Professional Documents
Culture Documents
1 s2.0 S0263822322005591 Main
1 s2.0 S0263822322005591 Main
Composite Structures
journal homepage: www.elsevier.com/locate/compstruct
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.
∗ 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. Composite Structures 295 (2022) 115786
2
A. Sarhadi et al. Composite Structures 295 (2022) 115786
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).
3
A. Sarhadi et al. Composite Structures 295 (2022) 115786
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).
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 ||
be precisely calculated via the kernel trick. Other linear and non-linear 2.2.5. Influence of the training set and image resolution
ML models (Gradient Boosting Regression, Support Vector Machine and The influence of the size of the training set, as well as of the
Lasso) were also tested but their performances were slightly worse than image resolution on the model’s accuracy were also investigated to
KRR (see the Supporting Information, SI). provide more insight on the ML model. For the influence of the size
4
A. Sarhadi et al. Composite Structures 295 (2022) 115786
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 3.3. Predictions for the elliptical disc source case
of training samples (𝑁) from the training set while predictions were
always performed on the same test set. The MAE and 𝑅2 parameters The inclusion of two additional parameters as target properties,
were monitored for different 𝑁 values, both for the point source and namely 𝑎 and 𝑏, slightly decreased the predictive accuracy of the ML
elliptic source cases. The influence of the image resolution on the model for 𝑄̇ and 𝑑 (see the SI). The initially chosen ranges used in
parameters MAE and 𝑅2 was also investigated for both source cases the Comsol/FEM simulations for 𝑎 and 𝑏 were relatively small (𝑎, 𝑏:
by monitoring these parameters for different image resolutions, which 5 × 10−3 m to 1.5 × 10−2 m), which resulted in a very noisy prediction
were in the range 20 × 20–165 × 165 pixels using the full training set of 𝑎 and 𝑏, as well as relatively small accuracies: 75 and 82% for 𝑎
to train the model and the test set to perform predictions. 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
3. Results and discussion 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
3.1. Artificial thermal images for point and elliptical disc source ranges.
In order to investigate the improvement of the predictions, a new
Fig. 3 shows representative examples of the three different cases set of wider and individually different ranges was then chosen (𝑎:
investigated. Fig. 3(a) shows the axisymmetric temperature distribution 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
due to a point source in the isotropic case. Conversely, Fig. 3(b) shows 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
that the surface temperature due to a point source in an orthotropic
can be seen in Fig. 5.
material appears elongated due to the higher thermal conductivity in
The prediction accuracy was in the range 85%–98%, and the lowest
𝑥-direction. The elongated temperature distribution shown in Fig. 3(c)
accuracy found, which corresponded to the depth parameter 𝑑 (accu-
is a combination of both, the elliptic shape of the source and the
racy = 85%) was comparable to the best accuracy reported in another
anisotropic heat conductivity. The pronounced temperature peak seen
related prediction reported in the literature [26]. The use of wider
in Fig. 3(d,e) is the result of a singular heat generation intensity
ranges for 𝑎 and 𝑏 to generate the thermal images seems to be the best
at the point source. On the other hand, the elliptical heat source is
compromise for achieving good accuracy for the prediction of all four
not singular which results in a lower heat generation intensity and parameters. However, this approach penalized the excellent accuracy
a more bell-shaped temperature profile. Comparison of Fig. 3(b) and found for 𝑄̇ (99%) and mainly for 𝑑 (94%) that were obtained using
(c) demonstrates the difficulty to predict the defect geometry since narrower ranges. For applications where the depth and the temperature
both the anisotropic heat conductivity and the heat source shape itself of the defect are the most important properties to be predicted, it is
contribute to the elongated surface temperature distribution observed recommended to train a ML model using thermal images generated
in both images. adopting narrower ranges for 𝑎 and 𝑏.
3.2. Predictions for the point source case 3.4. Influence of the training set and image resolution
5
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 related to the width and height of the image, which were the same.
appropriate. The trends are much noisier than those found in Fig. 6, especially for
The influence of the image resolution on the model’s performance the elliptic source case. MAE and 𝑅2 already stabilize from resolutions
is shown in Fig. 7 for the point source (subplots a and b) and elliptic of about 50 × 50 pixels in the point source case (Fig. 7(a,b)), but the
source (subplots c and d) cases. The resolution shown in the x-axes is same parameters stabilize with considerably higher resolutions of about
6
A. Sarhadi et al. Composite Structures 295 (2022) 115786
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 The proposed ML method is able to overcome the inherent com-
about 50 × 50 pixels a very large improvement of the model’s accuracy plexity of IHCPs resulting from the anisotropic heat transfer behaviour,
is also observed for the latter case. asymmetric thermal boundary conditions, and non-uniformly shaped
The trends shown in Fig. 7 were noisier at the low-resolution hotspots located deep inside a glass/epoxy composite structure. Pre-
regions. This is due to random mismatches between the pixels extracted diction is possible with temperature information from only one side
to train the ML model (i.e., every fifth pixel, see Section 2.2.1) and of the damaged plate. It should be noted, however, that the method
pixels carrying some information other than the zero intensity corre- presented was applied specifically to high cycle fatigue conditions
sponding to the black background. One fifth of all extracted pixels where quasi-stationary conditions prevail. Such heat transfer conditions
of images exhibiting close but different resolutions (e.g. 30 × 30 are less complex than in thermographic damage detection, depend-
ing significantly on time evolution. While the transient response in
pixels versus 32 × 32 pixels) might contain a very different number
thermographic damage detection requires higher order neural network
of non-black pixels that indeed contribute to the improvement of the
models, the steady-state response of a fatigue damage hotspot allows
corresponding ML models.
the use of a simpler ML model, which provides higher efficiency in such
applications.
4. General remarks 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
The ML model proved itself quite robust, where as little as 100 accuracy improved significantly up to about 100 training images, after
thermal images suffice for training to get a good performance. This which the improvement increased more slowly and asymptotically. This
suggests that building a ML model based on experimentally measured comparatively small training set proves the efficiency of the method.
IR images of defects in epoxy composites is feasible, especially if one Doubling the number of independent parameters from two to four
uses data augmentation also including artificially generated images. significantly expands the solution space, which affects the prediction
However, this ML model also has some limitations. For instance, accuracy of the elliptical disc source model compared to the point
no defect detection can be done to identify its exact location in the source. The prediction accuracy improved as the size of the elliptical
thermal image, and only one defect is expected to be present in the heat source dimensions increased. For a small mean relative heat source
same image. The geometry of real defects of composite materials can size of 𝑎𝑚𝑒𝑎𝑛 ∕𝑅 = 2%, the prediction was poor, while a larger relative
be very different from a simple ellipse. Therefore, the current ML model 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
could be further extended to address these problems, which is the
lower variation of the surface pixel temperature values in the images
subject of a future investigation.
and accordingly makes it more challenging for the ML algorithm to
predict the parameter. By the same argument, the lower prediction
5. Conclusion 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
In this work, it was shown that ML can be used to predict four key more than twice as large as 𝑘𝑦𝑦 .
physical and geometric parameters of a fatigue damage hotspot from Higher image resolutions resulted in better performance of the ML
surface thermal images, namely heat source intensity, depth, and size. model and the observed trend was more choppy in low resolution
7
A. Sarhadi et al. Composite Structures 295 (2022) 115786
regions, which is due to larger variations in the number of non-black [9] Chulkov AO, Nesteruk DA, Vavilov VP, Shagdirov B, Omar M, Siddiqui AO,
pixels effectively used to train the model. Prasad YLVD. Automated procedure for detecting and characterizing defects in
gfrp composite by using thermal nondestructive testing. Infrared Phys Technol
Future work will address arbitrarily shaped heat sources present at
2021. http://dx.doi.org/10.1016/j.infrared.2021.103675.
different depths of the composite material by utilizing convolutional [10] Marani Roberto, Palumbo Davide, Galietti Umberto, D’Orazio Tiziana. Deep
neural networks. learning for defect characterization in composite laminates inspected by
step-heating thermography. Opt Lasers Eng 2021. http://dx.doi.org/10.1016/j.
CRediT authorship contribution statement 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
Ali Sarhadi: Writing – original draft, Preparation of the dataset, J Fatigue 2021;150:106326. http://dx.doi.org/10.1016/j.ijfatigue.2021.106326.
Comsol simulations. Rodrigo Q. Albuquerque: ML modelling, Pa- [12] Gornet Laurent, Wesphal Ophélie, Burtin Christian, Bailleul Jean Luc, Rozy-
per writing and correction, Discussion. Martin Demleitner: Paper cki Patrick, Stainier Laurent. Rapid determination of the high cycle fatigue
limit curve of carbon fiber epoxy matrix composite laminates by thermography
correction, Discussion. Holger Ruckdäschel: Paper correction, Discus-
methodology: Tests and finite element simulations. In: Procedia Engineering.
sion. Martin A. Eder: FEM simulations, Paper writing and correction, 2013, http://dx.doi.org/10.1016/j.proeng.2013.12.123.
Discussion. [13] Peyrac Catherine, Jollivet Thomas, Leray Nolwenn, Lefebvre Fabien, West-
phal Ophélie, Gornet Laurent. Self-heating method for fatigue limit determination
Declaration of competing interest 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
The authors declare that they have no known competing finan- in fatigue damage problems. In: AIP Conference Proceedings. 2018, http://dx.
cial interests or personal relationships that could have appeared to doi.org/10.1063/1.5033668.
influence the work reported in this paper. [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.
Acknowledgements [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.
DTU Wind Energy is grateful for the financial support of this publi- Mech Compos Mater 2015. http://dx.doi.org/10.1007/s11029-015-9458-y.
cation by VILLUM FONDEN with grant number 36050. Furthermore, [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.
DTU Wind Energy acknowledges the financial support through the
1631/jzus.A2000380.
EUDP project ReliaBlade with grant number 64018-0068. [18] Nanthakumar SS, Lahmer T, Zhuang X, Zi G, Rabczuk T. Detection of
material interfaces using a regularized level set method in piezoelectric struc-
Appendix A. Supplementary data 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
Supplementary material related to this article can be found on- problems. J. Inverse Ill-Posed Prob 2010. http://dx.doi.org/10.1515/JIIP.2010.
line at https://doi.org/10.1016/j.compstruct.2022.115786. More de- 027.
tails about model performances and general trends are available as PDF. [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/
References 17415977.2018.1518444,
[22] Vamaraju Janaki, Sen Mrinal K. Unsupervised physics-based neural networks
[1] Persson Erik, Eriksson Ingvar, Zackrisson Leif. Effects of hole machining defects for seismic migration. Interpretation 2019. http://dx.doi.org/10.1190/INT-2018-
on strength and fatigue life of composite laminates. Composites A 1997. http: 0230.1.
//dx.doi.org/10.1016/S1359-835X(96)00106-6. [23] Deng S, Hwang Y. Applying neural networks to the solution of forward and
[2] Hörrmann Susanne, Adumitroaie Adi, Viechtbauer Christoph, Schagerl Martin. inverse heat conduction problems. Int J Heat Mass Transfer 2006. http://dx.doi.
The effect of fiber waviness on the fatigue life of CFRP materials. Int J Fatigue org/10.1016/j.ijheatmasstransfer.2006.06.009.
2016. http://dx.doi.org/10.1016/j.ijfatigue.2016.04.029. [24] Anitescu Cosmin, Atroshchenko Elena, Alajlan Naif, Rabczuk Timon. Artificial
[3] Bard S, Demleitner M, Häublein M, Altstädt V. Fracture behaviour of prepreg neural network methods for the solution of second order boundary value prob-
laminates studied by in-situ sem mechanical tests. In: Procedia Structural lems. Comput Mater Continua 2019;59(1):345–59. http://dx.doi.org/10.32604/
Integrity, Vol. 13. 2018, http://dx.doi.org/10.1016/j.prostr.2018.12.299. cmc.2019.06641, URL http://www.techscience.com/cmc/v59n1/27936.
[4] Bakis Gökhan, Wendel Jan Felix, Zeiler Rico, Aksit Alper, Häublein Markus, [25] Fang Qiang, Maldague Xavier. A method of defect depth estimation for simulated
Demleitner Martin, Benra Jan, Forero Stefan, Schütz Walter, Altstädt Volker. infrared thermography data with deep learning. Appl Sci 2020;10(19):6819.
Mechanical properties of the carbon nanotube modified epoxy–carbon fiber http://dx.doi.org/10.3390/app10196819.
unidirectional prepreg laminates. Polymers 2021. http://dx.doi.org/10.3390/ [26] Saeed Numan, King Nelson, Said Zafar, Omar Mohammed A. Automatic defects
polym13050770. detection in CFRP thermograms, using convolutional neural networks and
[5] Jolly Mr, Prabhakar A, Sturzu B, Hollstein K, Singh R, Thomas S, Foote P, transfer learning. Infrared Phys Technol 2019;102:103048. http://dx.doi.org/
Shaw A. Review of non-destructive testing (NDT) techniques and their appli- 10.1016/j.infrared.2019.103048, URL https://www.sciencedirect.com/science/
cability to thick walled composites. In: Procedia CIRP. 2015, http://dx.doi.org/ article/pii/S1350449519303135.
10.1016/j.procir.2015.07.043. [27] COMSOL Inc. Comsol. 2020, URL http://www.comsol.com/products/
[6] Kempf M, Skrabala O, Altstädt V. Acoustic emission analysis for characterisation multiphysics/.
of damage mechanisms in fibre reinforced thermosetting polyurethane and epoxy. [28] MATLAB 2021. Documentation, version 7.10.0 (R2020a). Natick, Massachusetts:
Composites B 2014. http://dx.doi.org/10.1016/j.compositesb.2013.08.080. The MathWorks Inc.; 2021.
[7] Koyama Kiyoshi, Hoshikawa Hiroshi, Kojima Gouki. Eddy current nondestructive [29] Vovk V. Empirical Inference. Springer; 2013, p. 105–16. http://dx.doi.org/10.
testing for carbon fiber-reinforced composites. J. Pressure Vessel Technology, 1007/978-3-642-41136-6.
Transactions of the ASME 2013. http://dx.doi.org/10.1115/1.4023253. [30] Hofmann T, Schölkopf B, Smola AJ. Kernel methods in machine learning. Ann
[8] Bang Hyun Tae, Park Solmoi, Jeon Haemin. Defect identification in composite Statist 2008;36(3):1171–220. http://dx.doi.org/10.1214/009053607000000677.
materials via thermography and deep learning techniques. Compos Struct 2020.
http://dx.doi.org/10.1016/j.compstruct.2020.112405.