Composite Structures 295 (2022) 115786

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Machine learning based thermal imaging damage detection in glass-epoxy

composite materials
Ali Sarhadi a , Rodrigo Q. Albuquerque b,c ,∗, Martin Demleitner b , Holger Ruckdäschel b,c , Martin
A. Eder a ,∗∗
a
Technical University of Denmark, Frederiksborgvej 399, Roskilde 4000, Denmark
b University of Bayreuth, Universitätsstrasse 30, Bayreuth 95447, Germany
c Neue Materialien Bayreuth GmbH, Gottlieb-Keim-Straße 60, Bayreuth 95448, Germany

ARTICLE INFO ABSTRACT

Keywords: Machine learning (ML) based fatigue damage detection from thermal imaging in glass-epoxy composites is
Machine learning an important component of remote structural health monitoring used for safety assessment and optimization
Thermal imaging of composite structures and components. However, accurate characterization of fatigue damage hotspots in
Damage detection
terms of size, shape, location, hysteretic heat, and local temperature deep inside the material using surface
3D thermal analysis
thermal images remains a challenge to date. This work aims at evaluating the theoretical accuracy level of
Composite material
hotspot characterization by training a ML model with artificially generated thermal images from 3D finite
element models with increasing complexity. Modelling the fatigue damage as an intrinsic heat source allowed
to significantly reduce the influence of thermal image noise and other uncertainties related to heat transfer. It
is shown that ML can indeed accurately recover the heat influx, depth, and geometry of the heat source from
the original thermal images of the composite materials with prediction accuracies in the range 85%–99%. The
effect of training set size and image resolution on the prediction error is also presented. The findings reported
in this work contribute to the advancement of accurate and efficient remote fatigue damage detection methods
for fibre composite materials.

1. Introduction detecting fatigue damage, where a distinct temperature signature, com-

Fibre-reinforced composites are increasingly used for load-bearing
structures in various industries, from aerospace and automotive to
rotor blades for wind turbines or sports applications, due to their high
specific strength and stiffness. However, defects arising during man-
ufacture or during service life can significantly impair their excellent
properties [1,2]. The anisotropic structure leads to complex damage
mechanisms such as delamination, cracks or fibre breakage, which
makes the detection and evaluation of defects in terms of location
and size a particular challenge, especially in the case of defects that
lie far below the surface and make visual inspection impossible [3,4].
Here, non-destructive testing (NDT) techniques are essential to enable
reliable operation and extend service life. Currently, methods such as
X-ray tomography [5], ultrasound [6] or eddy current [7] are used.
Infrared thermography is particularly well suited because it offers
coupling agent-free measurement and fast scan rates that allow scan-
ning of large components, which is not possible with the previously
mentioned methods [8–10]. In addition, thermal imaging is ideal for
monly referred to as a ‘‘hotspot’’, is known to form due to frictional
processes [11–13]. However, the current challenge in using thermal
sensors to detect damage in composites is the ability to accurately as-
sess the severity, size, location through thickness, and hysteretic energy
dissipation (growth stage) of a damage area deep within a normally
complex and intricate composite structure. As a result, thermal images
of the surface rarely give a direct picture of the damage zone embedded
deep within the material below the surface. This means that in the case
of thick laminates and/or laminates with complex stacking sequences,
sandwich panels, bonded joints, and other complex local geometric
features, the thermal signature produced by the damaged region may
be highly distorted, rendering direct damage assessment from surface
thermal images a nontrivial task.
Damage assessment using thermal images requires to solve an in-
verse thermal conduction problem (IHCP), which is formulated by a
higher order partial differential equation (PDE) [14–16]. Here, the
intensity and shape of the defect are related to the parameters in

∗ Corresponding author at: University of Bayreuth, Universitätsstrasse 30, Bayreuth 95447, Germany.
∗∗ Corresponding author at: Technical University of Denmark, Frederiksborgvej 399, Roskilde 4000, Denmark.
E-mail addresses: rodrigo.q.albuquerque@uni-bayreuth.de (R.Q. Albuquerque), maed@dtu.dk (M.A. Eder).

https://doi.org/10.1016/j.compstruct.2022.115786
Received 22 February 2022; Received in revised form 28 April 2022; Accepted 13 May 2022
Available online 21 May 2022
0263-8223/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Sarhadi et al.
Nomenclature
𝜌 Material density (kg∕m3 )
𝑇 Temperature (K)
𝑇0 Ambient reference temperature (K)
𝑄̇ Total heat source influx (W)
𝑞𝑉̇ Heat source influx per unit area (W∕m2 )
𝑞conv Convective heat surface flux per unit area
(W∕m2 )
𝑎∕𝑏 Semi-major/semi-minor axis of the elliptic
disc source (m)
𝑅 Disc perimeter radius (m)
ℎ𝑡𝑜𝑝 Film coefficient top surface (W/(m2 K))
ℎ𝑏𝑜𝑡 Film coefficient bottom surface (W/(m2 K))
𝑡 Time or thickness (s or m)
𝑑 Distance of heat source from bottom sur-
face (m)
𝑥, 𝑦, 𝑧 Cartesian coordinates (m)
𝑘𝑡𝑡 Coefficient of thermal conductivity in 𝑡-
direction (𝑡 = 𝑥, 𝑦 or 𝑧) (W/(m K))
𝑐𝑝 Specific heat capacity (J/(kg K))
𝑦 True target property (m or W)
𝑦̂ Predicted target property (m or W)
𝑦̄ Average target property (m or W)
𝑛 Number of samples (unitless)
the source term of the heat conduction PDE (c.f. Eq. (1)). A popular
and widely used approach to solving IHCPs is by utilizing the finite
element method (FEM). The latter is used to compute the sensitivities
of the parameter space which can then be used to iterate towards
an inverse solution through an optimization procedure via e.g. the
gradient descent method. Kien and Zhuang [17] proposed a deep La-
grange method to solve structural optimization problems. The method
is capable of finding the minimum of the objective function which
could be applied to inverse problems. However, such approaches can
lead to a bottleneck because IHCPs are usually ill-posed (the solution
space is non-convex, there are many local minima) and computationally
demanding, and therefore time-consuming to solve even with high
computational power. Nanthakumar et al. [18] suggested to utilize
several experimental setups and to use improved regularization to
overcome these challenges. In that work, an algorithm was developed
to detect inclusions in a piezoelectric material by solving an inverse
problem through the use of the extended finite element method. Other
approaches rely on semi-analytical temperature field fitting optimiza-
tion such as the Trefftz method [19]. This method however, is available
only for a few specific cases which narrows the versatility.
Recent advances in data science and machine learning (ML) have
opened a new way to efficiently and accurately handle such inverse
problems and reduce computation time through more economical ap-
proximations [20]. ML has already been used for photoacoustic to-
mography [21], geophysics [22], and heat conduction problems [23].
Here, elements of physics-based modelling complement ML models,
improving understanding of the underlying physical processes and in-
creasing prediction accuracy. Anitescu et al. [24] proposed an artificial
neural network method capable of solving inverse boundary value
problems. They adopted an adaptive collocation method that reduced
the computational effort.
Fang et al. [25] have reported the use of a relatively complex deep
learning model to detect the depth of artificially created defects in
carbon fibre reinforced polymers (CFRP) with a prediction accuracy
of about 95%. Saeed et al. [26] used a combination of convolutional
networks and deep feed-forward neural networks to detect and predict
2
Composite Structures 295 (2022) 115786
the depth of defects in CFRP in real time, achieving an accuracy of
88%. The present work shows that it is possible to use a much simpler
regression model trained on artificially generated IR images based
on finite elements to predict not only the depth of defects of epoxy
thermoset systems, but also other three geometric parameters of the
defects to achieve a prediction accuracy sometimes higher than 99%
(vide infra).
The novelty in this work is the training of the ML model with arti-
ficially generated thermal images obtained from 3D numerical thermal
analysis to perform the simultaneous prediction of four geometrical
parameters of defects in epoxy thermoset glass fiber reinforced (GFRP)
systems. This approach offers the advantage of well-controlled hotspot
conditions, where each thermal image is precisely associated with a
priori defined heat transfer conditions, size, shape and position of the
heat source. As a result, thermal image noise and other uncertainties
are significantly reduced, allowing a largely unbiased assessment of the
level of accuracy that can theoretically be achieved to determine key
hotspot parameters via ML. This work contributes to the advancement
of thermal imaging-based structural health monitoring techniques that
can accurately and efficiently obtain information about the state of
fatigue damage in thick fibre/epoxy composites. Moreover, this work
opens up new opportunities for accelerated fatigue testing methods on
coupon and sub-component scale by ML aided adaptive load frequency
control.
2. Methodology
2.1. Thermal analysis
Thermal analysis is conducted in order to phenomenologically
model the heat transfer involved in fatigue damage accumulation
processes with the purpose to create artificial thermal surface images
under controlled ideal conditions. The heat transfer models presented
in this work are described by the well-known general 3D heat equation
for anisotropic conductivity and internal heat source in its strong form
as follows
( )
𝜕𝑇 1 𝜕2 𝑇 𝜕2 𝑇 𝜕2 𝑇 𝑄̇
− 𝑘𝑥𝑥 + 𝑘𝑦𝑦 + 𝑘𝑧𝑧 = (1)
𝜕𝑡 𝑐𝑝 𝜌 𝜕𝑥2 𝜕𝑦2 𝜕𝑧2 𝑐𝑝 𝜌
The right hand side source term in Eq. (1) models the intrinsic heat
generated by a local defect e.g. a delamination or a crack located inside
a material due to predominantly frictional effects under cyclic fatigue
loading conditions. Such a concentrated heat source causing a distinct
temperature gradient that typically flares up in thermal images is com-
monly referred to as hotspot. The enthalpy increase of such a hotspot
is closely associated with hysteretic energy dissipation which acts like
a heat pump. It was shown elsewhere [11] that typically the hysteretic
heat input in composite materials under high cycle fatigue conditions
is low to moderate and constant, which eventually leads to thermal
equilibrium. Therefore, with good approximation the problem can be
modelled in the steady state approximation in which case the transient
part represented by the left hand side, temporal partial term in Eq. (1)
can be neglected, i.e., 𝜕𝑇
𝜕𝑡
→ 0. The size of the hotspot source term was
chosen small compared to the geometrical dimensions of the model.
The hotspot temperatures in glass/epoxy composite materials typically
range between 20 °C to 70 °C. In-plane hot spots parallel to the surface
were assumed since defects in composites usually occur between plies
in the form of debonding, delamination and microcracking.
Since radiative heat transfer is proportional to the fourth power of
the temperature, the effect of radiation can be considered insignificant
at these low temperatures. Instead, convection is the governing heat
transfer mechanism in the present problem which prevails on both
external surfaces that are in contact with the ambient air. The convec-
tive heat transfer boundary conditions assigned to both surfaces can be
( )
written in the form 𝑞conv = ℎ 𝑇 − 𝑇0 . The in-plane dimensions of the
thin walled problem are chosen ten times the thickness dimension of
A. Sarhadi et al.
Fig. 1. (a) Circular disc domain with 𝑅 ≫ 𝑡 with global coordinate system and location of point source on the 𝑧-axis and (b) axisymmetric representation of hatched rectangle
with thickness 𝑡, radius 𝑅 and heat transfer boundary conditions (Note: figure not to scale).
Table 1
Thermal modelling parameters common to all models presented in Section 2.
𝑅 𝑡 𝜌 𝑐𝑝 ℎ𝑡𝑜𝑝 ℎ𝑏𝑜𝑡
m m kg∕m3 J/(kg K) W/(m2 K) W/(m2 K)
0.5 0.05 1992.5 940.5 10.0 5.0
Table 2
Coefficients of thermal conductivity with values resembling a generic glass/epoxy
composite material in the global coordinate directions; in the orthotropic case the fibre
direction is assumed to be aligned with the global x-direction whereas the thermal
conductivity in the transverse and through-thickness direction was set equal.
Case 𝑘𝑥𝑥 𝑘𝑦𝑦
– W/(m K) W/(m K)
Isotropic 0.695 0.695
Orthotropic 0.805 0.374
the model. Therefore, it can be assumed with reasonable approximation
that conductive heat transfer across the model boundary remote from
the hotspot is small and has a negligible influence on the hotspot tem-
perature profile. Therefore, zero flux heat transfer boundary conditions
were assigned to the through thickness perimeter surface.
Table 1 lists the geometrical properties of the circular disc mod-
els as well as the thermal properties and the ambient temperature
consistently adopted throughout all cases.
2.1.1. Thermal model with point source
Fig. 1 depicts the geometry of the thermal model with the point heat
source and the heat transfer boundary conditions.
This geometry was discretized (in its weak form of Eq. (1)) within
the commercially available finite element software package Comsol [27]
with around 1.9 × 105 ten-noded quadratic tetrahedral elements. Zero
flux conditions were assigned to the nodes pertaining to the through-
thickness perimeter of the circumferential boundary. Convective heat
flux boundary conditions were assigned to either surface with different
film coefficients (see Table 1). Two different cases were analysed
namely an isotropic case and an orthotropic case whose coefficients of
thermal conductivity are provided in Table 2.
The concentrated heat generation source magnitude was randomly
chosen from the range 2 < 𝑄̇ < 4 W and was assigned to a single node
located on the z-axis at a randomly chosen distance within the range
0.005 < 𝑑 < 0.045 m (see red dot in Fig. 1). A generalized minimum
residual method (GMRES) was used to solve the linear problem. The
Comsol solver was run through a Matlab [28] code which was used to
change the parameters 𝑄̇ and 𝑑. The two parameters were randomly
chosen between their limits in such a way that unique parameter pairs
were created in each run. An artificial greyscale thermal image of the
top-surface was created with a view vector direction along the negative
z-axis (see camera symbol in Fig. 1(a)). A fixed temperature contour
range of 20 ≤ 𝑇𝑡𝑜𝑝 ≤ 70 °C was used for the thermal contour plots in all
the runs.
Composite Structures 295 (2022) 115786
2.1.2. Thermal model with elliptical heat source
Fig. 2 depicts the geometry of the thermal model with the elliptical
𝑇0 disc heat source of zero thickness and the heat transfer boundary
K conditions.
293.15 The numerical analysis procedure and discretization was the same
as previously described in Section 2.1.1. Orthotropic coefficients of
thermal conductivity were adopted as provided in Table 2.
A total heat generation magnitude was randomly chosen from the
range 2 < 𝑄̇ < 4 W and was uniformly distributed over the area of
the elliptic disc centred on the 𝑧-axis at a randomly chosen distance
within the range 0.005 < 𝑑 < 0.045 m (see red hatched ellipse and
𝑘𝑧𝑧
red line in Fig. 2). Both, the semi-major and semi-minor parameters
W/(m K) were randomly chosen from a range 3.0 × 10−2 < 𝑎 < 9.0 × 10−2 m
0.695 and 1.5 × 10−2 < 𝑎 < 4.5 × 10−2 m. The Comsol solver was run through
0.374 a Matlab code which was used to change the parameters 𝑄, ̇ 𝑑, 𝑎 and
𝑏. The four parameters were randomly chosen between their limits in
such a way that unique parameter quartets were created in each run.
The generation of the artificial greyscale thermal images as well as the
temperature scale bar are analogous to Section 2.1.1.
2.2. Machine learning modelling
All models were created using Python libraries (e.g. Numpy,
Sklearn) and Jupyter notebook. All codes were written in Python 3.
2.2.1. Data treatment
The greyscale thermal images were imported, cropped, resized and
converted into 2D arrays (116 × 115 pixels) with pixel intensity as
elements. Light-saturated images with pixel intensities equal to 255
were removed from the dataset. The 2D arrays were flattened to
become 1D arrays with dimension 13 340. Only every fifth element of
each 1D array was used to build the final data set, that contained 2668
features. Each feature was scaled to have zero mean and unit variance,
where scaling was first applied to the training set (80% of the data)
and then used to re-scale the features of the test set.
2.2.2. Model
The Kernel Ridge Regression (KRR [29]) model was trained to
predict the target properties 𝑄, ̇ 𝑑 and, when applicable, 𝑎 and 𝑏,
using the pixel intensities of thermal images as features. Although
image processing to predict target properties can be usually done using
convolutional neural networks, we have chosen KRR because it is
much simpler and faster to implement and to have its hyperparameters
optimized, while being considerably precise, as will be shown in the
results section. KRR combines the kernel trick [30] with a 𝐿2 regu-
larization term, which helps avoiding overfitting. This method has a
closed form that allows for efficient training and prediction for middle-
sized datasets. In particular, the use of the radial basis function kernel
means that the original features are expanded into an infinite number
of terms, which are not possible anymore to be individually evaluated,
but whose related inner product needed for performing predictions can
3
A. Sarhadi et al.
Fig. 2. (a) Circular disc domain with 𝑅 ≫ 𝑡 with global coordinate system and location of zero thickness elliptical disc source centred on the 𝑧-axis with semi-major axis 𝑎 and
semi-minor axis 𝑏 and (b) axisymmetric representation of hatched rectangle with thickness 𝑡, radius 𝑅 and heat transfer boundary conditions (Note: figure not to scale).
Fig. 3. Steady state top surface temperature distribution for (a) point source with
isotropic thermal conductivity with 𝑄̇ = 3 W and 𝑑 = 35 mm, (b) point source with
orthotropic thermal conductivity with 𝑄̇ = 3 W and 𝑑 = 35 mm, (c) elliptical source
with orthotropic thermal conductivity with 𝑄̇ = 3.5 W, 𝑑 = 45 mm, 𝑎 = 60 mm, 𝑏 = 30 mm
and the corresponding radial temperature profiles of both surfaces for (d) point source
— isotropic, (e) point source — orthotropic and (f) elliptical source — orthotropic. The
defects are highlighted in red.
be precisely calculated via the kernel trick. Other linear and non-linear
ML models (Gradient Boosting Regression, Support Vector Machine and
Lasso) were also tested but their performances were slightly worse than
KRR (see the Supporting Information, SI).
4
Composite Structures 295 (2022) 115786
2.2.3. Point source case
The thermal images generated using point defects were
pre-processed as already described. The hyperparameters 𝑘𝑒𝑟𝑛𝑒𝑙, 𝑎𝑙𝑝ℎ𝑎
and 𝑔𝑎𝑚𝑚𝑎 of the KRR method were optimized using a random grid
search method and an initial batch of 350 images corresponding to the
training set. During the hyperparameter optimization, the performance
of the trained model was evaluated using the leave-one-out cross-
validation (CV) technique with all images of the training set, as well
as the coefficient of determination (𝑅2 ) and the mean absolute error
(MAE), which are given by
1 ∑|
𝑛
MAE = 𝑦 − 𝑦̂𝑖 || (2)
𝑛 𝑖=1 | 𝑖
∑𝑛
(𝑦𝑖 − 𝑦̂𝑖 )2
̂ = 1 − ∑𝑖=1
𝑅2 (𝑦, 𝑦) 𝑛 (3)
̄2
𝑖=1 (𝑦𝑖 − 𝑦)
The mean absolute percentual error (MAPE) was also used to quan-
tify the accuracy of the models and is given by
100% ∑ || 𝑦𝑖 − 𝑦̂𝑖 ||
𝑛
MAPE = (4)
𝑛 𝑖=1 || 𝑦𝑖 ||
The leave-one-out CV technique starts with training the model with
all samples except only one and performing one single prediction with
that one-sample test set initially left out. This operation is performed
𝑛 times using every time a different sample of the dataset as the one-
sample test set. After this CV procedure, 𝑛 different models will have
been trained and evaluated using the same performance indicator (here,
MAE and 𝑅2 ). In order to evaluate the model with samples that were
never seen during the hyperparameter optimization, the trained model
was used to predict the target properties of the test set (88 samples).
The model performance was evaluated using the same CV technique
and the MAE and 𝑅2 parameters. The model was also evaluated by the
MAPE error.
2.2.4. Elliptic disc source case
For the elliptic source case, 925 thermal images were used in the
dataset, where 740 images were randomly selected as the training set
and used to train the ML model already with optimized parameters to
̇ 𝑑, 𝑎, and 𝑏 of the test set. The
predict the geometrical parameters 𝑄,
model was evaluated using MAE, 𝑅2 and MAPE.
2.2.5. Influence of the training set and image resolution
The influence of the size of the training set, as well as of the
image resolution on the model’s accuracy were also investigated to
provide more insight on the ML model. For the influence of the size
A. Sarhadi et al.
Fig. 4. Comparison between the predicted and true parameters for the test set using the trained ML model and point source case. The accuracies were 99 and 96% for 𝑄̇ and 𝑑,
respectively.
of the training set, ML models were built using a variable number
of training samples (𝑁) from the training set while predictions were
always performed on the same test set. The MAE and 𝑅2 parameters
were monitored for different 𝑁 values, both for the point source and
elliptic source cases. The influence of the image resolution on the
parameters MAE and 𝑅2 was also investigated for both source cases
by monitoring these parameters for different image resolutions, which
were in the range 20 × 20–165 × 165 pixels using the full training set
to train the model and the test set to perform predictions.
3. Results and discussion
3.1. Artificial thermal images for point and elliptical disc source
Fig. 3 shows representative examples of the three different cases
investigated. Fig. 3(a) shows the axisymmetric temperature distribution
due to a point source in the isotropic case. Conversely, Fig. 3(b) shows
that the surface temperature due to a point source in an orthotropic
material appears elongated due to the higher thermal conductivity in
𝑥-direction. The elongated temperature distribution shown in Fig. 3(c)
is a combination of both, the elliptic shape of the source and the
anisotropic heat conductivity. The pronounced temperature peak seen
in Fig. 3(d,e) is the result of a singular heat generation intensity
at the point source. On the other hand, the elliptical heat source is
not singular which results in a lower heat generation intensity and
a more bell-shaped temperature profile. Comparison of Fig. 3(b) and
(c) demonstrates the difficulty to predict the defect geometry since
both the anisotropic heat conductivity and the heat source shape itself
contribute to the elongated surface temperature distribution observed
in both images.
3.2. Predictions for the point source case
The trained kernel ridge model had the hyperparameters optimized
using the training set, evaluated via the leave-one-out cross validation
(see SI for more details), from where the final hyperparameters were
obtained: 𝑘𝑒𝑟𝑛𝑒𝑙 = ‘‘rbf’’, 𝑔𝑎𝑚𝑚𝑎 = 0.0041, 𝑎𝑙𝑝ℎ𝑎 = 0.00009. The model
trained with the whole training set was used to predict 𝑄̇ and 𝑑 for
the thermal images of the test set. For this simple point source case,
the agreement between the predicted and true target properties was
excellent (see Fig. 4).
The MAPE error of both predictions were (in %): 0.65 (𝑄) ̇ and
3.59 (𝑑). The corresponding model’s accuracy is larger than 99% for
the prediction of 𝑄̇ and still very high (96.4%) for the prediction of
𝑑. In addition, the predicted and true target properties were highly
correlated, as concluded from 𝑅2 values very close to 1.
5
Composite Structures 295 (2022) 115786
3.3. Predictions for the elliptical disc source case
The inclusion of two additional parameters as target properties,
namely 𝑎 and 𝑏, slightly decreased the predictive accuracy of the ML
model for 𝑄̇ and 𝑑 (see the SI). The initially chosen ranges used in
the Comsol/FEM simulations for 𝑎 and 𝑏 were relatively small (𝑎, 𝑏:
5 × 10−3 m to 1.5 × 10−2 m), which resulted in a very noisy prediction
of 𝑎 and 𝑏, as well as relatively small accuracies: 75 and 82% for 𝑎
and 𝑏, respectively. It seems that the larger value of 𝑘𝑥𝑥 compared to
𝑘𝑦𝑦 worsens the prediction accuracy of 𝑎 compared to 𝑏. The larger 𝑘𝑥𝑥
leads to a smaller variation between neighbouring pixels, decreasing
the sensitivity of the ML model towards the prediction of 𝑎. Moreover,
smaller ranges of 𝑎, 𝑏 worsened MAPE and 𝑅2 as compared with larger
ranges.
In order to investigate the improvement of the predictions, a new
set of wider and individually different ranges was then chosen (𝑎:
3 × 10−2 m to 9.0 × 10−2 m and 𝑏: 1.5 × 10−2 m to 4.5 × 10−2 m) to generate
the thermal images. The increase of the size of the ellipse improved the
accuracy of the ML predictions considerably in the case of 𝑎 and 𝑏, as
can be seen in Fig. 5.
The prediction accuracy was in the range 85%–98%, and the lowest
accuracy found, which corresponded to the depth parameter 𝑑 (accu-
racy = 85%) was comparable to the best accuracy reported in another
related prediction reported in the literature [26]. The use of wider
ranges for 𝑎 and 𝑏 to generate the thermal images seems to be the best
compromise for achieving good accuracy for the prediction of all four
parameters. However, this approach penalized the excellent accuracy
found for 𝑄̇ (99%) and mainly for 𝑑 (94%) that were obtained using
narrower ranges. For applications where the depth and the temperature
of the defect are the most important properties to be predicted, it is
recommended to train a ML model using thermal images generated
adopting narrower ranges for 𝑎 and 𝑏.
3.4. Influence of the training set and image resolution
The influence of the number of training samples on the accuracy of
the predictions is shown in Fig. 6. Note that the predicted parameters
on the y-axes of Fig. 6 are normalized for comparison reasons (their
absolute values are provided in the SI).
For the point source case (Fig. 6(a,b), solid lines) the model im-
proves considerably up to about 100 training samples and basically
stabilizes the performance from about 250 training samples. Adding
more samples to the training set from this point improves the model
more slowly. Both the MAE and 𝑅2 trends could be nicely reproduced
by a fitting of the type 𝑦 = 𝑐1 𝑥𝑐2 +𝑐3 (dashed lines in Fig. 6(a,b)), where
𝑐1 , 𝑐2 and 𝑐3 are adjustable parameters. For the elliptic source case
(Fig. 6(c,d)), MAE and 𝑅2 vary very little when the size of the training
set is close to 740 training samples, and the largest improvement occurs
up to 100 training samples. Fig. 6 reveals that the chosen sizes of
A. Sarhadi et al. Composite Structures 295 (2022) 115786

Fig. 5. Comparison between the predicted and true parameters for the test set using the trained ML model and elliptic source case. The accuracy values (in %) for these predictions
̇ 𝑑, 𝑎, and 𝑏, respectively.
were 98, 85, 94, and 90 for 𝑄,

Fig. 6. Influence of the number of training samples (𝑁) on the models’ performance for the point source (subplots a and b) and elliptic disc source (subplots c and d) cases in
an orthotropic material. MAE and 𝑅2 were normalized for comparison reasons and curve fittings were added for the point source case (dashed lines).

the training sets used in the predictions shown in Figs. 4 and 5 were
appropriate.
The influence of the image resolution on the model’s performance
is shown in Fig. 7 for the point source (subplots a and b) and elliptic
source (subplots c and d) cases. The resolution shown in the x-axes is
related to the width and height of the image, which were the same.
The trends are much noisier than those found in Fig. 6, especially for
the elliptic source case. MAE and 𝑅2 already stabilize from resolutions
of about 50 × 50 pixels in the point source case (Fig. 7(a,b)), but the
same parameters stabilize with considerably higher resolutions of about

6
A. Sarhadi et al.
Fig. 7. Influence of image resolution (in pixels) for the point source (subplots a and b) and elliptic disc source (subplots c and d) cases in an orthotropic material. MAE and 𝑅2
were normalized for comparison reasons.
130 × 130 pixels in the elliptic source case (Fig. 7(c,d)), although from
about 50 × 50 pixels a very large improvement of the model’s accuracy
is also observed for the latter case.
The trends shown in Fig. 7 were noisier at the low-resolution
regions. This is due to random mismatches between the pixels extracted
to train the ML model (i.e., every fifth pixel, see Section 2.2.1) and
pixels carrying some information other than the zero intensity corre-
sponding to the black background. One fifth of all extracted pixels
of images exhibiting close but different resolutions (e.g. 30 × 30
pixels versus 32 × 32 pixels) might contain a very different number
of non-black pixels that indeed contribute to the improvement of the
corresponding ML models.
4. General remarks
The ML model proved itself quite robust, where as little as 100
thermal images suffice for training to get a good performance. This
suggests that building a ML model based on experimentally measured
IR images of defects in epoxy composites is feasible, especially if one
uses data augmentation also including artificially generated images.
However, this ML model also has some limitations. For instance,
no defect detection can be done to identify its exact location in the
thermal image, and only one defect is expected to be present in the
same image. The geometry of real defects of composite materials can
be very different from a simple ellipse. Therefore, the current ML model
could be further extended to address these problems, which is the
subject of a future investigation.
5. Conclusion
In this work, it was shown that ML can be used to predict four key
physical and geometric parameters of a fatigue damage hotspot from
surface thermal images, namely heat source intensity, depth, and size.
7
Composite Structures 295 (2022) 115786
The proposed ML method is able to overcome the inherent com-
plexity of IHCPs resulting from the anisotropic heat transfer behaviour,
asymmetric thermal boundary conditions, and non-uniformly shaped
hotspots located deep inside a glass/epoxy composite structure. Pre-
diction is possible with temperature information from only one side
of the damaged plate. It should be noted, however, that the method
presented was applied specifically to high cycle fatigue conditions
where quasi-stationary conditions prevail. Such heat transfer conditions
are less complex than in thermographic damage detection, depend-
ing significantly on time evolution. While the transient response in
thermographic damage detection requires higher order neural network
models, the steady-state response of a fatigue damage hotspot allows
the use of a simpler ML model, which provides higher efficiency in such
applications.
In all cases studied, the predictive performance of the ML algorithm
was found to follow a power function of the size of the training set. The
accuracy improved significantly up to about 100 training images, after
which the improvement increased more slowly and asymptotically. This
comparatively small training set proves the efficiency of the method.
Doubling the number of independent parameters from two to four
significantly expands the solution space, which affects the prediction
accuracy of the elliptical disc source model compared to the point
source. The prediction accuracy improved as the size of the elliptical
heat source dimensions increased. For a small mean relative heat source
size of 𝑎𝑚𝑒𝑎𝑛 ∕𝑅 = 2%, the prediction was poor, while a larger relative
mean heat source size of 𝑎𝑚𝑒𝑎𝑛 ∕𝑅 = 9% gave a much more accurate
prediction. This means that a smaller size of the heat source results in
lower variation of the surface pixel temperature values in the images
and accordingly makes it more challenging for the ML algorithm to
predict the parameter. By the same argument, the lower prediction
accuracy of parameter 𝑏 compared to parameter 𝑎 for the larger set
of ranges is due to the smaller range of the former, even though 𝑘𝑥𝑥 is
more than twice as large as 𝑘𝑦𝑦 .
Higher image resolutions resulted in better performance of the ML
model and the observed trend was more choppy in low resolution
A. Sarhadi et al.
regions, which is due to larger variations in the number of non-black
pixels effectively used to train the model.
Future work will address arbitrarily shaped heat sources present at
different depths of the composite material by utilizing convolutional
neural networks.
CRediT authorship contribution statement
Ali Sarhadi: Writing – original draft, Preparation of the dataset,
Comsol simulations. Rodrigo Q. Albuquerque: ML modelling, Pa-
per writing and correction, Discussion. Martin Demleitner: Paper
correction, Discussion. Holger Ruckdäschel: Paper correction, Discus-
sion. Martin A. Eder: FEM simulations, Paper writing and correction,
Discussion.
Declaration of competing interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Acknowledgements
DTU Wind Energy is grateful for the financial support of this publi-
cation by VILLUM FONDEN with grant number 36050. Furthermore,
DTU Wind Energy acknowledges the financial support through the
EUDP project ReliaBlade with grant number 64018-0068.
Appendix A. Supplementary data
Supplementary material related to this article can be found on-
line at https://doi.org/10.1016/j.compstruct.2022.115786. More de-
tails about model performances and general trends are available as PDF.
References
[1] Persson Erik, Eriksson Ingvar, Zackrisson Leif. Effects of hole machining defects
on strength and fatigue life of composite laminates. Composites A 1997. http:
//dx.doi.org/10.1016/S1359-835X(96)00106-6.
[2] Hörrmann Susanne, Adumitroaie Adi, Viechtbauer Christoph, Schagerl Martin.
The effect of fiber waviness on the fatigue life of CFRP materials. Int J Fatigue
2016. http://dx.doi.org/10.1016/j.ijfatigue.2016.04.029.
[3] Bard S, Demleitner M, Häublein M, Altstädt V. Fracture behaviour of prepreg
laminates studied by in-situ sem mechanical tests. In: Procedia Structural
Integrity, Vol. 13. 2018, http://dx.doi.org/10.1016/j.prostr.2018.12.299.
[4] Bakis Gökhan, Wendel Jan Felix, Zeiler Rico, Aksit Alper, Häublein Markus,
Demleitner Martin, Benra Jan, Forero Stefan, Schütz Walter, Altstädt Volker.
Mechanical properties of the carbon nanotube modified epoxy–carbon fiber
unidirectional prepreg laminates. Polymers 2021. http://dx.doi.org/10.3390/
polym13050770.
[5] Jolly Mr, Prabhakar A, Sturzu B, Hollstein K, Singh R, Thomas S, Foote P,
Shaw A. Review of non-destructive testing (NDT) techniques and their appli-
cability to thick walled composites. In: Procedia CIRP. 2015, http://dx.doi.org/
10.1016/j.procir.2015.07.043.
[6] Kempf M, Skrabala O, Altstädt V. Acoustic emission analysis for characterisation
of damage mechanisms in fibre reinforced thermosetting polyurethane and epoxy.
Composites B 2014. http://dx.doi.org/10.1016/j.compositesb.2013.08.080.
[7] Koyama Kiyoshi, Hoshikawa Hiroshi, Kojima Gouki. Eddy current nondestructive
testing for carbon fiber-reinforced composites. J. Pressure Vessel Technology,
Transactions of the ASME 2013. http://dx.doi.org/10.1115/1.4023253.
[8] Bang Hyun Tae, Park Solmoi, Jeon Haemin. Defect identification in composite
materials via thermography and deep learning techniques. Compos Struct 2020.
http://dx.doi.org/10.1016/j.compstruct.2020.112405.
Composite Structures 295 (2022) 115786
[9] Chulkov AO, Nesteruk DA, Vavilov VP, Shagdirov B, Omar M, Siddiqui AO,
Prasad YLVD. Automated procedure for detecting and characterizing defects in
gfrp composite by using thermal nondestructive testing. Infrared Phys Technol
2021. http://dx.doi.org/10.1016/j.infrared.2021.103675.
[10] Marani Roberto, Palumbo Davide, Galietti Umberto, D’Orazio Tiziana. Deep
learning for defect characterization in composite laminates inspected by
step-heating thermography. Opt Lasers Eng 2021. http://dx.doi.org/10.1016/j.
optlaseng.2021.106679.
[11] Eder Martin A, Sarhadi Ali, Chen Xiao. A novel and robust method to quantify
fatigue damage in fibre composite materials using thermal imaging analysis. Int
J Fatigue 2021;150:106326. http://dx.doi.org/10.1016/j.ijfatigue.2021.106326.
[12] Gornet Laurent, Wesphal Ophélie, Burtin Christian, Bailleul Jean Luc, Rozy-
cki Patrick, Stainier Laurent. Rapid determination of the high cycle fatigue
limit curve of carbon fiber epoxy matrix composite laminates by thermography
methodology: Tests and finite element simulations. In: Procedia Engineering.
2013, http://dx.doi.org/10.1016/j.proeng.2013.12.123.
[13] Peyrac Catherine, Jollivet Thomas, Leray Nolwenn, Lefebvre Fabien, West-
phal Ophélie, Gornet Laurent. Self-heating method for fatigue limit determination
on thermoplastic composites. In: Procedia Engineering. 2015, http://dx.doi.org/
10.1016/j.proeng.2015.12.639.
[14] Chen Aizhou, Li Yuan, Yan Bo. Computational inverse methods of heat source
in fatigue damage problems. In: AIP Conference Proceedings. 2018, http://dx.
doi.org/10.1063/1.5033668.
[15] Peng Tishun, Liu Yongming. 3D crack-like damage imaging using a novel inverse
heat conduction framework. Int J Heat Mass Transfer 2016. http://dx.doi.org/
10.1016/j.ijheatmasstransfer.2016.06.018.
[16] Muramatsu M, Nakasumi S, Harada Y, Suzuki T. Application of the inverse heat
conduction analysis to the evaluation of defects in carbonfiber-reinforced plastics.
Mech Compos Mater 2015. http://dx.doi.org/10.1007/s11029-015-9458-y.
[17] Kien DN, Zhuang X. A deep neural network-based algorithm for solving structural
optimization. J Zhejiang Univ Sci A 2021;22(8):609–20. http://dx.doi.org/10.
1631/jzus.A2000380.
[18] Nanthakumar SS, Lahmer T, Zhuang X, Zi G, Rabczuk T. Detection of
material interfaces using a regularized level set method in piezoelectric struc-
tures. Inverse Problems Sci Eng 2016;24(1):153–76. http://dx.doi.org/10.1080/
17415977.2015.1017485.
[19] Ciałkowski Michał J, Grysa Krzysztof. Trefftz method in solving the inverse
problems. J. Inverse Ill-Posed Prob 2010. http://dx.doi.org/10.1515/JIIP.2010.
027.
[20] Adler Jonas, Öktem Ozan. Solving ill-posed inverse problems using iterative deep
neural networks. 2017, http://dx.doi.org/10.1088/1361-6420/aa9581,
[21] Antholzer Stephan, Haltmeier Markus, Schwab Johannes. Deep learning for
photoacoustic tomography from sparse data. 2019, http://dx.doi.org/10.1080/
17415977.2018.1518444,
[22] Vamaraju Janaki, Sen Mrinal K. Unsupervised physics-based neural networks
for seismic migration. Interpretation 2019. http://dx.doi.org/10.1190/INT-2018-
0230.1.
[23] Deng S, Hwang Y. Applying neural networks to the solution of forward and
inverse heat conduction problems. Int J Heat Mass Transfer 2006. http://dx.doi.
org/10.1016/j.ijheatmasstransfer.2006.06.009.
[24] Anitescu Cosmin, Atroshchenko Elena, Alajlan Naif, Rabczuk Timon. Artificial
neural network methods for the solution of second order boundary value prob-
lems. Comput Mater Continua 2019;59(1):345–59. http://dx.doi.org/10.32604/
cmc.2019.06641, URL http://www.techscience.com/cmc/v59n1/27936.
[25] Fang Qiang, Maldague Xavier. A method of defect depth estimation for simulated
infrared thermography data with deep learning. Appl Sci 2020;10(19):6819.
http://dx.doi.org/10.3390/app10196819.
[26] Saeed Numan, King Nelson, Said Zafar, Omar Mohammed A. Automatic defects
detection in CFRP thermograms, using convolutional neural networks and
transfer learning. Infrared Phys Technol 2019;102:103048. http://dx.doi.org/
10.1016/j.infrared.2019.103048, URL https://www.sciencedirect.com/science/
article/pii/S1350449519303135.
[27] COMSOL Inc. Comsol. 2020, URL http://www.comsol.com/products/
multiphysics/.
[28] MATLAB 2021. Documentation, version 7.10.0 (R2020a). Natick, Massachusetts:
The MathWorks Inc.; 2021.
[29] Vovk V. Empirical Inference. Springer; 2013, p. 105–16. http://dx.doi.org/10.
1007/978-3-642-41136-6.
[30] Hofmann T, Schölkopf B, Smola AJ. Kernel methods in machine learning. Ann
Statist 2008;36(3):1171–220. http://dx.doi.org/10.1214/009053607000000677.