R. DVOØÁK et al.: NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A MACHINE LEARNING APPROACH
561–565
NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A
MACHINE LEARNING APPROACH
NEPORU[NO TESTIRANJE CIPP NAPAK Z UPORABO PRISTOPA
STROJNEGA U^ENJA
Richard Dvoøák
*
, Lubo{ Pazdera, Libor Topoláø, Lubo{ Jakubka, Jan Puchýø
Brno University of Technology, Veveøí 331/95, Brno 602 00, Czech Republic
Prejem rokopisa – received: 2023-09-15; sprejem za objavo – accepted for publication: 2024-07-26
doi:10.17222/mit.2023.1000
In civil engineering, retrofitting actions involving repairs to pipes inside buildings and in extravehicular locations present com-
plex and challenging tasks. Traditional repair procedures typically involve disassembling the surrounding structure, leading to
technological pauses and potential work environment disruptions. An alternative approach to these procedures uses the
cured-in-place-pipe (CIPP) technology for repairs. Unlike standard repairs, CIPP repairs do not require a disassembly of the sur-
rounding structures; only the access points at the beginning and end of the pipe need to be accessible. However, this method in-
troduces the possibility of different types of defects.
1
This research aims to observe the defects between the host and newly
cured pipes. The presence of holes, cracks, or obstacles prevents achieving a desired close-fit state, ultimately reducing the life
expectancy of the retrofitting. This paper focuses on the non-destructive observation of these defects using the non-destructive
testing (NDT) impact-echo (IE) method. The study explicitly applies this method to the composite segments inside concrete host
pipes, forming a testing polygon. Previous results have indicated that the mechanical behaviour of cured composite pipes can
vary in stiffness depending on factors such as the curing procedure and environmental conditions.
2
The change in acoustic pa-
rameters such as resonance frequency, attenuation and other features of typical IE signals can describe the stiffness evolution.
This study compares different sensors used for the proposed IE testing, namely piezoceramic and microphone sensors. It evalu-
ates their ability to distinguish between the defects present in the body of a CIPP via a machine-learning approach using random
tree classifiers.
Keywords: retrofitting, cured-in-place pipes, non-destructive testing, impact-echo method, pipe defects, acoustic parameters,
machine learning, classification
Popravila in modifikacije (angl.: retrofitting actions) starih cevovodov v `e izdelanih zgradbah in na posebej obremenjenih
zunanjih vozli{~ih predstavlja poseben in zahteven gradbeni{ki poseg. Obi~ajni ali standardni postopki popravljanja cevovodov
so sestavljeni iz celotnega razstavljanja okoli{ke strukture. To vodi do tehnolo{kih premorov in potencialno tudi do ne`elenih
motenj v okolju. Alternativa k tem pristopom je uporaba tehnologije oz. postopka popravila cevovoda na licu mesta (CIPP;
angl.: cured-in-place pipes). Za razliko od standardnega postopka popravila cevovodov, CIPP ne zahteva celotne demonta`e
okoli{ke strukture. Potrebno je zagotoviti dostop do za~etka in konca cevovoda. Vendar pa s to metodo lahko popravimo le
dolo~ene vrste napak na cevovodih.
1
V ~lanku avtorji opisujejo opazovanje napak med obstoje~imi in na novo popravljenimi
cevovodi. Vendar pa prisotnost lukenj, razpok ali ovir prepre~uje, da bi dosegli popolen dostop, ki bi omogo~al idealen postopek
popravila. To pa dokon~no zmanj{uje pri~akovano dobo trajanja popravljenega cevovoda. V tem ~lanku se avtorji osredoto~ajo
na neporu{na opazovanja napak v cevovodih s pomo~jo NDT (angl.: non-destructive testing) udarno-odbojne zvo~ne metode
(IE; angl.: Impact-Echo method). Pri tej metodi mehanski udarci povzro~ajo valovanje zvoka skozi medij, ki je moteno zaradi
prisotnih napak v materialu. V pri~ujo~i {tudiji so avtorji to metodo uporabili direktno za testiranje posameznih segmentov
kompozitnih CIPP cevovodov. Predhodni rezultati raziskav
2
so pokazali, da je mehansko obna{anje oziroma togost popravljenih
CIPP kompozitnih cevovodov odvisno od izbranega postopka popravila in okoljskih pogojev. Sprememba akusti~nih parametrov
kot so resonan~na frekvenca, atenuacija (pojemanje) in druge karakteristike tipi~nih IE signalov lahko opi{ejo razvoj togosti. V
~lanku avtorji opisujejo primerjavo med razli~nimi senzorji (piezokerami~nimi in mikrofonskimi), ki so jih uporabili za
predlagan na~in IE testiranja. S tem so ovrednotili njihovo sposobnost razlikovanja med prisotnimi napakami v jedru CIPP s
pomo~jo pristopa strojnega u~enja in z uporabo treh naklju~nih klasifikatorjev.
Klju~ne besede: retrofiting, obdelava oziroma popravilo cevovodov na licu mesta, neporu{no testiranje, vpliv metode "mehanski
udarec in odboj zvoka", napake na cevovodih, akusti~ni parametri, strojno u~enje, klasifikacija
1 INTRODUCTION
The IE method has found wide application in the
construction industry due to its simplicity, low cost of
implementation and relatively wide range of applica-
tions.
3
At the same time, the IE method is dependent on
the correct interpretation of the measured data. In prac-
tice, it has been found to be applicable in measuring pile
lengths, localizing cracks in massive monolithic struc-
tures, delaminating bridge bodies, diagnosing the condi-
tion of concrete elements, etc. Due to its simple principle
of testing, there are many variations of the IE method,
for example, the low-frequency pulse-echo method.
So far, however, the IE method has not found signifi-
cant application in measuring defects in sewer lines and
product pipelines. These are linear structures made of
concrete, ceramics or steel, which can, in principle, be
measured using similar methods to those commonly used
for structural concrete. For example, a typical concrete
oven may be designed for type XF4 (highly saturated
Materiali in tehnologije / Materials and technology 58 (2024) 5, 561–565 561
UDK 532.542:608.32:001.891.55 ISSN 1580-2949
Original scientific article/Izvirni znanstveni ~lanek MTAEC9, 58(5)561(2024)
*Corresponding author's e-mail:
richard.dvorak@vutbr.cz (Richard Dvoøák)

with de-icing agents or seawater) and XA1 (weakly ag-
gressive chemical environments). These ovens com-
monly suffer from degradation due to abrasion, chemical
exposure, mechanical stresses, wear and tear, and need to
be repaired over time. This can be done by excavating
and replacing the entire pipe or with a trenchless tech-
nology. The CIPP method can be used in this area.
4
During this process, the host pipe is lined with a
resin-saturated insert forced into the pipe by water or air
pressure. This liner is then cured with elevated tempera-
ture (above 60 °C) using air or water. The high pressure
of the medium also allows the liner to make close contact
with the host pipe. The elevated temperature starts the
polymerization reaction, and the liner is cured, forming a
new pipe profile.
However, due to the nature of the curing process, de-
fects may occur. For example, in large diameter pipes,
resin washout may happen in the event of high water
pressure in the ground body. Caverns and longitudinal
thin cavities are formed in some areas of the pipe, reduc-
ing the pipe’s overall life.
Depending on the type of pipeline, holes will be cut
into the branch and connection pipelines using a manual
or robotic cutter. In some cases, however, not all of the
original connections are connected, creating blind con-
nections and, therefore, a cavity behind them.
The condition of pipelines is currently most often as-
sessed visually using camera surveys, and it must be
emphasised that the poor condition of a pipeline can only
be recognised when the degradation is intolerable. At se-
lected locations, for example, semi-destructive tear tests
can be carried out with some reliability, but often only to
confirm the poor condition of the pipeline. Other tools
are not used in the domestic market to assess the extent
of pipeline degradation or locate this degradation. From
this perspective, it would be helpful to have a tool that
can determine the state of degradation of a host pipeline
or defects in a newly repaired pipeline, and monitor se-
lected sections over time to assess the evolution of their
conditions.
2 EXPERIMENTAL PART
As part of field measurements on two pipe liners at
the test polygon, using the KAWO liner in collaboration
with company WOMBAT s.r.o., a test measurement was
carried out using the impact-echo method and a micro-
phone. This measurement was focused on trying to rec-
ognize different types of common defects on the CIPP
liner. The testing polygon with all simulated defects is
shown in Figure 1. The whole polygon is shown in Fig-
ure 1a. The KAWO liner tightly fitted to the host pipe
(referred to as the "wall") is shown in Figure 1b. The
cavern defect behind the KAWO liner is presented in
Figure 1c. When there is a big thin gap between the host
pipe and the liner, the defect is defined as a free-end,
shown in Figure 1d. Defects are defined based on retro-
fitting practices. The cavity is a typical defect, where the
space behind the liner wall allows the accumulation of
debris. The free-end represents the state where there is a
thin space between the host pipe and the liner and it is a
specific type of cavern defect. The whole polygon was
14 m long, the inner height was 1350 mm, and the width
of the largest span was 900 mm.
The microphone was hand-held, and the hammer im-
pacts were done from a 100-mm distance. Each signal
was composed of inputs from the microphone and the
force sensor.
The longitudinal axis of the measured pipe was cho-
sen along different parts of the estimated profile. The
axis is shown in Figure 2 and Figure 3. It was deter-
mined according to previous measurements on the free
part of the cured liner, accessible from the inside and
outside between the concrete tubes. A broadband piezo-
ceramic transducer measured the natural resonant fre-
quency on these parts. The frequency at the estimated
point E was 470 Hz. At points C and G, it was 731 Hz; at
the upper part, at point A, the resonant frequency was
789 Hz. This insert part is not dampened against self-vi-
bration, and its natural resonant frequency is measurable.
However, this situation is artificially created by the de-
sign of the test polygon.
R. DVOØÁK et al.: NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A MACHINE LEARNING APPROACH
562 Materiali in tehnologije / Materials and technology 58 (2024) 5, 561–565
Figure 1: Illustration of tested materials and structures: a) testing polygon with concrete sewage pipes and a cured-in-place pipe; b) representa-
tion of the close-fit state; b) cavern behind the polymer pipe; c) free-end without a close-fit state

A MEMS microphone ADMP401 and a modal ham-
mer with a built-in ring force sensor for exciting signals
were used to acquire signals. The signals were recorded
by a Handyscope HS3 digital oscilloscope with 16-bit
resolution and sampling frequency of 193 kHz. The re-
corded data were processed using the MATLAB soft-
ware, namely Machine Learning Toolbox, Statistics, and
Feature Extraction Toolbox, and for continuous wavelet
transforms, a fast wavelet library was used.
5
For the feature extraction, a Predictive Maintenance
Toolbox was used, which allows us to design functions
for extracting features from both signal and frequency
spectra. This toolbox is designed to create a feature ex-
traction algorithm, which focuses on several acoustic
metrics.
The root mean square (RMS) is a statistical measure
used in acoustics to quantify the magnitude of a varying
signal. It is particularly useful for measuring the average
power of an audio signal over time. By squaring the sig-
nal values, averaging them, and then taking the square
root, the RMS provides a single value that represents a
signal’s overall energy level. This metric is crucial in as-
sessing the sound intensity and ensuring consistent audio
levels in various applications, from music production to
noise monitoring.
6
The signal-to-noise ratio (SNR) is a key metric in
acoustics that quantifies the relationship between the de-
sired signal and the background noise. Expressed in
decibels (dB), the SNR measures how much a signal
stands out from the noise. A higher SNR indicates a
clearer, more distinguishable signal, essential for high-fi-
delity audio reproduction, speech intelligibility, and ef-
fective communication systems. Improving the SNR is a
primary goal in audio engineering, leading to better
sound quality and listener experience.
7
To further support the differences between the signals
of all three classes, band power and signal frequency are
also extracted from the signals. For training the models,
the whole set of variables, supported by the Predictive
Maintenance Toolbox was used: Clearance Factor, Sen-
sor Crest Factor, Sensor Impulse Factor, Sensor Kurtosis,
Sensor Mean, Sensor Peak Value, Root Mean Square
(RMS), Sensor SINAD, Signal-to-Noise Ratio (SNR),
Sensor Shape Factor, Sensor Skewness, Sensor Standard
Deviation (Std), Sensor Total Harmonic Distortion
(THD), Sensor Peak Amplitude 1 (PS), Sensor Peak Fre-
quency 1 (PS), Sensor Band Power (PS), Sensor Peak
Amplitude 1 (PS Spec 1), Sensor Peak Amplitude 2 (PS
Spec 1), Sensor Peak Amplitude 3 (PS Spec 1), Sensor
Peak Amplitude 4 (PS Spec 1), Sensor Peak Amplitude 5
(PS Spec 1), Frequency Peak, Sensor Peak Frequency 2
(PS Spec 1), Sensor Peak Frequency 3 (PS Spec 1), Sen-
sor Peak Frequency 4 (PS Spec 1), Sensor Peak Fre-
quency 5 (PS Spec 1), Band Power (BP).
To illustrate the variables with the highest relevance
for the classification, the Band Power, Frequency, SNR
and RMS are selected for the visualization of results.
R. DVOØÁK et al.: NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A MACHINE LEARNING APPROACH
Materiali in tehnologije / Materials and technology 58 (2024) 5, 561–565 563
Figure 2: Set-up for axis testing and the used measuring devices: a) testing points at the first cross-section of the CIPP; b) microphone used for
recording; c) hammer with a piezoelectric ring for signal excitation
Figure 3: Image from the inside of the pipe, with a testing grid of a
semi-section of the tube. The light is coming through the A4 testing
point, where a šcavern’ type of defect is present

3 RESULTS
In total, 379 signals were recorded, which consisted
of 196 close-fit signals, 76 free-end signals, and 107 cav-
ern signals. From the recorded signals, 27 variables were
extracted using the Feature Extraction Toolbox. These
features were then assessed with 4-way ANOVA to find
the variables with the highest overall variance. An exam-
ple of representative signals is presented in Figure 4a.
From the time-frequency domain, the main differences
between different states of the CIPP composite classes
can be recognized.
The healthy close-fit has a strong resonance response
at 1 kHz. The cavern can be distinguished by a lower
RMS and higher signal-to-noise ratio, where the domi-
nant frequency is scattered around 0.8–1.4 kHz. The
free-end is most distinguishable with the highest RMS,
band power and lowest signal-to-noise ratio. This corre-
sponds to a fairly stable oscillation when excited by an
impact hammer, with consistent response frequencies at
0.65 kHz and 1.2 kHz. A standard classification model
with the observed dataset can be designed using the Ma-
chine Learning Toolbox. Among the possible models, the
Tree Bag
8
reaches the highest accuracy of 0.93, using a
five-fold cross-validation algorithm on the given dataset.
9
Figure 5 shows the variance in the selected four vari-
ables. Each class is marked with a different colour, and
the overlapping and typical numeric range of the classes
can be observed. From the plot, it can be seen that the
RMS and band power allow us to clearly distinguish the
free-end and close-fit and these two classes are most dif-
ferent from each other. The cavern, on the other hand,
overlaps with both the close-fit and free-end classes. To
decide which model was the most optimal, a hyper-pa-
rameter tunning approach was utilized showing a com- parison between the first four most precise models. The
training was done using k-fold cross-validation with 5
folds. In this procedure, the whole dataset is divided into
five groups, where each time a different portion of the
dataset is the validation group and the rest is the training
group. The resulting accuracy is an average of these five
training sessions. This procedure ensures the robustness
R. DVOØÁK et al.: NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A MACHINE LEARNING APPROACH
564 Materiali in tehnologije / Materials and technology 58 (2024) 5, 561–565
Figure 4: Example of representative signals for each class, based on
lower Pearson residuals observed in all acquired signals: measured
signals and their time-frequency domain illustrated by fast continuous
wavelet transformation
Figure 5: Correlation plot of selected representative variables: band
power, root mean square, signal-to-noise ratio and dominant frequency
Table 1: Example of different classifiers and their accuracy used for
the dataset
Classifier Accuracy
Ensemble of bagged trees 93 %
Cubic support vector machine 88 %
Fine K-NN 87 %
Coarse Gaussian SVM 69 %
Figure 6: Confusion matrix of the designed model: bagged tree, vali-
dation accuracy of 0.93

of the training procedure, which could be biased if the
distribution of some of the classes is uneven.
The resulting model can be visualized using the con-
fusion matrix from Figure 6, which shows true-false pre-
dictions and prediction accuracy for each selected class.
The most successful classification can be observed in the
close-fit class, where only 5.1 % was misclassified. This
indicates that the designed model can be used to localize
healthy parts of the CIPP in the tested polygon. The low-
est accuracy can be seen in the cavern class, where
13.1 % of this class was misclassified as a close-fit or
free-end.
5 CONCLUSIONS
The paper presents a methodology for designing a
machine-learning algorithm for determining the state of
the inner wall of the cured-in-place pipe used for
trenchless retrofitting of sewage pipes. The methodology
uses data acquisition using the impact-echo NDT method
and a non-contact microphone, creating a training dataset
of signals and their features. The paper shows that train-
ing a machine learning classification model is possible
and can achieve high validation accuracy. However, con-
struction details and other defects, apart from the set
classes including close-fit, cavern and free-end, can in-
fluence classification accuracy. Possible accuracy and re-
liability of such a classification model can be achieved
by testing various types of pipes and their details regard-
ing the geometry of pipe, using the composition of the
CIPP and other conditions such as acoustic noise inside
the sewage system. The proposed method shows that the
heuristic models used for impact-echo classification can
be replaced by a machine-learning classification model,
which can increase classification accuracy and provide
testing procedures requiring less human experience as
they can be automated.
Acknowledgment
The paper was prepared under the internal grant
FAST-S-23-8275 of the Faculty of Civil Engineering,
Brno University of Technology. We are also thankful to
Wombat s.r.o. for granting us access to the testing poly-
gon and overseeing the technical details.
6 REFFERENCES
1
E. Kuliczkowska, A. Zwierzchowska, A qualitative analysis of early
defects present in PVC-U sewers but not observed in rigid pipes,
Tunnelling and Underground Space Technology, 56 (2016), 202–210
2
J. Hodul, et al., Effect of Chemical Aggressive Media on the Flexural
Properties of Cured-In-Place Pipes Supported by Microstructure Ob-
servation and Acoustic Emission, Materials, 13 (2020)14
3
N. J. Carino, The Impact-Echo Method: An Overview, Structures
(2001), 1–18
4
M. KAWO, Wombat s.r.o. [Online] 2022. https://www.wom-
bat.cz/metoda-kawo-bezvykopova-inverzni-technologie-pro-sanaci-k
analizacnich-sberacu/
5
L. P. A. Arts, E. L. van den Broek, The fast continuous wavelet trans-
formation (fCWT) for real-time, high-quality, noise-resistant
time–frequency analysis, Nature Computational Science, 2 (2022)1,
47–58
6
W. Xiong, L. Wu, F. Alleva, J. Droppo, X. Huang, A. Stolcke, The
Microsoft 2017 Conversational Speech Recognition System, 2018
IEEE International Conference on Acoustics, Speech and Signal Pro-
cessing (ICASSP) [online], IEEE, 2018, 5934–5938 [cit.
2024-07-25], ISBN 978-1-5386-4658-8, doi:10.1109/ICASSP.2018.
8461870
7
F. Lacy, A. Ruiz-Reyes, A. Brescia, Machinefor low signal-to-noise
ratio detection, Pattern Recognition Letters [online], 179 (2024),
115–122, [cit. 2024-07-25], ISSN 01678655, doi:10.1016/j.patrec.
2024.02.008
8
L. Breiman, Random forests, Machine Learning, 45 (2001) 1, 5–32
9
P. Refaeilzadeh, L. Tang, H. Liu, Cross-Validation, Encyclopedia of
Database Systems, Springer, Boston 2009, 532–538
R. DVOØÁK et al.: NON-DESTRUCTIVE TESTING OF CIPP DEFECTS USING A MACHINE LEARNING APPROACH
Materiali in tehnologije / Materials and technology 58 (2024) 5, 561–565 565