Proceedings of the ASME 2023

International Mechanical Engineering Congress and Exposition

IMECE2023
October 29-November 2, 2023, New Orleans, Louisiana

IMECE2023-112407

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

PREDICTION OF DELAMINATION LOCATION IN COMPOSITE STRUCTURES WITH DIFFERENT PLY
ORIENTATIONS: A FRAMEWORK INTEGRATING FINITE ELEMENT SIMULATION AND DEEP LEARNING

Junyan He1, ∗ , Linqi Zhuang1 , Adarsh Chaurasia1 , Ali Najafi1

1
Ansys Inc., Houston, TX, USA

ABSTRACT be able to detect the in-service damage when it occurs, provide

In this paper, a framework integrating guided-wave-based the location information of the damage and evaluate the damage
finite element (FE) simulation and deep learning is developed to severity. Data-driven approaches [1, 3–5] by leveraging a large
predict the location of internal delamination within a composite flow of existing sensor data and advanced AI/ML algorithms have
plate with different layup configurations. The data needed for shown promises in building the next-generation SHM systems.
training the prediction model are obtained through guided-wave- For a data-driven SHM system to be effective and accurate, a
based FE simulations. Convolutional neural network (CNN) significant amount of data are needed to train the model prop-
models are then built and trained to predict the location of the erly. To generate sufficient data, testing alone could be costly
existing delamination. The possibility of using transfer learning and unpractical, especially considering the different layups, ma-
is explored to account for composite plates with different layup terial variations, and types of in-service damages that could oc-
combinations while minimizing additional data needed for CNN cur within a composite structure. Compared to tests, simulations
model training. It is found that once a baseline CNN model is could provide a potentially cost-effective and efficient way to gen-
trained by the data extracted from a specific composite layup erate needed data. In the past, simulations, especially FE-based
case, transfer learning can significantly reduce the amount of simulations, have been used to understand the characteristics of
data required to achieve similar levels of prediction accuracy for guided wave propagation, wave-damage interactions, etc. [6–8].
other scenarios with different layup orientations, thus greatly im- Recently, researchers [9, 10] are also exploring the possibility of
proving the efficiency and versatility of the proposed framework. combining simulation and AI/ML algorithms for damage detec-
tion and classifications. As promising as it looks, however, there
Keywords: Guided wave, Composites Impact Dam-
are also some challenges for simulation, and one of the biggest
age, Structural health monitoring, Convolutional neural
ones is related to scaling. Despite the recent advancement in
networks, Transfer learning
high-performance computing, simulating the whole composite
structures such as an aircraft or wind turbine could still be com-
NOMENCLATURE putationally extensive. In terms of SHM, to resolve the scaling
FRPC Fiber Reinforced Polymer Composite issue, some researchers are looking into numerical methods that
FE Finite element are less computationally intensive compared to the FE method
SHM Structural health monitoring [7, 9], or simplified FE modeling method by modeling the impact
PZT Piezoelectric transducer region with homogenized but degraded stiffness properties[11].
CWT Continuous wavelet transform For a composite structure, in-service damages could occur in dif-
CNN Convolutional neural networks ferent forms at different locations with different composite layup
TL Transfer learning combinations. The training data must include all these situations
to properly train the AI/ML model. To the best of the authors’
1. INTRODUCTION knowledge, few studies focus on minimizing the amount of data
Guided wave is one of the most actively investigated struc- needed during the model training process. Aiming to improve
tural health monitoring (SHM) technologies [1] that could be a the understanding of this specific area, in this paper, we are ex-
key enabler for the damage-tolerant design of composite struc- ploring the usage of transfer learning to reduce the amount of
tures [2]. An ideal SHM system for a composite structure should training data needed. As a proof of concept, we will demonstrate
using the proposed framework to predict delamination location
∗ Corresponding author: [email protected]

1 Copyright © 2023 by ASME

 SYMM A-A Cross-sectional view

Ply#1 0˚
Delamination
PZT φ10mm Ply#2 θ1
X φ19mm Ply#3 θ2
Y 60mm Ply#4 θ3
Z A A
Ply#5 θ3
Ply#6 θ2
Ply#7 θ1
Ply#8 0˚

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

60mm

FIGURE 1: SCHEMATIC OF THE FE MODEL INCLUDING THE LAM-

INATE PLATE, PZT, AND A CIRCULAR DELAMINATION. THE COM-
POSITE LAYUP IS PARAMETERIZED BY THREE ANGLES θ1 , θ2 ,
AND θ3 .

in composite structures with different ply orientations.

FIGURE 2: A TYPICAL MESH OF THE LAMINATE PLATE WITH DE-
2. FE MODEL AND CNN SETUP LAMINATION.
2.1 General FE model settings
The FE model setting and material properties follow the
work of Zhuang et al. [8]. A schematic of the FE model and the second and third layers with randomly generated in-plane lo-
delamination location is shown in Fig. 1. An 8-layer laminate cations. Three-dimensional quadratic hexahedral elements with
plate is modeled, and the base unidirectional composite mate- a nominal mesh size of 2 𝑚𝑚 are used to discretize the sim-
rial properties are shown in Tab. 1. In the baseline case, we ulation domain to achieve computational efficiency for massive
choose 𝜃 1 = 0◦ and 𝜃 2 = 𝜃 3 = 90◦ , leading to a [0◦2 /90◦2 ] 𝑠 simulation data collections while maintaining satisfactory accu-
laminate. To evaluate the generalization capability of the de- racy compared to the results obtained using ideal element size of
veloped CNN model, two different composites with commonly 0.8 𝑚𝑚 in previous work [8]. A typical mesh is shown in Fig. 2.
adopted layup combinations, defined by (1) 𝜃 1 = 0◦ , 𝜃 2 = 45◦ A coupled-field transient simulation with a time step of 0.2 𝜇s is
and 𝜃 3 = −45◦ ([0◦2 /45◦ /−45◦ ] 𝑠 ) and (2) 𝜃 1 = 90◦ , 𝜃 2 = 45◦ performed using the FE model in Ansys.
and 𝜃 3 = −45◦ ([0◦ /90◦ /45◦ /−45◦ ] 𝑠 ), were also simulated and
TABLE 2: MATERIAL PROPERTIES FOR PZT-5H TRANSDUCER (IN
discussed in Sec. 2.2. As displayed in Fig. 1, the simulation do-
IEEE FORMAT)
main is a square with an edge length of 60 𝑚𝑚 and a thickness
of 0.92 𝑚𝑚, where each layer has a constant thickness of 0.115 Anisotropic Elastic [GPa]
𝑚𝑚. The origin of the FE model is located at the midpoint of the 𝐸 11 = 𝐸 22 =127, 𝐸 33 =117
left edge of the plate. Symmetry is applied on the left face of the 𝐸 12 = 𝐸 21 =80, 𝐸 13 =85
simulation domain, and infinite boundary conditions (INFIN 257 𝐸 55 = 𝐸 66 =24, 𝐸 44 =23
elements in Ansys) are applied on all truncated side surfaces to
represent an infinite laminate plate with a finite model. Density [𝑘𝑔/𝑚 3 ]
7500
TABLE 1: MATERIAL PROPERTIES FOR COMPOSITE PLATE Anisotropic Relative Permittivity (at constant strain)
𝜖11 = 𝜖22 =1700, 𝜖33 =1470
Young’s Modulus [GPa]
𝐸 𝑥 =121, 𝐸 𝑦 = 𝐸 𝑧 =8.6 Piezoelectric Coefficients (at constant stress) [𝐴𝑚/𝑠2 ]
𝑒 31 = 𝑒 32 =-6.5, 𝑒 15 = 𝑒 24 =17, 𝑒 33 =23.3
Shear Modulus [GPa]
𝐺 𝑥 𝑦 = 𝐺 𝑥𝑧 =4.7, 𝐺 𝑦𝑧 =3.1
Poisson’s Ratio
𝜈 𝑥 𝑦 = 𝜈 𝑥𝑧 =0.27, 𝐺 𝑦𝑧 =0.4 2.2 Training Data Generation and CNN Architecture
Using the FE simulation framework described in Sec. 2.1, a
Density [𝑘𝑔/𝑚 3 ] total of 1500 simulations were conducted with different in-plane
1490 delamination locations to train the baseline CNN model on the
[0◦2 /90◦2 ] 𝑠 layup. 300 simulations were conducted similarly on
To model the guided wave excited by a piezoelectric trans- each of the [0◦2 /45◦ /−45◦ ] 𝑠 and [0◦ /90◦ /45◦ /−45◦ ] 𝑠 layups for
ducer (PZT), a circular PZT with a radius of 9.5 𝑚𝑚 is placed transfer learning that will be discussed later. The X and Y co-
at the left side of the FE model along the symmetry line. The ordinates of the delamination centers were randomly sampled in
material properties of the PZT are shown in Tab. 2. The PZT ranges [−21, 21] 𝑚𝑚 and [−51, −18.5] 𝑚𝑚, respectively. The
is perfectly bonded with the laminate plate and is given a time- out-of-plane displacement at the top surface of the FRPC plate
varying excitation voltage signal to excite vibration in the plate. A was sampled at two sets of sensor locations. The first set of sen-
circular delamination with a radius of 5mm is placed in between sors is located at the four corners of a rectangle (𝑋 = ±26 𝑚𝑚,

2 Copyright © 2023 by ASME

 60mm

Delamination
centers FIGURE 5: CWT OF A TYPICAL SENSOR ARRAY SIGNALS.

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

Sensor array 1
Sensor array 2
is a 16 × 256 × 𝑁 𝑠𝑖𝑔𝑛𝑎𝑙 image, where 𝑁 𝑠𝑖𝑔𝑛𝑎𝑙 is the number of
60mm
sensor signals to include. Two different CNN models were built
FIGURE 3: RANDOMLY SAMPLED DELAMINATION LOCATIONS and trained to study the effect of including more sensor signals,
AND THE TWO SENSOR ARRAYS. one with 4 signals from array 1 and the other with 8 signals
from both arrays. Out of the 1500 baseline simulation cases,
1200 (80%) were used in training, and 300 (20%) were used in
testing. Once the baseline models were trained, transfer learning
11 7 3 was conducted with 720 data points. The transferred model was
tested on 180 data points unseen during transfer training, again
10 6 2 leading to an 80/20 data split. In both data sets, 31 of the data
points came from each baseline layup and the two new layups, so
the data contribution is equal. In all cases, the sparse categorical
9 5 1 cross entropy was used as the loss function, which is typical
for classification tasks. The CNN models were trained for 200
8 4 0 epochs with a batch size of 500.

3. RESULTS AND DISCUSSION

3.1 Baseline CNN model performance
FIGURE 4: A 4 × 3 GRID USED TO CLASSIFY DELAMINATION LO- The classification accuracy in the training and testing dataset
CATION. is shown in Fig. 7 for the CNN models using 4 and 8 sensor
signals, respectively. From the training history, it is obvious that
the model with 4 input sensor signals shows noticeable overfitting
𝑌 = −13.5/ −56𝑚𝑚), while the second is located at the midpoints after about 100 epochs. In comparison, the overfitting on the
of the four rectangle edges, as shown in Fig. 3. The sensor data model with 8 signals is much lower. This result indicates that
were extracted from the output database via pyDPF [12], a pub- including more sensor signals can improve the training process
licly available Ansys python library for post-processing simula- of the CNN model.
tion results (as a part of pyAnsys). 125 data points were recorded The highest classification accuracy on the testing dataset is
for each sensor throughout the duration of the FE simulation. 86.0% and 93.0% for using 4 and 8 sensor signals, respectively.
In the present paper, the composite plate was divided into a The confusion matrices for the two models are shown in Fig. 8.
4 × 3 grid to predict the delamination location, as shown in Fig. 4. To highlight the effect of including more sensor signals on the
A delamination is considered within a particular grid cell if its classification accuracy, the cells that have a significant change
center falls within the cell. Following the proposal of Badabagni in prediction accuracy between the two models are shown in
and Talreja [10], CWT was used to transform the time-domain Tab. 3. From the confusion matrices, we see that including four
sensor signals into two-dimensional images containing time and more sensors in the CNN provides more physical information
frequency information. To focus on the presence of the delami- about the delamination, leading to a more accurate classification
nation, sensor measurements on a pristine plate with no delami- of the delamination location. The highest prediction accuracy
nations were simulated in a separate simulation. We subtracted of 93% indicates that the performance of the proposed CNN
this pristine reference response from the sensor signals obtained model is satisfactory. However, when the individual cells are
in the simulations with delaminations to obtain the difference sig- inspected, it is noted that increasing the number of input sensors
nal. The difference signal was linearly interpolated to 256 time does not always increase classification accuracy for all cells. For
steps and sent to CWT with 16 wavelets of different frequencies, example, in Tab. 3, we see that the classification accuracy for cell
converting a 1 × 125 time-domain signal into a 16 × 256 two- 2 decreased by 17.2% (incorrectly classified as cell 1, one of its
dimensional image. A typical CWT for sensor array 1 (Fig. 3) is neighboring cells) after more sensor signals were added to the
shown in Fig. 5. CNN, although for many other cells, the classification accuracy
Since CWT provides two-dimensional images as outputs, was increased. This accuracy decrease on certain cells could
a CNN model was used to extract the spatial features in the be due to the delaminations near the cell boundaries (e.g., near
signal images. The architecture of the proposed CNN is shown cells 1 and 2), where the additional sensors (sensor array 2) were
in Fig. 6. The input data for each delamination center location placed. For those cases, a small delamination shift that leads to a

3 Copyright © 2023 by ASME

 100

80

Accuracy [%]
60

40
Training accuracy, 4 sensors
20
Testing accuracy, 4 sensors
0 25 50 75 100 125 150 175 200
Epoch

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

(a) 4 sensor signals
100

80

Accuracy [%]
60

Input CWT 40
Training accuracy, 8 sensors
signals 20
Testing accuracy, 8 sensors
0 25 50 75 100 125 150 175 200
Epoch
(b) 8 sensor signals

Three FIGURE 7: CNN MODEL TRAINING HISTORIES OF THE TWO CNN

MODELS.
convolution
layers
change in its cell classification index may not lead to a significant
difference in the measured signal at the nearest sensor; therefore,
the added sensors cannot provide more insights and can even lead
to increased confusion.
Normalize
TABLE 3: PERFORMANCE COMPARISON ON SELECTED CELLS

Flatten Cell index 2 4 6 9

4 sensors [%] 93.1 66.7 63.0 66.7
8 sensors [%] 75.9 100.0 96.3 80.0
Difference [%] -17.2 33.3 33.3 13.3
Three dense
layers with
drop-out 3.2 Transfer learning performance
One of the key advantages of a laminated composite is that
the ply orientation can be engineered to meet specific design re-
quirements. As a result, the composite layup design may differ
across various locations within the same structure. For the pro-
Output posed framework to be practical for real-world applications, the
logits prediction model must be effective for different composite layup
designs, not just a fixed one, which requires additional training
FIGURE 6: CNN MODEL ARCHITECTURE WITH 4 INPUT SENSOR data that include all the layup variations for the CNN model. To
SIGNALS. avoid going through the same data-intensive training process for
the baseline case discussed in the previous section, we explored
transfer learning to reduce the number of simulation cases needed
for different composite layups to generate accurate neural network
predictions in this section. Since a CNN model trained using 8
input sensor signals provides higher delamination identification
accuracy, all models in the present section were trained using 8
sensor signals.
To compare the effectiveness of transfer learning from a
previously trained CNN, a CNN model was first trained on the
720 training data points (240 data points for each layup case)
starting from randomly initialized weights and bias. Then, the
pre-trained weights and bias from Sect. 3.1 were used as the

4 Copyright © 2023 by ASME

 100
0 88.0%

80

Accuracy [%]
84.6%

2 93.1%
60

Training accuracy, from random

40
80.0% Testing accuracy, baseline layup
Testing accuracy, [02/45/ 45]s
Ground truth cell index
4 66.7% 20
Testing accuracy, [0/90/45/ 45]s
96.2% 0 25 50 75 100 125 150 175 200
Epoch
6 63.0%

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

84.0% (a) Training from scratch
8 90.6% 100

90

Accuracy [%]
66.7%
80
10 100.0%
70

60
Training accuracy, from random
91.3%
Testing accuracy, baseline layup
0 2 4 6
Predicted cell index
8 10 50 Testing accuracy, [02/45/ 45]s
40 Testing accuracy, [0/90/45/ 45]s
0 25 50 75 100 125 150 175 200
(a) 4 sensor signals Epoch
0 88.0%
(b) Transfer learning from a pre-trained model
92.3%

2 75.9%
FIGURE 9: CNN MODEL TRAINING HISTORIES. TESTING ACCU-
96.7% RACY FOR THREE DIFFERENT COMPOSITE LAYUPS ARE RE-
Ground truth cell index

4 100.0% PORTED SEPARATELY.

96.2%

6 96.3%
The highest classification accuracy of both CNN models on
88.0%
the baseline and new layup datasets is reported in Tab. 4. As
8 90.6%
a comparison, the randomly initialized and pre-trained models’
80.0%
prediction accuracy before any training was also calculated and
10 100.0%
included in Tab. 4. The confusion matrices for the two models as
100.0%
0 2 4 6 8 10
evaluated on the new layup data, are shown in Fig. 10.
Predicted cell index From Tab. 4, it is obvious that transfer learning from a pre-
(b) 8 sensor signals viously trained model leads to the highest prediction accuracy on
the two new composite layups, especially compared to a CNN
FIGURE 8: CONFUSION MATRICES FROM THE TWO CNN MOD- model trained from scratch. From the confusion matrices, we
ELS. again confirm that transfer learning is more effective than train-
ing the model from scratch when identical training data was used.
It is worth noting that when inspecting the prediction accuracy
starting point for transfer learning on identical training data. The for each index, both models show consistently lower accuracy on
classification accuracy is shown in Fig. 9 for the CNN models bins 5 and 6, which are located at the center of the laminate plate
trained from random and pre-trained weights. In this case, the and are farthest away from any sensor locations. This indicates
training accuracy was evaluated using data points from all three that the prediction accuracy can be improved by adding a sensor at
composite layups, while the testing accuracy in three layups is the center of the plate to bring in additional physical information.
reported separately. The training histories show that the randomly
initialized CNN model could learn with a limited amount of 4. CONCLUSION
data, although showing some degree of overfitting after around A guided-wave-based structural health monitoring system
60 epochs. In comparison, the CNN model resumed from a was simulated using 3D coupled-physics transient FE models.
previously trained one showed overfitting much earlier, at around Out-of-plane displacements from different sensor locations were
15 epochs. This is expected as the sensor signals obtained from extracted. Continuous wavelet transform was used to transform
the new composite layup differed from those obtained in the the time-domain signals into two-dimensional images, which
baseline layup, and the CNN model was already trained on the were used to train a convolutional neural network to identify
baseline data to obtain accurate delamination identification. Note delamination locations. It was found that for a given composite
that in both CNN models, the prediction accuracy on the baseline layup, having 8 sensor input signals instead of 4 provided addi-
layup is higher than the new layups introduced from transfer tional physical information to the CNN model, thus allowing it to
learning, even for the model trained completely from scratch. It is achieve higher prediction accuracy.
also important to note that the prediction accuracy on the baseline The current study also investigated the use of transfer learn-
layup data has only decreased slightly during transfer learning. In ing to extend accurate delamination location prediction to new
contrast, the prediction accuracy on the new layup data increased composite layups with limited additional data. With only 300
significantly. This indicates that catastrophic forgetting has not additional cases (compared to the 1500 cases used in the orig-
occurred during transfer learning. inal training) for each new layup, transfer learning was able to

5 Copyright © 2023 by ASME

 TABLE 4: CLASSIFICATION ACCURACY COMPARISON OF TWO MODELS BEFORE AND AFTER TRAINING

[0◦2 /90◦2 ] 𝑠 [0◦2 /45◦ /−45◦ ] 𝑠 [0◦ /90◦ /45◦ /−45◦ ] 𝑠

Randomly initialized
(before training) [%] 11.7 10.0 11.7

Randomly initialized

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

(after training) [%] 72.9 73.3 65.0
Pre-trained
(before training) [%] 91.7 33.3 33.3

Pre-trained
(after training) [%] 85.6 81.7 73.3

0 88.9% on unseen composite layups.

71.4%

2 66.7% REFERENCES
80.0% [1] Güemes, Alfredo, Fernandez-Lopez, Antonio, Pozo, An-
Ground truth cell index

4 83.3% gel Renato and Sierra-Pérez, Julián. “Structural health

22.2%
monitoring for advanced composite structures: a review.”
6 76.9%
Journal of Composites Science Vol. 4 No. 1 (2020): p. 13.
40.0%
[2] Talreja, Ramesh and Phan, Nam. “Assessment of damage
8 50.0%
tolerance approaches for composite aircraft with focus on
92.3%
barely visible impact damage.” Composite Structures Vol.
10 71.4%
219 (2019): pp. 1–7.
100.0%
[3] Mardanshahi, A, Nasir, V, Kazemirad, S and Shokrieh,
0 2 4 6 8 10 MM. “Detection and classification of matrix cracking in
Predicted cell index
laminated composites using guided wave propagation and
(a) Training from scratch artificial neural networks.” Composite Structures Vol. 246
0 100.0% (2020): p. 112403.
85.7% [4] Qian, Cheng, Ran, Yunmeng, He, Jingjing, Ren, Yi, Sun,
2 66.7% Bo, Zhang, Weifang and Wang, Rongqiao. “Application
80.0% of artificial neural networks for quantitative damage detec-
tion in unidirectional composite structures based on Lamb
Ground truth cell index

4 91.7%

44.4%
waves.” Advances in Mechanical Engineering Vol. 12 No. 3
6 61.5%
(2020): p. 1687814020914732.
60.0%
[5] Wu, Jun, Xu, Xuebing, Liu, Cheng, Deng, Chao and Shao,
8 75.0%
Xinyu. “Lamb wave-based damage detection of compos-
ite structures using deep convolutional neural network and
92.3%
continuous wavelet transform.” Composite Structures Vol.
10 85.7%
276 (2021): p. 114590.
100.0%
0 2 4 6 8 10
[6] Mitra, Mira and Gopalakrishnan, S. “Guided wave based
Predicted cell index
structural health monitoring: A review.” Smart Materials
(b) Transfer learning from a pre-trained and Structures Vol. 25 No. 5 (2016): p. 053001.
model [7] Leckey, Cara AC, Wheeler, Kevin R, Hafiychuk, Vasyl N,
Hafiychuk, Halyna and Timuçin, Doğan A. “Simulation
FIGURE 10: CONFUSION MATRICES FROM THE TWO CNN MOD-
of guided-wave ultrasound propagation in composite lam-
ELS. SCORES WERE CALCULATED ON THE NEW LAYUP DATA.
inates: Benchmark comparisons of numerical codes and
experiment.” Ultrasonics Vol. 84 (2018): pp. 187–200.
[8] Zhuang, Linqi, Chaurasia, Adarsh and Najafi, Ali. “Im-
leverage previously learned knowledge and produce much higher pact Damage Evaluations in a Composite Laminate Using
prediction accuracy on unseen layups compared to a CNN model Guided Wave-Based Simulation.” ASME International Me-
trained from scratch using identical data. This result indicates chanical Engineering Congress and Exposition, Vol. 86656:
that transfer learning is an effective strategy to reduce the amount p. V003T04A024. 2022. American Society of Mechanical
of new data needed to accurately predict delamination locations Engineers.

6 Copyright © 2023 by ASME

 [9] Rautela, Mahindra and Gopalakrishnan, S. “Ultrasonic 2022, Vol. 12046: pp. 159–166. 2022. SPIE.
guided wave based structural damage detection and local- [11] Deng, Peiwen, Saito, Osamu, Okabe, Yoji and Soejima,
ization using model assisted convolutional and recurrent Hideki. “Simplified modeling method of impact damage for
neural networks.” Expert Systems with Applications Vol. numerical simulation of Lamb wave propagation in quasi-
167 (2021): p. 114189. isotropic composite structures.” Composite Structures Vol.
[10] Badabagni, Sai and Talreja, R. “Damage classification 243 (2020): p. 112150.
based on stiffness reduction in cross-ply laminates with con- [12] “pyDPF-Core.” https://dpf.docs.pyansys.com/version/

Downloaded from http://asmedigitalcollection.asme.org/IMECE/proceedings-pdf/IMECE2023/87615/V004T05A014/7238787/v004t05a014-imece2023-112407.pdf by Brunel University user on 08 February 2025

volution neural networks.” Sensors and Smart Structures stable/. Accessed: 04/07/2023.
Technologies for Civil, Mechanical, and Aerospace Systems ,

7 Copyright © 2023 by ASME