by a bolt. Furthermore, studies on vi­
machines and structures, it is vital to know if a system with bolted joints bration methods using non-contact laser excitation, piezoelectric
is experiencing bolt loosening as this can significantly reduce the load transducer and changes in natural frequencies can be found in [9–13].
bearing capacities of the system. This can result in catastrophic failures While these techniques have their advantages, they may also have
for machines and structures. certain limitations. For an example, using large and heavy equipment
Up to date, various authors have investigated on monitoring of the may produce results with high precision and accuracy but it would be
bolted connection to ensure safety of the system in service. These stu­ difficult to apply the technology in real fields due to its large size. Using
dies include acoustic emission methods, optical inspections, fiber low frequency vibration methods can maximize the sensing area but is
Bragg-grating (FBG) sensors, electrical conductivity and vibration based difficult to detect small damage at an early stage.
methods [1–8]. Amerini and Meo [1] developed a reliable index for In this study, an impedance based method known as the electro­
assessing the loosening/tightening health state of a bolted structure mechanical impedance (EMI) technique is used with an artificial neural
based on acoustic/ultrasonic parameters. Guarino et al [2] demon­ networks (ANN) technique to predict torque loss of three different sized
strated a simple technique for acquiring data from an impact test using bolts for bolted structure specimens made for this study. Research on
a basic electronic stethoscope to identify loosened bolts. Huang et al [3] using the EMI technique for damage detection of structures can be
utilized a washer as an optical force transducer for direct tension found in various literatures including in Huo et al [14] and Ritdum­
measurement of bolted joints. Khomenko et al [4] developed a reusable rongkul et al [15] where the authors conducted experiments to monitor
FBG sensors to achieve precise monitoring of the bolt. Furthermore bolted joints with PZT transducers. Studies by Lim et al [16] and Fan
studies based on FBG can be found in [5,6,7] where Yeager et al [5] et al [17] also showed successful damage detection of aluminum truss

E-mail address: wongi@seoultech.ac.kr.

https://doi.org/10.1016/j.engstruct.2020.111228
Received 24 February 2020; Received in revised form 30 June 2020; Accepted 13 August 2020
Available online 12 October 2020
0141-0296/ © 2020 Published by Elsevier Ltd.
W.S. Na Engineering Structures 226 (2021) 111228

specimen and pin connection loosening using the EMI technique, re­
spectively. Corrosion monitoring was also proven to be possible with
the EMI technique where Li et al [18,19] made a smart corrosion
coupon which identified corrosion induced thickness loss in [18] and
developed theoretical modeling of the EMI instrumented circular pie­
zoelectric-metal transducer for corrosion monitoring in [19]. Regarding
machine learning and visual images, Chen et al [20] proposed a loos­
eness diagnosis method for connecting bolt on sensitive mixed-domain
features of excitation response and manifold learning, Cha et al [21]
used Hough transform and support vector machines to detect loosening
of bolts based on vision, Zhao et al [22] combined deep learning and
machine vision and proposed a bolt loosening angle detection tech­
nology, Nguyena et al [23] proposed an algorithm using image pro­
cessing technique for bolted steel structures and Ramana et al [24]
introduced fully automated vision-based method for detecting loosened
bolts using the Viola-Jones algorithm.
An interesting attempt is made for the study where a training data
from a test specimen is used to predict torque loss of a totally different
specimen with the same designed dimensions and properties. The
purpose of this was that in civil structures such as in cable stayed Fig. 1. Experimental setup for the EMI method.
bridges, there are countless number of bolts where bolted joint parts
have similar dimensions and properties throughout the structure. As it is fully powered by the USB cable connected to a computer. A typical
acquiring data for every single parts to train the ANN algorithm would experimental setup for conducting the EMI technique can be seen in
be unrealistic, this study investigates the possibility of using a training Fig. 1 using the AD5933 evaluation board where it is connected to a
data from a specimen to identify torque loss of a different specimen. computer and a bolted structure test specimen made for this study.
This would significantly reduce the damage identification process as Details of this bolted structure will be explained in the later section. The
only a single set of training data is required to monitor other parts PZT model PSI-5A4E used for this study was purchased from Piezo
assuming that the design is identical. In addition, an idea of applying Systems (http://piezo.com) where all the tests were conducted at a
centered mean averaging method on acquired impedance signatures room temperature of 24 °C ( ± 0.2 °C) to minimize impedance signature
were also evaluated to improve the accuracy of the torque prediction changes due to variations in temperature. The PZT patch was cut into
predicted by the designed ANN algorithm for this study. With the 15 mm square patches to be attached on top of the bolt head using a
findings from this research, one can create a damage detection system commercialized epoxy (Loctite Quick Set) and left at room temperature
that can automatically predict torque values (using ANN algorithm) for 48 h to ensure full curing of the adhesive.
with high accuracy at low cost.
2.2. Data identification index
2. Damage detection and data processing methods
Since the principle of the EMI technique is to monitor any changes
2.1. Electromechanical impedance technique in impedance signatures to identify any presence of damage, one needs
to quantify this change to evaluate the severity of such damage. To do
The principle of the EMI method is to attach a PZT transducer onto a this, there are three different statistical metrics used in general to
target structure and monitor changes in electrical impedance sig­ quantify damage severity when performing the EMI technique. These
natures. The 1-D equation (equation (1)) proposed by Liang et al [25] are RMSD (root mean square deviation), MAPD (mean average per­
shows that the electrical admittance (inverse of impedance), Y ( ) is centage deviation) and CCD (correlation coefficient deviation) which is
directly related to the mechanical impedance of the structure Zs ( ) and shown in Eqs. (2) to (4). The variables Zio and Zi represent reference and
the attached PZT transducer Za ( ). Thus any changes in the structure corresponding impedance signatures where real part of impedance
(such as damage) will affect the mechanical impedance of the structure were used throughout the study as it has been proven to perform better
which can be detected by simply monitoring the electrical impedance of compared to the imaginary part when detecting structural changes
the PZT transducer. The remaining variables , a , 33 T
, , d32x , and Ȳ xx
E
, [27]. The remaining variables N , Z̄ , and Z represent number of im­
stands for the input frequency, geometric constant, dielectric constant, pedance signature data, mean value and standard deviation, respec­
dielectric loss tangent, coupling constant and complex Young’s mod­ tively. In a previous study by Tseng and Naidu [28], it was shown that
ulus, respectively. RMSD and MAPD were more suitable for locating and characterizing
damage growth where CCD was more suitable for identifying damage
Zs ( ) size increase at a fixed location. Thus in this work, all three statistical
Y( )=i a T
33 (1 i ) d32x Y¯ xx
E
Zs ( ) + Za ( ) (1) metrics were used for designing the artificial neural networks algorithm
for this study.
In general, there are three different type of devices for measuring
the electrical impedance of the PZT transducer. The most widely used
RMSD = [Re (Zi ) Re (Zio )]2 / [Re (Zio )]2
device is using an impedance analyzer (eg. Agilent 4294a) which is N N (2)
relatively expensive. A lower version of measuring impedance was
proposed by Peairs et al [26] by connecting a function generator with
MAPD = |[Re (Zi ) Re (Zio)]/ Re (Zio )|
an oscilloscope and applying a simple circuit to approximate the elec­ N (3)
trical impedance of the attached PZT patch. The cost of performing the
EMI technique can be further reduced by using AD5933 evaluation 1
CCD = 1 CC, whereCC = [Re (Zi ) Re (Z¯ )].
board manufactured from Analog Devices. Co. The cost of the device is Z Zo N
under US$100 with the ability of measuring the impedance up to
100 kHz with more than 500 data points. The device is portable where [Re (Zio) ReZ¯ o )] (4)

2
W.S. Na Engineering Structures 226 (2021) 111228

2.3. Probabilistic neural networks work. However, although that loading effects may differ, they will
eventually result in bolt loosening (decreasing torque). Thus only the
Up to date, application of artificial neural networks in the field of torque value will be considered and investigated in this work.
civil engineering for detecting damage has been widely discussed by For the actual experiment, the bolt was tightened to specific values
various researchers. If one were to monitor a structure using the EMI using a digital torque wrench (model ETD210-AW42) purchased from a
technique, the surrounding environment such as rain, wind and load domestic company Torque World (http://www.torqueworld.com)
during service period may affect the impedance signature with time. shown in Fig. 3. The digital torque wrench allows the user to input a
Thus if one were to use artificial neural networks technique to detect desired torque value and when this value is reached during tightening,
damage of structures, retraining of the algorithm may be required over the wrench gives out a beeping sound notifying the user that the desired
a certain time period. For this reason, probabilistic neural networks torque value has been reached. However, one could argue that hand
(PNN) introduced by Donald F. Specht (1990) [29] was used for this tightening method of bolts may result in fluctuations in the actual
study. Compared to other artificial neural network methods, the torque value as the user heavily relies on beeping sound before stopping
training process is much faster than backpropagation. Also, it is guar­ (eg. tightening a bolt up to 30 N m and repeating this process will not
anteed to converge to an optimal classifier as training set size increases. always be exactly 30 N m every time). Since the EMI technique uses a
Furthermore, training samples can be added or removed without ex­ high frequency range in general, small fluctuations in torque values will
tensive retraining which makes PNN suitable for EMI technique on civil created variations in impedance signatures, possibly causing fault
infrastructures. Such PNN algorithm could also be used to automate the alarms. Thus considering this factor, the designed PNN algorithm for
damage detection process as it can be programmed to automatically this study will be evaluated. The purpose of this work includes ensuring
give decisions. In this study, an attempt to detect loosening of bolted repeatability of the results, applying the PNN algorithm and to increase
structures based on single training data set is investigated. In addition, the accuracy of the prediction results. Although only 9 bolts with lim­
torque (N m) predictions during bolt loosening are also conducted to ited torque values were used for this study, the results in the later
test the accuracy of the PNN algorithm developed for this study. Further section experimentally shows acceptable outcome for predicting the
details and evaluation of this algorithm will be discussed in the later torque values.
section.
3.2. Repeatability test using the digital torque wrench
3. Bolt loosening detection using the EMI method
To investigate variations in impedance signatures at same target
3.1. Test specimen torque using the digital torque wrench, a large bolt from one of the test
specimen was tightened to 30 N m to measure the impedance from
Fig. 2 shows that test specimen (bolted structure) used for this study 30 kHz to 90 kHz with 500 data points. Although that the torque value
where dimensions for each part is labelled in the figure. The specimens of 30 N m selected for this study can be said to be a small value, the
were specifically manufactured for this study from a domestic company main purpose of the study was to investigate the possibility of the
Partners Lab (http://www.partners-lab.com) using a stainless steel and finding the torque value using the PNN algorithm. In addition, a
standard size bolts. Three specimens with the same dimensions and method of increasing the accuracy can be found later in the study. Thus
properties were made with labels #1-1, #2-1 and #3-1 shown in the this torque value was adequate enough for the experiments carried out
figure. Two stainless metal plates of size 200 mm × 90 mm with 8 mm in this study. After acquiring the impedance signature, the bolt was
thickness was joined with the three different sized bolts. The top head completely loosened and tightened again to 30 N m to measure another
surface of all bolts were grinded to achieve a smooth flat surface for the impedance signature where this process was repeated until 5 im­
PZT patch to be attached using the epoxy adhesive. One needs to note pedance signatures were acquired. The results for this can be seen in
that for bolts on real structures are usually subjected to both tensile and Fig. 4 where variations in the signatures can be seen especially in the
shear stresses, which is different from the test specimen shown in this resonance frequency ranges between 50 kHz and 70 kHz. This

Fig. 2. Test specimen (bolted structure) with dimensions.

3
W.S. Na Engineering Structures 226 (2021) 111228

Fig. 3. Digital torque wrench (model: ETD210-AW42).

with 500 data points where this process was repeated 4 more times. By
observation, a large peak exist for the 0 N m impedance signature at
around 50 kHz. Such large peaks are only seen for 0 N m where the
remaining impedance signatures (10 –30 N m) have smaller peaks.
Such result is not surprising as 0 N m means complete loosening of the
bolt which leaves the bolt to vibrate freely when conducting the EMI
technique, resulting in resonance with high amplitude as seen in the
figure. In the figure, it can be seen that decrease in torque (bolt loos­
ening) causes the resonance peaks to shift in the left direction. The rest
of the tests (which were conducted 4 more times) showed similar trend
compared to Fig. 5. Thus the figures were not displayed in this study.
However, all data for the 5 tests were used for predicting the torque of
the bolt later in the study (Section 4.2). Regarding the 3 statistical
metrics (RMSD, MAPD and CCD) for Fig. 5, the calculated values when
the torque decreased to 20 N m from 30 N m were 2.017%, 1.323% and
0.288%, respectively. Furthermore, the values when the torque of the
bolt is decreased to 10 N m and 0 N m from 30 N m for the 3 statistical
metrics were 3.412%, 1.876%, 0.419% and 7.151%, 3.184%, 1.162%,
respectively. The RMSD values resulted in the highest values here
Fig. 4. Five impedance signatures measured at 30 N m. where one could say that this metric performs the best as a higher
threshold value can be defined to distinguish between a damaged cases
experimentally proves that the torque value is not exactly 30 N m every to an undamaged case.
time as the bolt was hand-tightened. To quantify signature variations
using the three statistical metrics (RMSD, MAPD and CCD), combina­ 3.3. Impedance signatures subjected to torque loss
tions of 2 impedance signatures from the 5 signatures were used
(5C2 = 10 combinations). From the calculated results, it was found that In this section, test specimens #1-1, #2-1 and #3-1 were used to
the RMSD values ranged from 0.387% to 1.904%, MAPD values from observe how the impedance signatures change subjected to bolt loos­
0.294% to 1.085% and CCD values from 0.202% to 0.266%. ening. Figs. 6 and 7 shows the impedance signatures with different
Fig. 5 shows another test regarding repeatability performance of the torque values for the three different size bolts. First looking at Fig. 6(a)
digital torque wrench. This time the same large bolt used for previous to (c) which are the impedance signatures from the large size bolts, the
test was used where it was tightened to 30 N m and then loosened 10 difference between different test specimens can be clearly seen. How­
N m each time until the bolt was completely loosened (0 N m) with ever, the resonance frequency ranges for all figures are the same where
impedance signatures being measured at each step. Here, impedance three resonance frequency regions can be found around 51 kHz, 62 kHz
signatures in the frequency range from 30 kHz to 90 kHz was measured and 79 kHz. For Fig. 6(a), we can see a large peak at 62 kHz which is
not seen for Fig. 6(b) and (c). This experimentally proves that even that
all three bolted structures were manufactured with the same dimen­
sions and properties, the impedance signature is highly sensitive to very
small changes that might have occurred during the manufacturing of
the test specimen. The stacked column graph Fig. 6(d) shows the three
statistical metrics calculated for Fig. 6(a)–(c) where the reference sig­
nature used was the impedance signature acquired at 30 N m. The first
three stacked columns (from left) represents the results calculated from
Fig. 6(a), the next three columns Fig. 6(b) and the last three columns
Fig. 6(c). Observing at the RMSD values for all three test specimens,
values increase as the torque value of the bolt decreases in general. This
can be seen for the first two test specimens #1-1 and #2-1. However,
for the test specimen #3-1 (Fig. 6(c) results), the RMSD drops from
3.96% to 3.52% although that the bolt was loosened from 20 N m to 10
N m. This is an unwanted outcome as such result can cause false alarms
when monitoring bolted structures using the EMI technique. Regard­
less, the RMSD values for all three test specimens when the bolts are
complete loose (0 N m) range from 6% to 7% which is a wanted out­
come as one could identify complete loosening of the bolt when the
RMSD value reaches this value.
Next examining the MAPD values, the first observation made is that
Fig. 5. Impedance signatures from 30 N m to 0 N m. the first two results (20 N m and 10 N m) for all three test specimens

4
W.S. Na Engineering Structures 226 (2021) 111228

Fig. 6. Impedance signatures subjected to bolt loosening for large size bolts: (a) large #1-1; (b) large #2-1; (c) large #3-1 and (d) calculated statistical metrics.

show very small difference in values. For #1-1 and #2-1 results, the Regarding Fig. 7(a)–(f) which shows the impedance signatures for
MAPD values at 20 N m is 1.32% and 1.42%, and at 10 N m is 1.88% medium and small size bolts, clear differences can be visually identified
and 1.82%, respectively. Thus one could setup an EMI technique based compared to the impedance signatures for the large size bolts (Fig. 6).
monitoring system where MAPD values around 1.32% ± 0.1% notifies First with the medium size bolt (Fig. 7(a)–(c)), the resonance frequency
the user that the torque is 20 N m and 1.88 ± 0.1% is 10 N m if these ranges are similar compared to the large size bolts where resonance is
results were reliable. However, the results for #3-1 is quite different located at around 55 kHz, 65 kHz and 80 kHz. Regarding the resonance
compared to #1-1 and #2-1 as the MAPD values are 1.84% and 2.01% at 80 kHz for Fig. 7(a), the size of the resonance (amplitudes of the
at 20 N m and 10 N m, respectively. This can complicate the damage impedance signatures) is noticeably small compared to the resonance
identifying process when predicting torque as values around 1.8% can shown in Fig. 7(b). With such large difference in signatures, one may
mean either 20 N m or 10 N m. For the MAPD values at 0 N m for all think that these two set of impedance sigantures are from two test
three specimens, they are 3.18%, 4.19% and 3.46% where one could specimens with completely different properties. One of the possible
define a threshold value of 3% to identify complete loose of the large explanation of this large difference could be that the resonance in this
size bolt (eg. MAPD resulting in a values over 3% is 0 N m). Finally, region is highly sensitive to the bolt itself rather than change in torque.
observing at the CCD values, loosening of the bolt from 30 N m down to In other words, one of the reasons for the signature difference could be
0 N m results in very low values overall where the difference is difficult due to the very small difference in bolt dimension during the manu­
to visually recognize. The CCD values from 20 N m to 0 N m for #1-1 facturng process as standard size bolts have dimension error ranges.
were 0.29%, 0.42% and 1.16%, increasing with torque loss as expected. However, one could argue that the difference in resonance size between
For #2-1, the values were 0.29%, 0.35% and 1.04%, and for #3-1, the the two specimen could be caused from difference in the adhesive
values were 0.49%, 0.43% and 0.92% proving that the CCD values do bonding layers. But if we were to observe the resonance at 65 kHz for
not always increase with torque loss. In addition, due to the low values both Fig. 7(a) and (b), the size of resonance is quite the same between
compared to the other two statistical metrics (RMSD and MAPD), this the two figures, experimentally proving that the adhesive layer has no
experimentally proves for this study that CCD metric is not the best problems. So for these medium size bolts, it would be ideal to avoid
choice when identifying loosening of bolted structures. analyzing the resonance region at 80 kHz to minimize false alarms

5
W.S. Na Engineering Structures 226 (2021) 111228

Fig. 7. Impedance signatures subjected to bolt loosening for three medium and three small size bolts: (a) medium #1-1; (b) medium #2-1; (c) medium #3-1; (d) small
#1-1; (e) small #2-1; and (f) small #3-1.

when predicting torque loss. Next regarding the results for the small likely that such signatures will not detect damage as successful as the
size bolts (Fig. 7(d)–(f)), resonance with a multiple number of im­ large and medium size bolts due to absence of multiple number of
pedance peaks are difficult to observe. Although that a small resonance peaks.
is found near 90 kHz for all the bolts from all specimens, it is highly To evaluate the performance of the EMI technique for Fig. 7(a)–(f),

6
W.S. Na Engineering Structures 226 (2021) 111228

Fig. 8. Statistical metrics calculated for: (a) medium size bolts and (b) small size bolts.

the three statistical metrics were calculated and plotted as stacked either 20 N m, 10 N m or 0 N m (complete loosening).
column graph in Fig. 8(a) and (b). First observing at the results for the
medium size bolts shown in Fig. 8(a), results for #1-1 is quite promising 4.2. Torque prediction subjected to repeatability
as all RMSD, MAPD and CCD values increase with decrease in torque.
The results for #2-1 show similar trend with the values at 0 N m being In Section 3.2, bolt loosening test carried out on a large size bolt 5
quite high (17.4% for RMSD, 8.37% for MAPD and 6.11% for CCD). The times by measuring impedance signatures from 30 N m to 0 N m. Here,
results for #3-1 displays the lowest values when compared to #1-1 and the first test was inserted into the PNN algorithm to set up the pattern
#2-1 specimens. Here, it can be seen that the three statistical metric layer (training set) and rest of the 4 tests were used to evaluate the
values at 20 N m and 10 N m are very close to the results from #1-1 but performance of the algorithm where all data used for this process is
the values at 0 N m shows large difference. The values for #3-1 is very shown in table 1. The numbers in the brackets represent the torque
small with RMSD, MAPD and CCD values resulting in 2.05%, 0.75%, value where the impedance signature obtained from this torque was
and 0.28%, respectively (compared to the #1-1 values of 9.32%, 4.12%, used as the corresponding signature. Thus “RMSD(20)” values were
and 1.80%). Although that this is an unwanted outcome as this in­ calculated using the impedance signature acquired at 20 N m against
dicates that the impedance signatures at 30 N m and 0 N m have small the reference signature at 30 N m. Rest of the values were calculated
difference, it would be interesting to see how the PNN algorithm per­ using the reference signature as the impedance signature at 30 N m for
forms under such circumstance. Observing the results for Fig. 8(b), the each test. Using the first test as the training data (the first column from
metric values generally increase with decrease in torque for all speci­ the table), the predictions for the 4 tests show overall accuracy of
mens. In addition, results between the specimens do not show large 91.7% (11/12) with only the second test failing to achieve 100% (2/3)
difference except for the metric values for the #3-1 where it is relatively correct prediction. Looking at the table, the calculated values using the
low compared to the other two specimens (#1-1 and #2-1). impedance signatures of 20 N m and 10 N m are quite similar (eg.
“RMSD(20)” values are close to “RMSD(10)” values) and this is one of
4. Application of the PNN the reasons for the wrong prediction for the second test. Although it is
not shown in the table, the wrong prediction was made for the first
4.1. Training set prediction where the PNN algorithm predicted the bolt torque to be 10
N m when the actual torque was 20 N m. From the PNN algorithm, the
Fig. 9 depicts the classification algorithm with 3 categories each difference between the two values at the summation layer resulted in
representing the torque values for the bolts (S1 = 20 N m, S2 = 10 N m very small difference as the summed values were 2.94 for S1 and 2.95
and S3 = 0 N m). The PNN algorithm is designed to use one impedance for S2. Thus this experimentally proves that the PNN algorithm can
signature input where each category contains 3 units (R11, M12, and C13 predict the torque with very high accuracy of 91.7% for the same bolt
etc). Using this single input, the three metrics RMSD, MAPD and CCD (specimen).
are calculated for each category using the impedance signature at 30
N m as the reference signature (treated as undamaged bolt for this 4.3. Torque prediction subjected to bolt loosening
study). Here, each trained data point corresponds to the unit which is a
Gaussian function with a peak centered on the parameters location. The In Section 3.3, bolt loosening tests were carried out for the three
variables x 0 , y0 , and z 0 are the trained data calculated from the RMSD, different bolted structures #1-1, #2-1 and #3-1. Although that the
MAPD and CCD equations, respectively. x , y , and z are the input three test specimens were manufactured with the exact same dimen­
parameters from a different impedance signature for predicting torque sions and properties, it was visually confirmed by previous figures that
values. is a constant that controls the width of the function that re­ the impedance signatures subjected to bolt loosening had differences.
quires an educated guess based on the knowledge of the data when The resonance frequency ranges were virtually identical between spe­
determining the value. Thus for this study, of 0.01 was used for all cimens but the shapes were noticeably different. Taking this in mind,
tests. At completion of the training process, the PNN algorithm is ready the goal of this section was to use impedance signatures to train the
for predicting bolt torque values where just one impedance signature PNN algorithm and then predict bolt loosening of a different specimen
measured from 30 kHz to 90 kHz is required for the prediction. The (eg. using the training data from #1-1 to predict the torque values of
algorithm then goes through the pattern layer where the values are #2-1 and #3-1). As previously mentioned in the introduction section,
summed to create a summed category unit (summation layer). The the reason for this is that for large structures, there can be many bolted
maximum value from this layer is then selected to predict the torque of parts and it would be extremely difficult to acquire training data for all

7
W.S. Na Engineering Structures 226 (2021) 111228

Fig. 9. Probabilistic neural networks algorithm designed for this study.

Table 1 N m. This is not surprising as we have used a training data from a

PNN input data for repeatability test. different specimen (#1-1) to predict the torque for a bolt in #2-1. In
1st 2nd 3rd 4th 5th
addition, the statistical metric values for the medium size bolt are re­
latively high compared to the results for #3-1, which is the main cause
RMSD(20) 2.50 2.73 2.38 2.32 2.25 of wrong prediction. Overall for this specimen, 7 correct predictions
MAPD(20) 1.52 1.60 1.42 1.41 1.33 were made out of 9 cases resulting in 77.8% accuracy of the PNN al­
CCD(20) 0.32 0.34 0.31 0.30 0.29
RMSD(10) 2.81 2.63 2.91 2.71 2.71
gorithm. Next with the #3-1 specimen, the correct torque value pre­
MAPD(10) 1.82 1.66 1.85 1.86 1.84 dictions for the large, medium, small size bolts were 2/3 for all pre­
CCD(10) 0.35 0.33 0.36 0.34 0.34 dictions. Thus 6 correct predictions were made out of 9 cases resulting
RMSD(0) 7.47 8.32 7.76 8.50 7.52 in 66.7% accuracy of the PNN algorithm. Here, wrong predictions were
MAPD(0) 3.57 3.75 3.76 3.76 3.68
made subjected to 20 N m, 0 N m and 0 N m for large, medium and
CCD(0) 1.24 1.50 1.32 1.54 1.25
Predictions 2/3 3/3 3/3 3/3 small size bolts, respectively.
Overall, the prediction results for both #2-1 and #3-1 look quite
promising as the 13 correct predictions were made out of 18 cases re­
parts. For this reason this study investigates the possibility of using just sulting in 72.2% accuracy. However, one can say that this accuracy is
one part of the bolted area to represent other areas. not high enough to be reliable to be used for real field applications.
Taking #1-1 impedance signatures to train the PNN algorithm, the Thus in the next section, an idea of increasing the accuracy is propose to
calculated input data and prediction results for torque values are dis­ further evaluate and improve the performance of the PNN algorithm.
played in table 2. The last row shows the number of correct predictions
for the two specimens #2-1 and #3-1. First with the #2-1 results, the
predictions for the large and small size bolts show 100% (3/3) accuracy 5. Improving the accuracy of the PNN algorithm
correctly identifying the torque. The medium size bolt shows 33.3% (1/
3) accuracy where the PNN algorithm predicted all the torques to be 0 In the previous section, the overall accuracy of the PNN algorithm
for predicting torque values for #2-1 and #3-1 was 72.2%. Thus in this
Table 2
part, an attempt to increase accuracy of the prediction is investigated.
PNN input data for bolt loosening test. Fig. 10(a) shows 3 impedance signatures from medium size bolts of #1-
1 at 20 N m, #3-1 at 0 N m and 20 N m (the reasons for choosing these
Large bolts Medium bolts Small bolts
signatures will be explained later in this section). As shown earlier,
#1-1 #2-1 #3-1 #1-1 #2-1 #3-1 #1-1 #2-1 #3-1 although that the resonance frequency ranges were the same, the shape
of the impedance signatures were generally different with many small
RMSD(20) 2.17 2.26 3.96 5.74 9.44 5.31 1.26 2.59 1.76 and large random peaks. One could argue that these small random
MAPD(20) 1.32 1.42 1.84 1.96 3.33 1.95 1.01 1.40 1.13
peaks can be treated as noise and eliminating these could possibly
CCD(20) 0.29 0.30 0.49 0.81 1.97 0.73 0.22 0.32 0.25
RMSD(10) 3.41 2.82 3.52 7.12 10.16 7.06 2.21 3.59 2.00 improve accuracy of the PNN algorithm. Thus to minimize this dif­
MAPD(10) 1.88 1.82 2.10 2.96 3.99 2.87 1.54 2.04 1.58 ferent, a smoothing technique known as centered moving average
CCD(10) 0.42 0.35 0.43 1.14 2.25 1.14 0.28 0.43 0.27 method was applied by averaging 20 data points. The result for this can
RMSD(0) 7.15 6.72 6.20 9.32 17.43 2.05 7.99 11.07 4.19 be seen in Fig. 10(b) where smoothing the signature resulted in elim­
MAPD(0) 3.18 4.19 3.46 4.12 8.37 0.75 3.98 5.10 2.94
inating small random peaks observed in Fig. 10(a). So to further in­
CCD(0) 1.16 1.04 0.92 1.80 6.11 0.28 1.34 2.43 0.51
Predictions 3/3 2/3 1/3 2/3 3/3 2/3 vestigate the effect of smoothing the signature on the accuracy of the
PNN algorithm, centered moving average technique was applied to all

8
W.S. Na Engineering Structures 226 (2021) 111228

applications. However, by averaged the data with 40 and 80 points, the

accuracy of the PNN algorithm decreases to 15/18 (83.3%) and 13/18
(72.2%) proving that there is an optimum value when choosing the
number of data points to be averaged.
From table 3, averaging 20 data points seems to be the best choice
for the experiment. With 94.4% accuracy, one could hope for a 100%
accuracy by further analyzing the problem and overcoming it.
Fig. 10(b) shows why achieving 100% for this study is difficult. One of
the training data used for the PNN algorithm was the signature at 20
N m from the #1-1 specimen and the prediction for #3-1 at 20 N m
failed to correctly identify the actual torque. The algorithm predicted 0
N m at this stage. Fig. 10(b) shows why this happened as impedance
signature at 0 N m is close to the training data compared to the im­
pedance signature at 20 N m. Thus averaging impedance signature data
will not overcome this problem and another solution is required which
will need further investigation in the future.

6. Conclusions

In this study, electromechanical impedance (EMI) technique was

combined with probabilistic neural networks (PNN) technique to detect
bolt loss of bolted structure test specimens. Three test specimens were
made with the same dimensions and properties where each specimen
had three different sized bolts. The first part of the study was to observe
how the impedance signature changed subjected to torque values from
30 N m to 0 N m in 10 N m step and investigate the difference between
other specimens. The results show that also that the three test speci­
mens were created with the same dimensions and properties, the shape
of the impedance signatures varied. However, the resonance frequency
ranges were virtually identical between specimens. The second part of
the study was to set up the PNN algorithm for predicting torque values
subjected to change in torque. Here, only the training data from #1-1
was used to predict the torque values of #2-1 and #3-1. Regardless of
the differences in impedance signatures between the test specimens, the
PNN algorithm prediction accuracy was 77.8% for #2-1 and 66.7% for
#3-1 subjected to torque of 20 N m, 10 N m and 0 N m. Thus overall
accuracy of 72.2% was achieved with the PNN algorithm. The last part
of the study proposed an easy way of increasing the accuracy for pre­
dicting the torque. Using centered mean averaging method for
smoothing the impedance signature, this eliminated small random
Fig. 10. (a) Impedance signatures compared to a reference signature and (b) peaks that existed for most of the impedance signatures. With these
smoothed impedance signature. signatures, the PNN algorithm was evaluated again to achieve a pre­
diction accuracy of 94.4% by averaging 20 data points. Increasing the
Table 3
number of data point for averaging resulted in decrease in prediction
Prediction results with the application of the centered mean averaging method. accuracy back to 72.2% proving that an optimum value exist when
selecting the number of data points to be averaged.
Large bolts Medium bolts Small bolts
Although limited number of specimens with torque values were
#2-1 #3-1 #2-1 #3-1 #2-1 #3-1 Predictions used for this study, the results were promising where application of the
PNN algorithm has been experimentally proved in this study to be ef­
0 averaged 3/3 2/3 1/3 2/3 3/3 2/3 13/18 (72.2%) fective. Furthermore, increasing the prediction accuracy with the cen­
20 averaged 3/3 3/3 3/3 2/3 3/3 3/3 17/18 (94.4%)
tered mean averaging method improved the prediction outcome. Thus
40 averaged 3/3 3/3 2/3 2/3 3/3 2/3 15/18 (83.3%)
80 averaged 3/3 3/3 2/3 1/3 3/3 1/3 13/18 (72.2%)
in a future study, the goal of achieving an accuracy of around 100% will
be conducted with various bolt sizes and torque values to bring this idea
one step closer for real applications.
signatures by averaging 20, 40 and 80 data points where the final
prediction results are shown in table 3. The first row represents the CRediT authorship contribution statement
result for the previous section where 72.2% (13/18) accuracy was
calculated. The rest of the rows are results from averaging the im­ Wongi S. Na: Conceptualization, Methodology, Software,
pedance signatures by 20, 40 and 80 data points. The last column re­ Validation, Resources, Data curation, Writing - original draft, Writing -
presents the overall accuracy performance of the PNN algorithm con­ review & editing, Visualization, Supervision, Project administration,
ducted using the averaged signatures. The best performance was Funding acquisition.
produced when the impedance signatures were averaged by 20 data
points where 17 correct predictions were made out of 18 cases, which is
Declaration of Competing Interest
94.4%. With such high accuracy of the result, one could say that this
work brings the EMI technique a one step closer for real field
The authors declare that they have no known competing financial

9
W.S. Na Engineering Structures 226 (2021) 111228

interests or personal relationships that could have appeared to influ­ [14] Huo L, Chen D, Liang Y, Li H, Feng X, Song G. Impedance based bolt pre-load
ence the work reported in this paper. monitoring using piezoceramic smart washer. Smart Mater Struct
2017;26(5):057004.
[15] Ritdumrongkul S, Abe M, Fujino Y, Miyashita T. Quantitative health monitoring of
References bolted joints using a piezoceramic actuator–sensor. Smart Mater Struct
2003;13(1):20.
[16] Lim YY, Bhalla S, Soh CK. Structural identification and damage diagnosis using self-
[1] Amerini F, Meo M. Structural health monitoring of bolted joints using linear and
sensing piezo-impedance transducers. Smart Mater Struct 2006;15(4):987.
nonlinear acoustic/ultrasound methods. Struct Health Monit 2011;10(6):659–72.
[17] Fan S, Li W, Kong Q, Feng Q, Song G. Monitoring of pin connection loosening using
[2] Guarino J, Hamilton R, Fischer W. Acoustic detection of bolt detorquing in struc­
eletromechanical impedance: Numerical simulation with experimental verification.
tures. In Proceedings of Meetings on Acoustics 157ASA Acoustical Society of
J Intell Mater Syst Struct 2018;29(9):1964–73.
America 2009;6(1): 065002.
[18] Li W, Liu T, Zou D, Wang J, Yi TH. PZT based smart corrosion coupon using elec­
[3] Huang YH, Liu L, Yeung TW, Hung YY. Real-time monitoring of clamping force of a
tromechanical impedance. Mech Syst Sig Process 2019;129:455–69.
bolted joint by use of automatic digital image correlation. Opt Laser Technol
[19] Li W, Wang J, Liu T, Luo M. Electromechanical impedance instrumented circular
2009;41(4):408–14.
piezoelectric-metal transducer for corrosion monitoring: modeling and validation.
[4] Khomenko A, Koricho EG, Haq M, Cloud GL. Bolt tension monitoring with reusable
Smart Mater Struct 2020;29(3):035008.
fiber Bragg-grating sensors. J Strain Anal Eng Des 2016;51(2):101–8.
[20] Chen R, Chen S, Yang L, Wang J, Xu X, Luo T. Looseness diagnosis method for
[5] Yeager M, Whitaker A, Todd M. A method for monitoring bolt torque in a composite
connecting bolt of fan foundation based on sensitive mixed-domain features of
connection using an embedded fiber Bragg grating sensor. J Intell Mater Syst Struct
excitation-response and manifold learning. Neurocomputing. 2017;219:376–88.
2018;29(3):335–44.
[21] Cha YJ, You K, Choi W. Vision-based detection of loosened bolts using the Hough
[6] Chen D, Huo L, Li H, Song G. A Fiber Bragg Grating (FBG)-Enabled Smart Washer for
transform and support vector machines. Autom Constr 2016;71:181–8.
Bolt Pre-Load Measurement: Design, Analysis. Calibr Experim Valid
[22] Zhao X, Zhang Y, Wang N. Bolt loosening angle detection technology using deep
2018;18(8):2586.
learning. Struct Control Health Monit 2019;26(1):e2292.
[7] Ho SCM, Li W, Wang B, Song G. A load measuring anchor plate for rock bolt using
[23] Nguyen TC, Huynh TC, Ryu JY, Park JH, Kim JT. Bolt-loosening identification of
fiber optic sensor. Smart Mater Struct 2017;26(5):057003.
bolt connections by vision image-based technique. In Nondestructive
[8] Argatov I, Sevostianov I. Health monitoring of bolted joints via electrical con­
Characterization and Monitoring of Advanced Materials, Aerospace, and Civil
ductivity measurements. Int J Eng Sci 2010;48(10):874–87.
Infrastructure, vol. 9804. International Society for Optics and Photonics; 2016, p.
[9] Huda F, Kajiwara I, Hosoya N, Kawamura S. Bolt loosening analysis and diagnosis
980413.
by non-contact laser excitation vibration tests. Mech Syst Sig Process
[24] Ramana L, Choi W, Cha YJ. Fully automated vision-based loosened bolt detection
2013;40(2):589–604.
using the Viola-Jones algorithm. Struct Health Monit 2019;18(2):422–34.
[10] Wang F, Huo L, Song G. A piezoelectric active sensing method for quantitative
[25] Liang C, Sun FP, Rogers CA. Coupled electro-mechanical analysis of adaptive ma­
monitoring of bolt loosening using energy dissipation caused by tangential damping
terial systems—determination of the actuator power consumption and system en­
based on the fractal contact theory. Smart Mater Struct. 2017;27(1):015023.
ergy transfer. J Intell Mater Syst Struct 1994;5(1):12–20.
[11] Wang F, Ho SCM, Song G. Monitoring of early looseness of multi-bolt connection: a
[26] Peairs DM, Park G, Inman DJ. Improving accessibility of the impedance-based
new entropy-based active sensing method without saturation. Smart Mater Struct
structural health monitoring method. J Intell Mater Syst Struct 2004;15:129–39.
2019;28(10):10LT01..
[27] Sun FP, Chaudhry Z, Liang C, Rogers CA. Truss structure integrity identification
[12] Wang F, Chen Z, Song G. Monitoring of multi-bolt connection looseness using en­
using PZT sensor–actuator. J Intell Mater Syst Struct 1995;6(1):134–9.
tropy-based active sensing and genetic algorithm-based least square support vector
[28] Tseng KK, Naidu ASK. Non-parametric damage detection and characterization using
machine. Mech Syst Sig Process 2020;136:106507.
smart piezoceramic material. Smart Mater Struct 2002;11:317.
[13] He K, Zhu WD. Detecting loosening of bolted connections in a pipeline using
[29] Specht DF. Probabilistic neural networks. Neural Networks. 1990;3(1):109–18.
changes in natural frequencies. J Vib Acoust 2014;136(3):034503.

10