Transfer Learning for Optimization of Carbon
Fiber Reinforced Polymer Production
Institute for Software & Systems Engineering
Abstract. The main problem that keeps many areas of research from
using Deep Learning methods is the lack of suﬃcient amounts of data.
We propose transfer learning from simulated data as a solution to that
issue. In this work, we present the industrial use case for which we plan
to apply our transfer learning approach to: the production of economic
Carbon Fiber Reinforced Polymer components. It is currently common
practice to draw samples of produced components statistically and per-
form destructive tests on them, which is very costly. Our goal is to predict
the quality of components during the production process. This has the
advantage of enabling not only on-line monitoring but also adaptively
optimizing the manufacturing procedure. The data comes from sensors
embedded in a tooling in a Resin Transfer Molding press.
Keywords: Transfer Learning, Simulation-based Machine Learning, Or-
ganic Computing, Carbon Fiber Reinforced Polymer
Machine Learning (ML) and especially Deep Learning (DL) algorithms have been
successfully applied to certain ﬁelds of research recently: Computer Vision  or
Speech Recognition . Moreover, Reinforcement Learning (RL) in combination
with DL was used to solve games (e.g., Go  and Atari games ). Other
areas, for instance, industrial production or medical sciences, try to beneﬁt from
these algorithm families as well. The problem that keeps those and many other
areas from adapting DL to the respective problems is the immense need for
data when using deep neural networks. The gathering of enough data and the
subsequent labeling phase is costly or can be almost impossible for certain kinds
The use case we present is the automation, optimization, and adaptation of
industrial production of Carbon Fiber Reinforced Polymer (CFRP) that should
be improved by transfer learning from simulated data. Components made of
Carbon ﬁber make vehicles lighter by 50-70 % compared to conventional steel
and therefore more energy-eﬃcient: A 1 % weight reduction reduces the en-
ergy consumption by approximately 0.6-0.7 % . Currently, the production
of CFRP is too expensive for automotive volume scale, because it is by far
2 Simon Stieber
more complicated to produce CFRP than press steel or aluminum . The
factors that drive the cost the most are diﬃcult automation, high cycle times,
personnel, industrial manufacturing equipment, and material cost. In low-scale
production industries such as avionics and aerospace, even more cost-ineﬃcient
factors such as increased manual labor and intermediate products, called “car-
bon pre-pregs”1increase costs even further. Our solution addresses several of
these problems: we use an automated Resin Transfer Molding (RTM) process
(see Fig. 1) with more cost-eﬃcient materials: non-wovens made of recycled
carbon ﬁbers and a small share of natural ﬁbers and caprolactam as resin.
Fig. 1. Schematic RTM press with sym-
bolic sensors and actuators.
This combination of an RTM pro-
cess with non-wovens and online in-
situ monitoring during the produc-
tion process is unique to our knowl-
edge and makes the whole endeavor
very complex. The manufacturing of
components for mere training purposes
is non-economical, since they are dis-
posed of thereafter. The press and
its components are shown in Fig. 1.
A and B are containers for the two-
component resin. The mixing unit
is marked with C. D illustrates the
press. The carbon mat E is about
to be injected with resin. F repre-
sents the tooling that holds the neg-
ative of the future component and is
equipped with acoustic, electric, tem-
perature and pressure sensors.
The process steps are:
1. The non-woven carbon mat is
brought into the press, which has
around 100 degrees Celsius.
2. Resin is injected into the mold.
3. The Press is pressed and resin
spreads in the mat.
4. Sensors measure the spread distri-
bution and the hardening of the
4.1. If necessary, the press keeps
shut for a longer period or ad-
ditional pressure is applied.
5. The press is opened and the ﬁn-
ished component is extracted.
1Pre-preg stands for “pre-impregnated” composite ﬁbers. Within these, a polymer
matrix material, e.g., epoxy, is already present in the ﬁber.
Transfer Learning for Optimization of CFRP 3
In this position paper, we present the initial goals of the project and suggested
paths towards them. We propose learning on simulated data (as in robotics 
or gaming [15,21]) and reﬁning that knowledge with few real-world samples as a
solution. While the dataset of real-world samples within our project is too small
for DL by a magnitude, as many simulated runs as necessary can be produced.
Additionally, the time-consuming data labeling step is no longer necessary since
the data is already labeled when it comes from the simulation: all stages of resin
distribution and curing are available for every spot of the component.
Researching transfer learning from simulated to real data is a major focus of
our work. We want to identify and tackle the Machine Learning (ML) challenges
at the gap between simulation and the real world.
The smart CFRP press that is constructed during my dissertation is self-
learning and self-adaptive. These self-X capabilities are a key concept of organic
computing systems . The restore invariant approach , another key notion
of organic computing that describes the corridor of correct behavior can be
applied to this machine. Initially, the press has to be parametrized to get into this
corridor of correct behavior: it produces CFRP components with a high quality.
Whenever the quality is declining—moving out of the corridor—the process is
adapted by the machine itself.
In Section 2 we present the expected challenges and the goals of the project.
Section 3 entails recent works on transfer learning on the one hand and works on
non-destructive testing of CFRP components via acoustic and electric sensors on
the other hand. Methods, algorithms, and ideas on how to tackle the challenges
of this project are shown in Section 4. The last part, Section 5 gives an overview
of future research activities. Additionally, our work is put into context with
respect to our project partners’ contributions.
Bridging the gap between simulated ﬂow fronts2and real sensor data will be
a major challenge. Gr¨ossing et al.  compared the simulation software we are
going to use, PAM RTM3, a subset of PAM Composites by ESI, and an open
source alternative with regards to how realistic they model ﬂow fronts. During
the work on the project, we will test how suitable PAM RTM is for our setup.
One key question is how to merge the two types of data. Are we going to be
able to learn on raw sensor data, time series of ultrasonic and electric sensors,
or do we use abstracted ﬂow fronts from real data for training? Both types of
training data have to be aligned to make it possible to coherently train on them.
2The ﬂow front is the edge between the part of the mat that has been impregnated
with resin and the rest that has to be impregnated yet. The ﬂow front simulation
software is able to simulate the distribution process of the resin.
virtual-manufacturing/composites/pam-composites/pam-rtm, Accessed on
August 29, 2018
4 Simon Stieber
The research on transfer learning from simulated and real data is an important
aspect of my dissertation.
Partners in the project have been working on similar questions in the past.
Kalafat and Sause  sensed damages in carbon components with acoustic sen-
sors. In our project, similar sensors, as well as piezo-electric sensors, will be used
to sense the distribution of the resin in the carbon ﬁber mat and the polymer-
Another challenge is to handle varying environmental factors such as air
temperature and humidity. These can interfere with the production progress:
especially the caprolactam is sensitive to humidity. In a controlled production
environment, such environmental factors should become tamable.
Carbon long ﬁbers, the raw material, have a variance in weight per meter of
10 %4. That makes the material’s behavior less predictable than that of steel
or aluminum. Variations in the production material will be a central issue. In
addition to the variance in weight of the raw material, a mix of recycled carbon
and natural ﬁbers will be used to produce the non-woven mats for our project.
Diﬀerent distributions of these materials in the non-wovens will cause diﬀerent
permeabilities. Permeability is a very important property when predicting the
distribution of the resin in the mat. We want to tackle this challenge by varying
the input parameters of the simulation to cover problematic non-woven mats.
The goals for our part in this project, that build on top of each other, are:
1. Classifying the quality5of one component from mere sensor input
2. Optimizing the quality over the production process of multiple components
3. Get a real-time prediction of one component as it is still in the press and
adjust certain parameters if the outcome is expected to be sub-optimal, see
Fig. 1, step 4.1.
Furthermore, the optimization of the algorithm is part of our work, as soon as
the goals regarding functionality are met. The main priority is to ensure that
the inference step meets the requirements for online adaption of parameters at
production speed. The hard time limit needs to be determined ﬁrst, once the
timing of the process is clear. As soon as the timing of the process is deﬁned, we
can set the speciﬁcations for our algorithm and start optimizing.
3 State of the art - Related work
There have been preliminary works on optimizing the production of CFRP prod-
ucts. Sorg  compared several data mining algorithms on data coming from
4Source: Voith Composites, http://voith.com/composites-de/index.html, Ac-
cessed on November 15, 2018
5The quality of a CFRP component is determined by several factors. The resin has
to be distributed equally and has to reach all parts of the ﬁber mat. Additionally,
no air or other materials shall be enclosed in the component. Then the stability is
Transfer Learning for Optimization of CFRP 5
the production of a carbon car roof. The sources of the data were already es-
tablished process logs which held data about several time spans, thicknesses of
the product at certain points, temperatures, and pressures. Within our project
these sorts of data will be available as well as data that is speciﬁcally gathered
for our needs: sensor data that shows the ﬂow front and curing progress.
These acoustic and electric sensors have already been used to detect dam-
ages in CFRP components in-situ on one hand. Kostka et al.  and Larrosa
et al.  both have been working on monitoring the health status of an already
ﬁnished CFRP component with sensors (acoustic and electric), that were em-
bedded in the component during production. Kostka et al. used decision trees
and Larrosa et al. implemented Gaussian discriminative analysis to reach their
goals. Kalafat and Sause have shown that the localization of defects in CFRP
via neural networks is feasible and outperforms other techniques  .
Opposed to these authors that used sensor data to determine the health sta-
tus of a CFRP component, we will use acoustic and electric sensors to determine
the ﬂow front and the curing process within in the press. Their works are impor-
tant with regard to using sensors to determine the state of a component already
A research company named Synthesites, on the other hand, has published
several works on online monitoring of the production of CFRP. Pantelelis et
al.  have shown that it is possible to monitor not only the distribution of the
resin but also the curing of the product with electrical and temperature sensors.
In another work which was carried out in a EU project6, they have shown that
they can use neural networks for the optimization of the curing process . In
their work, the heating proﬁle for a CFRP component was optimized so that
production could be accelerated by 36 % with 5 % less heating energy involved
compared to no curing monitoring and control. They used very shallow neural
networks with only one hidden layer and used bootstrap aggregating to combine
the output of 30 networks. Besides the data from online monitoring and the
neural networks involved, there is a third similarity of their work to our project:
they used simulated data for initial training.
We delimit our work from theirs by several points. First, the online moni-
toring in our project concerns not only the distribution of resin and curing of
the component at several points of interest, but the exact ﬂow front process,
damage detection during deforming, and monitoring of temperature, pressure,
and acoustics in the tooling. We operate on a much larger scale regarding data
compared to the other project: a sensor network made of many more sensors will
be used in our project. They used only ﬁve single sensors. We will use more so-
phisticated ﬂow front simulations that are specially made for CFRP production
reﬁned with data coming from the aforementioned sensor network. Fr the other
works, a simpler, one-dimensional simulation method by Pantelelis et al.  was
used. Another diﬀerence lays within the utilized materials: we use caprolactam as
resin and non-woven carbon mats whilst they were using epoxy and wovens. As
6IREMO (intelligent REactive polymer composites MOulding) https://cordis.
europa.eu/project/rcn/94018_en.html, Accessed on August 22, 2018
6 Simon Stieber
mentioned in Section 2, Gr¨ossing et al. compared PAM RTM, and OpenFOAM7
regarding their ability to model ﬂow fronts. To test how realistic the simulations
are, they installed a transparent upper half on an RTM press to make the ﬂow
front advancement visible. In contrast to our work, they used a resin with high
viscosity and carbon pre-pregs. Their experiments showed that PAM RTM is
computationally more eﬃcient because of the diﬀerent underlying mathematical
model. It is easier to operate, but has the disadvantage that it cannot model the
transition between porous and non-porous/solid material. Once we start using
PAM RTM, we will see if this issue is relevant to our work and, if so, we need
to check if that issue is solved in a recent version of the software since the paper
appeared in 2016.
After covering the related work concerning CFRP, the following part will
present research on learning methods that have been utilized for other use-cases.
We expect for at least some of these approaches to work well in the context of
Weiss et al.  give a comprehensive and in-depth overview of the methods
of transfer learning. They present diﬀerent domains where algorithms of this sort
have been implemented successfully, i.e. for image classiﬁcation, human-activity
classiﬁcation, and software defect classiﬁcation. They do not explicitly treat
simulation-to-real data transfer learning but give a good overview of common
techniques to adapt the knowledge of an algorithm from one domain to the other.
The main method to do so is to ﬁne-tune a neural network, a method that
is described more thoroughly in Section 4. Progressive Neural Networks  are
another interesting idea to handle the transfer of knowledge from one domain
to another. They have been used for transferring knowledge from simulation to
real data in a robot use case , which makes them well suited for our problem.
Section 4 gives an introduction to the operating principle behind this type of
Other sources focus on diﬀerent applications of DL algorithms to speciﬁc
use cases with the help of simulations. These include traﬃc ﬂow prediction 
and the understanding of dynamical scenes from video data . Both works
have shown that it is possible to learn from simulated data and to adapt that
knowledge to real-world applications. They lack the link to CFRP or industrial
production in general, a gap we plan to ﬁll.
Initially, we will train our model to predict the quality (our deﬁnition of quality
is described in Footnote 5) of a small and simple carbon component. In a second
step, a chicane component that exhibits a complex geometry, for instance ascents
or round edges, will be produced. These three steps are necessary due to the fact
that the production equipment will be installed in parallel to the ﬁrst steps in the
project and the ﬁnal system will be available late in the project. This stepwise
7https://www.openfoam.com/, Accessed on August 29, 2018
Transfer Learning for Optimization of CFRP 7
approach makes it inevitable to devise an algorithm that is capable oﬀ keeping
some of its knowledge from previous steps and adapt it for the next level. That
makes it unnecessary to learn from scratch for each new component and lowers
development and production costs.
The idea of transfer learning was already described in Section 3: adapting a
pre-trained algorithm to a new domain to save time and training eﬀort, especially
when there is not enough data available in the new domain. The common and
straightforward approach to do so for neural networks is called ﬁne-tuning. The
following four steps sketch the approach in a simpliﬁed manner.
1. Load network topology and initialize it with pre-trained weights
2. Cut oﬀ last layer - the output layer
3. Add new output layer that ﬁts the new problem
4. Train the last network again and keep all old layers in their original state,
do not update their weights
This approach, originally described by Hinton et al.  has the advantage that
the training converges and it takes much less time to train compared to training
from scratch. In our case, we use simulated data for pre-training and the real data
of one type of component for ﬁne-tuning. For the more sophisticated parts, the
network can be adapted again. An approach that could be an even better ﬁt to
Fig. 2. Progressive neural network with three columns. From 
our multi-component/task problem are progressive neural networks  because
their topology is especially made for transfer learning from one to multiple tasks.
The principle works as follows and is shown in Fig. 2:
8 Simon Stieber
1. Train a neural network with an arbitrary topology (CNN, RNN, ...); this is
the ﬁrst column.
2. Add a second network/column that is trained for another task alongside the
3. Add adapter connections from the layers of the ﬁrst to the second column
(denoted (a) in Fig. 2).
4. Column one is disabled for training.
5. Column two uses features from column one to solve the new task.
This approach has some great advantages: it circumvents catastrophic forget-
ting , an unwanted behavior that happens to ﬁne-tuned networks very often:
once a new task is learned, the algorithm performs poorly in the original task.
Another interesting feature for our purpose is that new tasks can be added
easily: one new column per task. Helpfully, new tasks do not have to contain as
many parameters as the previous ones and thus need less data. For our project
a good ﬁrst start is to train the ﬁrst column for simulated data for the simple
component. The second column could be the real data for the simple compo-
nent. This learning from previous stages continues when switching to the chicane
component: it is very desirable to keep the training eﬀort low when transferring
The disadvantage of these networks is the massive amount of parameters
they utilize. For every new task, a new column is added, which lets the number
of weights grow enormously when training many tasks. In our project, we do not
have to meet hard storage constraints but keeping an eye on the dimensions of
a network is always reasonable.
In the explanation above, one detail was skipped: the adapter connections
between the layers of diﬀerent columns. These connections are single layer Multi
Layer Perceptrons (MLPs), which means they have weights and activation func-
tions and thus can be disabled. As the number of columns grows, the number
of adapters and their related parameters grows exponentially. Since the number
of parameters in a progressive neural network is high anyway, disabling certain
connections between layers helps to keep the size of the network in check.
For the diﬀerent components, we obtain simulated data from ﬂow front sim-
ulations. As described in Section 3, the preferred simulation software is PAM
RTM by ESI.
Flow front simulations have certain similarities to videos: the single images
are 2D if the component is not too complicated and they have a time series
nature. Colors indicate diﬀerent stages of wetting and/or polymerization. One
obvious path to handle this type of data is to treat it similar to video data
which opens the possibility to adapt approaches for video analysis that have
proven successful: previous studies include e.g., video classiﬁcation as shown
by Karpathy et al.  and Ng et al. . Another idea based on the video
path is to use image segmentation methods  to pre-train a network in a
ﬁrst step to let it gain knowledge on segmenting ﬂow fronts and the optimal
distribution of resin and curing. In a second step, that network is retrained for
regression, i.e. inferring one value for the quality of the component. Convolutional
Transfer Learning for Optimization of CFRP 9
Neural Networks (CNNs) are neural networks that work very well for image
segmentation and overall analysis of single images. If we follow the image path,
a mixture of CNNs is the network topology we choose, no matter if we ﬁne-
tune a network or use several networks together in a progressive neural network.
For CNNs, the hyperparameters to tune are: the overall topology, which can
be a residual network, or made of inception modules or simpler topologies or
another architecture. The next parameter is the number of layers and neurons.
For training, the learning rate and the optimizer are interesting parameters to
tune. We will start with proposed values and start optimizing training as soon
as progress is visible, that means the training is working.
The idea of using raw sensor data input would make it necessary to select
of a diﬀerent kind of topology that is well suited for that data because it is
uncertain if raw data has the same properties as images: repeating patterns at
certain areas, areas of higher interest than others and likewise. Since we do not
know yet if we can use the raw data for learning, this path of raw data is not
described further here.
Despite the fact that deep neural networks have proven to work for many
diﬀerent kinds of problems, we will consider other types of algorithms for our
approach. Sorg  showed that a method for constructing decision trees, which
was invented a long time ago, CHAID , was more suitable for his optimization
problem of the process mining the production of a carbon car roof. When it
comes to the acceptance of an algorithm within production and management, a
very important aspect of an algorithm is its interpretability. Decision trees are
easily understandable whilst neural networks are black boxes with next to no in-
terpretability. If the performance ﬁgures of these algorithms are similar and even
if the decision tree is a little weaker than the neural network, chances are that
the overall decision rules in favor of the understandable decision trees. Decision
trees carry another interesting feature: they do not need as much training data
as deep neural networks. Opting for them, or any other data frugal algorithm
would have the advantage of setting aside using simulated data, as well as trans-
fer learning and just use the real data collected by the sensors. That would be a
much simpler approach. For our goals (the detailed construction of a ﬂow front,
curing, etc.), these algorithms are likely not capable enough, but we will give
them a try as well. One other simple approach are shallow neural networks as
utilized by Pantelesis et al. . But even for these networks carrying only one
hidden layer with a handful of neurons, simulated data was used for training.
This reduces the hope that only real data will help to successfully train any kind
of learning algorithm.
The real data is coming from sensors within the press: ultrasonic, electric,
pressure and temperature sensors, which are also shown in Fig. 1. Deriving the
ﬂow front and the curing and all other target values from this data is another
challenge, which will be tackled with the help of experts in material science.
After this ﬁrst step of predicting quality of a ﬁnished component from sensor
input, optimizing the quality of components over the course of time is the next
level of complexity. With the acquired quality inference method, a RL approach
10 Simon Stieber
to increase the quality of the component will be developed and applied. The
parameters to change, e.g., the actions for RL, are the pressure of resin, the
pressure of the press and the temperature of the press. Ultimately, we will opti-
mize our algorithm to recognize possible defects in the component in the press
and to consequently take countermeasures such as increasing the pressure of the
press or leaving the component in the press for a longer time.
To obtain a baseline to compare to, approximately 100 units of the simple
sample components will be produced. Their quality will be tested and they will
therefore be destroyed. This quality assessment of 100 simple components will
work as our labels and we will split this set into training, test, and validation
data sets. We will then compare the outcome our algorithm produces with this
real data. For the more complex component, not as many units will be produced.
Nonetheless, the real data will also be used for training and testing, but at a
We suggested transfer learning from simulated data as a tool to make DL usable
for domains with small datasets. The optimization of the production of CFRP
is the use case our project is dedicated to. We showed the expected challenges
and the proposed methodology for that project.
After a training for PAM Composites, enough data for initial neural network
trainings have to be produced and the ﬁrst trainings have to be carried out.
These two steps, gathering data and training, including tweaking of parameters
have to be repeated until the results satisfy our demands.
When the ﬁrst batch of sample components and related quality data is com-
pleted, we can start our transfer learning eﬀort. Depending on the success of
these attempts on transferring knowledge, the following steps include, ﬁrst, to
incrementally optimize the quality of components with RL over time, and second,
to interfere in the production process, if necessary. At last, adapting knowledge
from the simple to the complicated component has to be accomplished.
Our part in this project is comprised of all data science and ML tasks. At
the moment, especially the work on transfer learning and the combination of
various learning methods are scientiﬁc challenges. Others will follow during the
lifetime of the project. The research of these problems will be a major part of
1. French: Catastrophic forgetting in connectionist networks. Trends in cognitive sci-
ences 3(4), 128–135 (apr 1999)
2. Graves, A., Mohamed, A.R., Hinton, G.: Speech recognition with deep re-
current neural networks. 2013 IEEE International Conference on Acous-
tics, Speech and Signal Processing (ICASSP) (6), 6645–6649 (2013).
Transfer Learning for Optimization of CFRP 11
3. Gr¨ossing, H., Stadlmajer, N., Fauster, E., Fleischmann, M., Schledjewski, R.: Flow
front advancement during composite processing: predictions from numerical ﬁlling
simulation tools in comparison with real-world experiments. Polymer Composites
37(9), 2782–2793 (sep 2016). https://doi.org/10.1002/pc.23474
4. G¨udemann, M., Nafz, F., Ortmeier, F., Seebach, H., Reif, W.: A speciﬁcation and
construction paradigm for organic computing systems. In: Proceedings - 2nd IEEE
International Conference on Self-Adaptive and Self-Organizing Systems, SASO
2008. pp. 233–242 (2008). https://doi.org/10.1109/SASO.2008.66
5. Hinton, G.E., Salakhutdinov, R.R.: Reducing the dimensionality
of data with neural networks. Science 313(5786), 504–507 (2006).
6. Ingrid Bonde Akerlind, Parisa Bastani, Irene Berry, Kandarp Bhatt, Alice Chao,
E.C., Valerie Karplus, David Keith, Michael Khusid, Eriko Nishimura, S.Z.: On
the Road toward 2050: Potential for Substantial Reductions in Light-Duty Vehi-
cle Energy Use and Greenhouse Gas Emissions John. Massachusetts Institute of
7. Kalafat, S., Sause, M.G.: Localization of Acoustic Emission Sources in Fiber Com-
posites Using Artiﬁcial Neural Networks. 31st Conference of the European Working
Group on Acoustic Emission pp. 1–8 (2014)
8. Karpathy, A., Toderici, G., Shetty, S., Leung, T., Sukthankar, R., Fei-Fei, L.: Large-
Scale Video Classiﬁcation with Convolutional Neural Networks. 2014 IEEE Confer-
ence on Computer Vision and Pattern Recognition (CVPR) pp. 1725–1732 (2014).
9. Kass, G.V.: An Exploratory Technique for Investigating Large Quan-
tities of Categorical Data. Applied Statistics 29(2), 119 (1980).
10. Kostka, P., Holeczek, K., Filippatos, A., Langkamp, A., Hufenbach, W.: In situ
integrity assessment of a smart structure based on the local material damping.
Journal of Intelligent Material Systems and Structures 24(3), 299–309 (2013).
11. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet Classiﬁcation with Deep
Convolutional Neural Networks. Advances In Neural Information Processing Sys-
tems pp. 1–9 (2012)
12. Larrosa, C., Lonkar, K., Chang, F.K.: In situ damage classiﬁcation for compos-
ite laminates using Gaussian discriminant analysis. Structural Health Monitoring
13(2), 190–204 (2014). https://doi.org/10.1177/1475921713517288
13. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic
segmentation. In: Proceedings of the IEEE Computer Society Conference on Com-
puter Vision and Pattern Recognition. vol. 07-12-June, pp. 3431–3440. IEEE (jun
14. Lv, Y., Duan, Y., Kang, W., Li, Z., Wang, F.Y.: Traﬃc Flow Prediction With Big
Data: A Deep Learning Approach. IEEE Transactions on Intelligent Transporta-
tion Systems 16(2), 865–873 (2014). https://doi.org/10.1109/TITS.2014.2345663
15. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie, C.,
Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Hassabis,
D.: Human-level control through deep reinforcement learning. Nature 518(7540),
529–533 (feb 2015). https://doi.org/10.1038/nature14236
16. Nafz, F., Seebach, H., Stegh¨ofer, J.p., Anders, G., Reif, W.: Constraining Self-
organisation Through Corridors of Correct Behaviour: The Restore Invariant Ap-
12 Simon Stieber
proach. In: Organic Computing — A Paradigm Shift for Complex Systems, pp.
79–93. Springer Basel, Basel (2011). https://doi.org/10.1007/978-3-0348-0130-0 5
17. Pantelelis, N., Vrouvakis, T., Spentzas, K.: Cure cycle design for compos-
ite materials using computer simulation and optimisation tools (April) (2003).
18. Pantelesis, N., Bistekos, E.: Process monitoring and control for the production of
CFRP components. SAMPE Conference pp. 5–9 (2012)
19. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J.,
Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive Neural Networks. arXiv
20. Rusu, A.A., Vecerik, M., Roth¨orl, T., Heess, N., Pascanu, R., Hadsell, R.: Sim-
to-Real Robot Learning from Pixels with Progressive Nets. In: 1st Conference on
Robot Learning (CoRL 2017). pp. 1–9 (2016)
21. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez,
A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui,
F., Sifre, L., van den Driessche, G., Graepel, T., Hassabis, D.: Mastering the
game of Go without human knowledge. Nature 550(7676), 354–359 (oct 2017).
22. Sorg, C.: Data Mining als Methode zur Industrialisierung und Qualiﬁzierung neuer
Fertigungsprozesse f¨ur CFK-Bauteile in automobiler Großserienproduktion. Disser-
tation, Technische Universit¨at M¨unchen (2014)
23. Weiss, K., Khoshgoftaar, T.M., Wang, D.D.: A survey of transfer learning,
vol. 3. Springer International Publishing (2016). https://doi.org/10.1186/s40537-
24. Wu, J., Yildirim, I., Lim, J., Freeman, W., Tenenbaum, J.: Galileo : Perceiving
Physical Object Properties by Integrating a Physics Engine with Deep Learning.
Advances in Neural Information Processing Systems 28 (NIPS 2015) pp. 1–9 (2015)
25. Yue-Hei Ng, Joe and Hausknecht, Matthew and Vijayanarasimhan, Sudheendra
and Vinyals, Oriol and Monga, Rajat and Toderici, G.: Beyond Short Snip-
pets : Deep Networks for Video Classiﬁcation. Proceedings of the IEEE con-
ference on computer vision and pattern recognition pp. 4694—-4702 (2015).
26. Zhang, J., Pantelelis, N.: Modelling and optimisation control of poly-
mer composite moulding processes using bootstrap aggregated neural net-
work models. 2011 International Conference on Electric Information and
Control Engineering, ICEICE 2011 - Proceedings 1(2), 0–3 (2011).