Advanced Deep Learning Techniques for Industry 4.0: Application to

Mechanical Design and Structural Health Monitoring

Fakhreddine Ababsa

PIMM, Arts et Métiers ParisTech, CNRS, CNAM, HESAM University, Paris, France

Keywords: Generative Design, Deep Learning, Additive Manufacturing, Topology Optimization, Structural Health

Monitoring (SHM), MLP, GAN.

Abstract: Nowadays, Deep Learning (DL) techniques are increasingly employed in industrial applications. This paper

investigate the development of data-driven models for two use cases: Additive Manufacturing-driven

Topology Optimization and Structural Health Monitoring (SHM). We first propose an original data-driven

generative method that integrates the mechanical and geometrical constraints concurrently at the same

conceptual level and generates a 2D design accordingly. In this way, it adapts the geometry of the design to

the manufacturing criteria, allowing the designer better interpretation and avoiding being stuck in a time-

consuming loop of drawing the CAD and testing its performance. On the other hand, SHM technique is

dedicated to the continuous and non-invasive monitoring of structures integrity, ensuring safety and optimal

performances through on-site real-time measurements. We propose in this work new ways of structuring data

that increase the accuracy of data driven SHM algorithms and that are based on the physical knowledge related

with the structure to be inspected. We focus our study on the damage classification step within the aeronautic

context, where the primary objective is to distinguish between different damage types in composite plates.

Experimental results are presented to demonstrate the effectiveness of the proposed approaches.

1 INTRODUCTION

Many industrial sectors have embraced the concepts

and technologies of Industry 4.0. These technologies

offer the opportunity to promote efficient, on-demand

production through digitisation and the integration of

intelligent machines. This transition is primarily

driven by technological advances, which include

artificial intelligence (AI) and machine learning,

sensor networks, Internet of Things (IoT)

technologies, cloud computing, additive

manufacturing and the accessibility of large data sets

that can be leveraged by these innovations. However,

the degree of adoption of AI technologies for the

industry 4.0 varies considerably from one industrial

sector to another (Khelalef et al., 2019). Many

research studies have been carried out last years on

the use of AI in Industry 4.0. They highlight how the

different AI technologies aim to solve problems

concerning the acquisition, processing, modelling

and interpretation of data results. For example, in

(Nath et al., 2021) an AI-based feature engineering

framework was introduced for rotor fault diagnosis.

This study utilized a range of AI/ML methodologies

to extract significant data features, enabling the

detection of rotor faults. (Lockner et al., 2021)

introduced an enhanced injection molding process

through transfer learning. The authors employed

neural networks and genetic algorithms to optimize

the molding process, meeting customer demands for

waste reduction and operational speed. In a study

(Oehlmann et al., 2021), researchers employed an

artificial neural network (ANN) to predict the

pressure inside the nozzle of an FFF printer with the

aim of optimizing the printing process. In (Mohan et

al., 2021), a machine learning-based approach for

predictive maintenance was proposed. They

introduced a data-driven regression system to

transition from Industry 3.0 to Industry 4.0 with

minimal alterations to the existing setup. In another

study (Arden et al., 2021), the necessity of

implementing Industry 4.0 and AI-driven processes in

the pharmaceutical sector was examined. They

suggested that the pharmaceutical industry should

https://orcid.org/ 0000-0003-3862-2449

Ababsa, F.

Advanced Deep Learning Techniques for Industry 4.0: Application to Mechanical Design and Structural Health Monitoring.

DOI: 10.5220/0012364300003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 535-542

ISBN: 978-989-758-680-4; ISSN: 2184-433X

535

Figure 1: DL-AM-TO Training procedure.

leverage emerging cutting-edge technologies to

transition into the next generation of manufacturing.

In this paper, we aim to elaborate on two use cases of

AI within the scope of Industry 4.0, offering a

complementary perspective to the state of the art. The

first case focuses on harnessing Generative

Adversarial Networks (GANs) for design,

optimization, and numerical validation in the field of

additive manufacturing. The second case concerns the

utilization of a Convolutional Neural Network

(ConvNet) for the classification of damage cases in

the context of Structural Health Monitoring (SHM) of

aeronautical composite structures.

The rest of the manuscript is organized as follows:

Section 2 describes the methodology proposed for the

construction of the design algorithm using GAN’s.

Section 3 provides the details of the default

classification method for SHM. Section 4 gives

concluding remarks and suggestions for future work.

2 DL FOR ADDITIVE

MANUFACTURING-DRIVEN

TOPOLOGY OPTIMIZATION

In this work, we propose an advanced approach

utilizing Deep Learning (DL) to incorporate not only

mechanical constraints but also geometric constraints

inherent to Additive Manufacturing (AM) at a

conceptual level. DL's intrinsic capability to discern

spatial correlations obviates the necessity for deriving

explicit analytical formulas for geometric constraints,

relying instead on the availability of a substantial

dataset comprising numerous instances that

encapsulate these constraints.

We propose a topological optimisation (TO)

approach for data-driven AM (DLAM- TO). DL-AM-

TO is a generative model that takes mechanical

(Boundary conditions (BC), loads (F), and the volume

fraction (V)) and geometrical conditions (the

minimum thickness ( 𝑡ℎ



), the maximum length

(𝑙𝑒𝑛



), the minimum overhang (𝜃



), the number

of bars ( 𝑁𝑏𝑟



)) as inputs and generates a 2D

structure following these constraints. It is trained

within a five-discriminator-GAN (Goodfellow et al.

2014) framework consisting of a generator (DL-AM-

TO) and five discriminators: the traditional

adversarial discriminator and four geometric

discriminators, a bar counter,𝑡ℎ



, 𝑙𝑒𝑛



and 𝜃



predictors (Fig.1).

The four geometric constraints were chosen

among the existing set of constraints as a use case to

validate the proposed methodology.

2.1 Architecture

DL-AM-TO inherits the residual-convolutional

encoder-decoder architecture (Zhang et al., 2018)

presented in our previous work (Almasri et al., 2021)

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

536

Figure 2: Performance of geometric discriminators showing predicted vs true values of 𝑡ℎ



, 𝑙𝑒𝑛



, 𝜃



, & 𝑁𝑏𝑟



from left to right respectively. The worst performant discriminator is clearly the 𝑡ℎ



predictor.

with one difference; the skip-connections between the

outputs of encoder layers and the inputs of decoder

layers were eliminated here. The traditional

discriminator consists of seven down-sample

convolutional layers followed by a dropout and a final

fully connected layer. The geometric discriminators’

network consists of a stem, an Inception/Reduction

Resnet-v1- block-A, an Inception/Reduction Resnet-

v1-block-B, an Inception Resnet-v1-block-C

followed by an average pooling layer, a dropout layer,

and a fully connected layer1. We would like to point

out that the input of the three geometric

discriminators (𝑡ℎ



, 𝑙𝑒𝑛



and 𝜃



) consists of

the 2D design only, and the bar counter’s input

consists of the design alongside the mechanical

conditions.

2.2 Training Loss Function

The most challenging aspect of GANs is to find an

equilibrium between the generator and the

discriminator and avoid the dominance of one over

the other. The loss function with other training

parameters play an important role into stabilizing the

training and condemning the phenomenon of

oscillating losses. In this work, the loss function is

further challenging; it has to also account for the

geometrical (𝑡ℎ



, 𝑙𝑒𝑛



, 𝜃



and 𝑁𝑏𝑟



) and

mechanical (BC, F, V) constraints. Thus

𝐿









𝐿



𝜆



𝜆



𝐿







𝐿





𝐿





𝐿







(1)

Where the reconstruction loss 𝐿









∑

𝑥



𝑥











𝑥



and 𝑥



are the real and generated 2D design and N

is the batch size, 𝐿





∑|

𝑐



𝑐̂



, 𝑐∈







𝑁𝑏𝑟



,𝑡ℎ



,𝑙𝑒𝑛



,𝜃





𝑐



𝑎𝑛𝑑 𝑐̂



are the

input and predicted geometrical values respectively,

and 𝐿



is the Binary Cross Entropy (0 𝐿





100 in PyTorch). Hence, 𝐿



was set to 0.01,

so 𝐿



becomes of the same order of magnitude of

all other losses varying between 0 and 1.

2.3 Results

In this section, we present the performance of DL-

AM-TO with regard to geometric constraints. Thus,

we not only evaluate the aesthetics of the generated

designs (Fig. 2), but also test DL-AM-TO's ability to

respond to geometric changes (Fig. 3).

2.3.1 Training and Test Dataset

11719 samples of GMCAD (Almasri et al., 2021) are

used for training and 4405 samples for test. It consists

of 2D designs (in a .png format) alongside their

mechanical and geometrical constraints. GMCAD’s

features are detailed in (Almasri et al., 2021)

2.3.2 Geometric Discriminators’

Performance

To train the geometric discriminators, we augmented

the training dataset with three rotations of 90°,

180°and 270°. The predictive performance of the

geometrical discriminators is presented in Fig. 2. In

order to evaluate a predictor, an admissible error

interval is set (predictions within the green lines in

Fig.2 are considered correct). As we can clearly see,

the 𝑡ℎ



predictor shows the highest number of

inadmissible predictions (predictions outside the

green intervals). To quantify this observation, the

percentage of erroneous predictions for every

geometrical discriminator is computed.

We choose for the 𝑡ℎ



and 𝑙𝑒𝑛



lenmax the

relative prediction error defined as 𝑒







 100, and for 𝜃



and 𝑁𝑏𝑟



Δ

𝑇𝑟𝑢𝑒 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑

. The percentage of

predictions that fall within 𝑒





5% is 46%,

𝑒





5% is 1%, Δ𝜃



5

is 3,15% and

Δ𝑁𝑏𝑟



1 𝑏𝑎𝑟 is 0,15%. Consequently, we can

Advanced Deep Learning Techniques for Industry 4.0: Application to Mechanical Design and Structural Health Monitoring

537

conclude that all geometric discriminators are

sufficiently precise except for the 𝑡ℎ



one, which

needs further improvement. In fact, if we tolerate a

higher error interval of 10% for 𝑡ℎ



, we would end

up with 29,1%of inadmissible predictions

2.3.3 Dl-AM-TO’s Performance

Figure 3 shows a sample of real versus generated

designs alongside their skeletons and the geometrical

metrics: Δ𝜃



, Δ𝑁𝑏𝑟



, and 𝑒





. In fact,

designers are more interested in the design’s

geometry, which is best defined by the skeletons,

which explains their use here for comparison.

Figure 3: Comparison between the real and generated

designs in their full and skeleton formats on the test set.

As we can clearly see, DL-AM-TO captures the

geometrical information; Δ𝑁𝑏𝑟



0, Δ𝜃



rarely exceeds 5°, similarly, 𝑒





does not

exceed 10%. Aesthetically, the generated designs’

skeletons are similar to the real ones; we used the

structural similarity index (SSIM) to compare them

(Wang et al., 2004). SSIM is  0.7. The overall

geometrical performance was evaluated manually

over a sample of 100 designs of the test set; in other

terms, we counted the 𝑁𝑏𝑟



and measured the

𝑙𝑒𝑛



, and 𝜃



manually. We define a design

complying with (i) the 𝑁𝑏𝑟



constraint if

Δ𝑁𝑏𝑟



1 , (ii) the 𝑙𝑒𝑛



constraint if

𝑒





10% , and (iii) the 𝜃



constraint if

Δ𝜃



5%. We find that 83% of the designs respect

the 𝑁𝑏𝑟



constraint, 76% comply with the 𝑙𝑒𝑛



constraint, and 90% with the 𝜃



constraint.

In order to further investigate the geometrical

understanding of DL-AM-TO, we realized the

following experiment: We fixed the mechanical

constraints and altered one geometrical variable at a

time (𝑙𝑒𝑛



and 𝜃



). To calibrate the design’s

geometry, we simply need to modify the input value

of the de-sired geometrical condition. As we can see,

every time we increase 𝑙𝑒𝑛



𝜃



⁄

, the design’s

shape is modified in order to comply with this

variation while always conforming with mechanical

constraints (the F and BC). However, we can notice

that some geometrical constraints are correlated;

increasing the 𝜃



alters the 𝑙𝑒𝑛



, and at a certain

value, an additional bar appears.

To sum up, DL-AM-TO captures the geometrical

and mechanical constraints concurrently and

responds to geometrical changes creatively; the

obtained results encourage the further improvement

of the model

2.3.4 Discussion

The performance of DL-AM-TO relies on several

factors. In this study, the traditional SIMP method

was used to create the GMCAD dataset for training,

with the main objective being the integration of

mechanical and geometrical conditions through DL

architectures, rather than developing a new TO

algorithm. The performance of geometrical

discriminators is a key consideration. These

discriminators, trained within GAN frameworks, are

sensitive to oscillating losses, particularly with the

"𝑡ℎ



" variable, which, when integrated into the

model, resulted in performance deterioration. The

"𝑡ℎ



" discriminator struggled with precision due to

the transformation of CAD designs into image-like

representations using computer vision filtering

techniques, which can inadvertently change the

design's thicknesses.

DL-AM-TO represents a flexible and adaptable

approach, as it allows for the incorporation of

additional geometric constraints to enhance the

generated designs without necessitating significant

modifications to the model's architecture. Each input

geometric condition can be integrated as an additional

objective function of the discriminator. Thus,

changing multiple conditions at a time for multiple

times will generate a set of optimal Pareto front

solutions.

3 DL FOR DATA DRIVEN SHM

APPLICATION

Structural Health Monitoring (SHM) is dedicated to

the continuous and non-invasive monitoring of

structures integrity, ensuring safety and optimal

performances through on-site real-time

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

538

Figure 4: The work dataflow, from data acquisition and transformation into 2D images to MLP classification.

measurements (Worden et al., 2007). Typically, a

SHM approach for thin structures relies on the

utilization of transducers, such as piezoelectric

(PZT), bonded on the structure under examination

and that can either send and received ultrasonic waves

within the structure to be inspected. Thus, the data

collection process associated with the SHM of thin

structures using ultrasonic waves generate large

datasets composed of many time series associated

with all the PT paths available on the structure that

need to be processed using robust and efficient

techniques in order to determine the presence of the

damage and their properties (nature, position, size,

etc.). In this study, we investigate new ways of

structuring data that increase the accuracy of data

driven SHM algorithms and that are based on the

physical knowledge related with the structure to be

inspected. In particular, the generation of key images

from physics-based signals corresponding to all the

available PZT paths.

The overall dataflow of the developed SHM

algorithm is presented in Figure 4. First, the received

signals will be processed to compute the DIs for

various tested damage configuration. For each damage

case, reference images will then be generated from

these DIs, while on the other hand, a white Gaussian

noise will be added to the reference signals to compute

a new list of DIs. “Noisy” images will then be

generated from these noisy DIs and then compared

with the reference images in order to keep only images

that are relevant for supervised learning. The process is

repeated until there are enough images to form image

collections for each damage. The images are then

separated into training, validation, and test bases

depending on the methods, and fed into a MLP to

perform a classification task. The methods will finally

be compared according to their performance.

3.1 Lamb Waves Raw Data Acquisition

A complex composite aeronautic plate made of

composite carbon epoxy is considered (Fig. 5). It is

made with 4 plies oriented along [0°/ 45°/ 45°/ 0°]

and its dimension are 400x300 mm. A five

piezoelectric elements array is used, and each element

is bonded to the plate and acts sequentially as sensor

and actuator in order to emit and receive Lamb waves

signals. They are numbered from 1 to 5 and have a

12.5 mm radius. Three types of damage (impact,

delamination, magnet) are considered as shown in

Figure 4.

Figure 5: Epoxy plate under study, where “dam1” refers to

impact, “dam”2 refers to delamination and “dam3” refers to

magnet damage.

The first damage is made by an impact, the second

damage is made by placing magnet from both sides of

the plate, and the third damage is an artificial

delamination achieved by placing a small Teflon tape

between two composite plies. Lamb wave

measurements were employed for their ability to

interact with these damages (Mechbal et., 2017). The

signals were 5-cycle bursts with specific parameters,

and each signal was collected and emitted with

multiple repetitions for statistical robustness. Signal

pre-processing included denoising through a discrete

wavelet transform and time alignment to account for

hardware-induced time misalignments.

Following the collection of signals for all

receiver-actuator pairs, a set of 40 candidate Damage

Indexes (DIs) designed for damage detection is

computed by comparing the damaged state signals

with reference signals. The goal was to determine an

average score across three sub-scores that assess the

Advanced Deep Learning Techniques for Industry 4.0: Application to Mechanical Design and Structural Health Monitoring

539

coherence, range, and consistency of these DIs across

various scenarios. The study revealed that DIs based

on residual energy and signal correlation exhibit the

highest robustness, accommodating a wide range of

scenarios, environmental conditions, and potential

uncertainties, such as variations in the placement of

piezoelectric sensors. It's worth noting that for

reference-damage configurations, 100 DIs can be

generated (comparing 10 healthy signals to 10

damaged ones), whereas for ref-ref comparisons, only

45 DIs are computed due to the exclusion of

redundant comparisons within the available

repetitions.

3.2 Image Generation from DI’s

The developed technique is based on damage index

(DI) datasets computed for various damage

configurations. It exploits information about the

location of piezoelectric transducers (PZTs) to

structure a 2D image to match the acquired Lamb

wave signals. This approach, called sensor placement

(SP), uses the physical positioning of the sensors on

the plate, independently of the location of the

damage. The image generation process is as follows:

in the columns, from left to right, we position the

pixels corresponding to the PZT actuator located

furthest to the left and right of the plate. The rows of

pixels are in the same order as the columns, but this

concerns the positions of the receivers. For example,

a pixel at coordinates (3,4) in the image matrix

characterises the damage index (DI) between the third

leftmost PZT, acting as an actuator on the plate, and

the fourth leftmost PZT, acting as a sensor. In

particular, the actuator positioned in the column is not

taken into account, as actuator-receiver pairs

involving the same PZT are excluded from the

analysis. The available 40 DI’s would only provide us

with 40 images per damage, which is clearly

insufficient for performing supervised deep neural

network learning tasks. Therefore, we need to

generate more images from the original ones. We

chose to use additive white Gaussian noise (AWGN)

but in a manner different from what is commonly

found in the literature. Indeed, data augmentation

techniques often aim to manipulate image support to

generate new images. These techniques include

flipping, rotation, resizing, cropping, or even adding

noise directly to the source image. In our case, it is

the temporal signals that are being modified. We add

Gaussian noise directly to a signal, stored in our case

in matrix. AWGN is characterized by its signal-to-

noise ratio (SNR) whose expression is:

𝑆𝑁𝑅



10𝑙𝑜𝑔





𝐴



𝐴







20𝑙𝑜𝑔





𝐴



𝐴





(2)

Where 𝐴



and 𝐴



represent the amplitudes of

the reference signal and the additive noise,

respectively.

The AWGN models random noise distributed

according to a normal distribution and simulates the

background noise, caused for example by

interference, of a physical channel, in our case the

plate, which serves as an information transmission

gateway between a transmitter and a receiver. It is

therefore perfectly suited for this kind of situation.

Signals for the 3 damages that were initially recorded

were used to generate 40 reference images that we

will call vanilla. At the same time, these signals were

then used as a basis for applying AWGN that differ in

SNR. A limit SNR above which signal noise does not

create images that are sufficiently different from the

references was found empirically. Signals are noised

with SNRs below the SNR limit and then new DIs are

generated and batches of 40 (SP method) noisy

images are created per SNR value.Images generated

for the SP method with the noisy signals are shown in

Figure 6. It can be observed that according to the

damage, the reference image and the artificially

created AWGN image have the same structure and

differ only in the light intensity of some pixel blocks.

Figure 6: SP vanilla (1) vs. noisy (2) images with AWGN

for SNR=50.5 dB and for the DI Welch-based Power

Spectral Density (WPSD). Damage is by delamination (a),

impact (b) and magnet (c).

The main information in the artificial image, i.e., the

white pixel that informs about the largest DI, is

(a)

(b)

(c)

(1) (2)

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

540

located at the same place and is of the same intensity

as in the corresponding vanilla image. After several

batches, we finally managed to generate 418 images

per class, i.e. 1254 images.

3.3 Classifier Architecture

Deep Learning is increasingly being adopted in the

field of Structural Health Monitoring (SHM) (Tabian

et al., 2019). In our case, we propose to train a MLP

neural network using the SP image dataset that we

presented in the previous section. The aim is to use

this network to perform a classification of the three

types of damage. We have therefore flattened the SP

images into a single vector which is used as the input

layer of a fully connected hidden MLP. We have

proposed a model where the hyper parameters have

been finely tuned to obtain the best classification rate.

For that, we employed a grid search approach and

conducted a series of experiments to systematically

determine the optimal hyper parameters of our MLP

model. The SOFTMAX activation function at the

final layer represents the categorical distribution,

assigning scores to each output class and

transforming them into probabilities.

3.4 Classification Results

The evaluation of the MLP model's performance will

involve analyzing the loss function across epochs.

The loss function is a crucial indicator of the model's

behavior after each optimization iteration, reflecting

its predictive accuracy. Additionally, we will

consider the accuracy metric, which measures the

model's predictive precision in comparison to the

ground truth data. Table 1 summarizes the accuracy

and loss results for the proposed MLP model, along

with the learning settings applied across multiple

training sessions.

Table 1: MLP model classification results.

Total number of ima

1254

umber of trainin

data

1003

umber of test ima

125

Epochs

Learnin

rate

0.001

Avera

e loss

0.372

Avera

e accurac

0.928

15 training sessions were conducted for each damage

to raise the average information depicted in table 1.

When we assess our model performance in

classifying the three types of defects, we can see that

the results are quite impressive, despite the dataset's

size limitations. After training for 30 epochs, the

accuracy approaches nearly 1, which means almost

perfect classification.

Figure 7: Performance of the MLP model after 500 epochs.

However, we observe a relatively high loss with

respect of that high accuracy. This can be understood

by examining the loss function mathematically. When

we plot a histogram of the loss values for individual

samples, it's evident that most samples have very low

loss, but a few outliers have significantly higher

losses. The MLP results are shown in figure 7. The

proposed model performs very well and does not

overfit. In fact, the SP images show to fit well for this

kind of task that is not too complex. The network

manages to class efficiently the three types of

damage, with few examples misclassified for the

magnet damage as shown in the confusion matrix

(figure 8).

Figure 8: The confusion matrix.

Advanced Deep Learning Techniques for Industry 4.0: Application to Mechanical Design and Structural Health Monitoring

541

3.5 Discussion

SP images appear to be perfectly suited to the task of

classifying defects using an MLP model.

Nevertheless, our primary challenge has been the

limited availability of data, primarily due to high

computational costs. Addressing the issue of data

scarcity is a common concern in deep learning. Data

augmentation techniques are often employed to

artificially expand datasets. Transfer learning, an

approach that leverages knowledge gained in solving

one problem to address another, is also of interest.

While our results suggest that the SP images performs

well with limited data, there is potential in exploring

how incorporating knowledge from complex neural

networks could benefit both methods.

4 CONCLUSIONS

In this paper, we presented two use cases for artificial

intelligence in Industry 4.0. The first is an additive

manufacturing-driven topological optimisation

approach based on deep learning, called DLAM- TO.

This technique integrates mechanical and geometric

constraints at the same level and generates 2D

designs. More interestingly, it easily adapts the

geometry of the design to propose several additive

manufacturing compliant geometries corresponding

to the needs of the design engineer while maintaining

a mechanical performance similar to that proposed by

SIMP. The second example demonstrated the

feasibility of a data-driven approach based on images

generated by physics (lamb wave propagation) to

classify three types of damage in the context of SHM

application. The next step involves finalizing the

demonstrators and planning their evaluation by end

users, namely, designers and maintenance operators.

This phase is crucial to refine these approaches and

transform them into fully operational tools intended

for use in the context of the factories of the future. It

also requires an optimization and adaptation process

to ensure their alignment with the specific needs of

these professionals before proceeding with the actual

deployment in industrial environments.

REFERENCES

Almasri, W., Bettebghor, D., Ababsa, F., Danglade, F. and

Adjed, F., 2021, July. Deep Learning Architecture for

Topological Optimized Mechanical Design Generation

with Complex Shape Criterion. In International Con-

ference on Industrial, Engineering and Other

Applications of Applied In-telligent Systems, pp. 222-

234, (2021)

Almasri W., Bettebghor D., Adjed F., Ababsa F., Danglade

F., GMCAD: an original Synthetic Dataset of 2D

Designs along their Geometrical and Mechanical

Conditions. In International Conference on Industry 4.0

and Smart Manufacturing, (2021).

Arden, N. S., Fisher, A. C., Tyner, K., Lawrence, X. Y.,

Lee, S. L., & Kopcha, M.. Industry 4.0 for

Pharmaceutical Manufacturing: Preparing for the Smart

Factories of the Future. International Journal of

Pharmaceutics, (2021)

Bendsoe, MP., Kikuchi N., Generating optimal topologies

in structural design using a homogenization method,

Computer Methods in Applied Mechanics and

Engineering, vol. 71, no. 2, pp. 197-224, (1988)

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Bengio, Y.. Generative

adversarial nets. In: Advances in neural information

processing systems, pp. 2672-2680, (2014)

Khelalef, A., Ababsa, F., Benoudjit, N., An Efficient

Human Activity Recognition Technique Based on Deep

Learning. Pattern Recognition and Image Analysis 29,

702-715 (2019)

Lockner, Y., Hopmann, C. Induced network-based transfer

learning in injection molding for process modelling and

optimization with artificial neural networks. Int J Adv

Manuf Technol 112, 3501–3513 (2021)

Mechbal, N., Rébillat, M., Damage indexes comparison for

the structural health monitoring of a stiffened

composite plate. 8th ECCOMAS Thematic Conference

on Smart Structures and Materials (SMART 2017) pp.

436-444, (2017)

Mohan, T. R., Roselyn, J. P., Uthra, R. A., Devaraj, D., &

Umachandran, K.. Intelligent machine learning based

total productive maintenance approach for achieving

zero downtime in industrial machinery. Computers &

Industrial Engineering, (2021)

Nath, A.G., Udmale, S.S. & Singh, S.K. Role of artificial

intelligence in rotor fault diagnosis: a comprehensive

review. Artif Intell Rev 54, 2609–2668 (2021).

Oehlmann, P., Osswald, P., Blanco, J.C. et al. Modeling

Fused Filament Fabrication using Artificial Neural

Networks. Prod. Eng. Res. Devel. 15, 467–478 (2021).

Tabian, I.; Fu, H.; Sharif Khodaei, Z. A Convolutional

Neural Network for Impact Detection and

Characterization of Complex Composite Structures.

Sensors, 19, 4933, (2019)

Wang, Z., Bovik, A.C., Sheikh, H.R. and Simoncelli, E.P.,

Image quality assessment: from error visibility to

structural similarity. IEEE trans-actions on image

processing, 13(4), pp.600-612, (2004)

Worden, K.; Farrar, C. R.; Manson, G. & Park, G. "The

Fundamental Axioms of Structural Health Monitoring"

Proceedings: Mathematical, Physical and Engineering

Sciences, The Royal Society, pp. 1639-1664, (2007)

Zhang, Z., Liu, Q.,Wang, Y., Road extraction by deep

residual u-net. IEEE Geoscience and Remote Sensing

Letters, 15(5), 749-753 (2018).

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

542