PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Carbon fiber reinforced polymers (CFRP) are light yet strong composite materials designed to reduce the weight of aerospace or automotive components-contributing to reduced greenhouse gas emissions. Resin transfer molding (RTM) is a manufacturing process for CFRP that can be scaled up to industrial-sized production. It is prone to errors such as voids or dry spots, resulting in high rejection rates and costs. At runtime, only limited in-process information can be made available for diagnostic insight via a grid of sensors (e.g. ultrasound or pressure). We propose FlowFrontNet, a deep learning approach to enhance the in-situ process perspective by learning a mapping from sensors to flow front "images" (using upscaling layers), to capture spatial irregularities in the flow front to predict dry spots (using convolutional layers). On simulated data of 6 million single time steps resulting from 36k injection processes, we achieve a time step accuracy of 91.7% when using a 38 × 30 sensor grid with 1 cm sensor distance in x-and y-direction. On a sensor grid of 10 × 8, with a sensor distance of 4 cm, we achieve 83.7% accuracy. In both settings, FlowFrontNet provides a significant advantage over direct end-to-end learning models.
Content may be subject to copyright.
FlowFrontNet: Improving Carbon Composite
Manufacturing with CNNs
Simon Stieber[0000000237538264], Niklas Schr¨oter[0000000271543374],
Alexander Schiendorfer[0000000252835304], Alwin
Hoffmann[0000000251233918] and Wolfgang Reif
Institute for Software & Systems Engineering, University of Augsburg, Germany
Abstract. Carbon fiber reinforced polymers (CFRP) are light yet strong
composite materials designed to reduce the weight of aerospace or auto-
motive components – contributing to reduced greenhouse gas emissions.
Resin transfer molding (RTM) is a manufacturing process for CFRP
that can be scaled up to industrial-sized production. It is prone to errors
such as voids or dry spots, resulting in high rejection rates and costs. At
runtime, only limited in-process information can be made available for
diagnostic insight via a grid of sensors (e.g. ultrasound or pressure).
We propose FlowFrontNet, a deep learning approach to enhance the in-
situ process perspective by learning a mapping from sensors to flow front
“images” (using upscaling layers), to capture spatial irregularities in the
flow front to predict dry spots (using convolutional layers). On simulated
data of 6 million single time steps resulting from 36k injection processes,
we achieve a time step accuracy of 91.7% when using a 38 ×30 sensor
grid with 1 cm sensor distance in x- and y-direction. On a sensor grid
of 10 ×8, with a sensor distance of 4 cm, we achieve 83.7% accuracy. In
both settings, FlowFrontNet provides a significant advantage over direct
end-to-end learning models.
1 Introduction to Composite Manufacturing via RTM
Carbon fiber reinforced polymers (CFRP) are extremely strong composite ma-
terials despite their low weight. That makes them attractive for the construction
of lighter aerospace and automotive parts (conventionally made from steel or
aluminum) to reduce fuel consumption and CO2 emissions [5]. In essence, these
composites are made from a so-called polymer matrix that is reinforced with tex-
tiles containing carbon fibers. To produce CFRP parts industrially, resin transfer
molding (RTM, [1]) is a commonly applied manufacturing process for medium
volumes (1,000s to 10,000s of parts) and is depicted in Figure 1a): A liquid ther-
moset polymer (called a resin) is injected under pressure into a mold cavity that
contains reinforcement material such as textiles with carbon fibers. This results
in a “flow front” that separates impregnated material from dry material, shown
in Figure 1b. Thermoset resins are converted from a liquid to a solid through
heat – they are “cured”.
2 S. Stieber et al.
(a) Front view of the hardware involved in RTM. (b) Top view of the evolving flow
front inside (B), with a dry spot.
Fig. 1. Overview of resin transfer molding (RTM): The resin (2.) is injected into a
mold cavity (B) filled with textiles containing carbon fibers (C). A press (A) applies
the necessary pressure (1.).
During the RTM process, several errors can render the result useless and,
thus, make the overall production expensive [17,4,9]. They occur, inter alia, due
to high input variances of the fiber contents in the textile (the “preform”). Dry
spots refer to areas of the preform that are not impregnated by the liquid, as
shown in the top right of Figure 1b. In some cases, these dry spots irreparably
invalidate the stability and stiffness properties required for the manufactured
part. In others, they can be repaired manually. Either way, automated process
monitoring based on sensors applied to the mold would significantly improve
the quality assurance – called in-situ monitoring. These sensors (e.g., tempera-
ture, dielectric, ultrasound, or pressure) are able to track the flow front of the
fluid and, consequently, predict spatial deviations from proper RTM runs. This
may indicate problems such as dry spots or voids (i.e., enclosures where air is
trapped), providing diagnostic insight or even control actions to avoid rejects.
In this paper, we propose to use machine learning, in particular convolutional
neural networks (CNN), to get a binary classifier f:Rn→ {0,1}from nmold-
integrated sensors to labels describing whether, at a given point in time, there is a
dry spot present or not. As an intermediate step, we train the network to generate
higher-resolution images of the flow front (extracted from an RTM simulation in
PAM-RTM1) from sensor data. These images capture spatial irregularities in the
flow. Since such full images are only available during simulation or specialized
permeability studies [3] and not in real-world closed molds, a trained model could
substantially enhance the spatial information transmitted by actual flow front
sensors in productive settings – in the sense of a digital twin [8]. In this paper,
pam-composites/pam-rtm-composites-molding- simulation-software
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 3
we focus on detecting dry spots from sensor input obtained from simulated data
as a first step towards a transfer to real data. These are major quality concerns
during the injection process [7].
Our model, FlowFrontNet, first uses several deconvolutional and convolu-
tional layers to create the image representation from pressure sensor data and
proceeds to perform the binary classification using convolutional and dense lay-
ers. The overall classification is performed individually for every frame of a sim-
ulated injection sequence such that a warning of dry spots can be issued at any
time. Our central scientific question is if we can detect dry spots from simu-
lated sensor data and whether the spatial flow-front information captured in
convolutional layers improves the classification. We compare the accuracy the
simulation-enhanced FlowFrontNet achieves to that of a standard feed-forward
network (both based on sensor data) and find major improvements in Section 4.
Our goal is to offer FlowFrontNet as a starting point for future research in com-
posite manufacturing: code2, checkpoints [19] and data [20] are available.
Following a discussion of related work, we present the data regime including
input variation for simulation and automated label acquisition in Section 2,
present the neural network models and training methodology in Section 3, and
conclude with experimental results in Section 4.
1.1 Related Work
To cover the relevant related work, two aspects have to be addressed: The tech-
nical process and the machine learning models.
There have been several publications on in-situ monitoring the RTM process.
Pantelelis et al. [12] present a system on how to detect the curing state but only
mention ML-based analyses as future work. Zhang et al. [22] use simulated and
actual sensor data to detect the curing rate of the process. They use aggregated
shallow neural networks to achieve this based on two sensors. They further use
the model to adapt the heating of the process. Our work, by contrast, focuses on
the resin injection and detect dry spots in the flow front from multiple sensors.
Leveraging the spatial capabilities inherent to CNNs has never been applied to
dry spot detection in the RTM process before.
Unrelated to composite production, deconvolutional layers that increase the
spatial dimensions have been used for image enhancement tasks. Xu et al. [21]
used them to deblur images. Shi et al. [15] approach superresolution, the task of
upscaling pictures from low to high resolution, with deconvolutional networks.
For semantic segmentation, Noh et al. [11] used a combination of convolutional
and deconvolutional networks. They compress the input image with a VGG-16
to afterward decompress the encoding with deconvolutional layers.
Out of these approaches, [15] comes closest to our task of extracting infor-
mation from a grid of sensors since the compressed data is known before and
not a learned encoding as in [11]. However, our approach works with more het-
erogeneous data in that it generates images from sensors rather than improving
4 S. Stieber et al.
the resolution of conventional photos. Furthermore, the dimensions of the input
are enlarged and not kept the same as in [21].
2 Creating Training Data from Simulation
To train a detector for dry spots from RTM sensor data, we need to observe
sufficiently many training instances. Generally speaking, industrial ML use cases
based on sensor data from actual production cycles tend to suffer from limited
data quality and quantity. Moreover, setting up a new process (including design
of the mold, choice of the resin, etc.) takes time until sufficiently many training
runs have been processed, not to mention the material costs. Simulation is a
proven remedy for the lack of data [13] which is our focus in this paper. Basic
RTM processes are well supported by existing engineering tools [3].
The lack of real high-quality data is only one reason to opt for simulation.
Another more pressing reason is that we can create a spatial representation of
the flow front, i.e., a “flow front image” (see Figure 3a), that is not observable
in real closed molds. Those images will serve as the target for the generative
Deconv/Conv part of FlowFrontNet, described in Section 3.
Being able to simulate the process might raise the question why one has to
apply machine learning at all – instead of just running the simulation online.
First, the trained models will encapsulate only those aspects of the simulation
that are needed to make good dry spot predictions for a given RTM setup. Online
simulation would take much longer and be infeasible for real-time monitoring.
Second, we can reasonably anticipate that real runs will contain aspects that
are not properly captured by simulation. For instance, variability in the process
parameters such as the textile might not be observable. Using an ML-model will
eventually enable us to add real data to the training.
In the following, we describe our process of obtaining enough (6 million)
domain-specific training instances from simulations executed with randomized
initial conditions (variances in the input textile) and automatically deriving the
corresponding dry spot labels that are images with the filling level as intensities.
2.1 Simulated RTM Runs in PAM-RTM
PAM-RTM is a software package designed for laying out RTM processes, in-
cluding fluid dynamics simulation. Modeling flow transport in porous media
mathematically is most commonly carried out based on Darcy’s law [2] for one-
dimensional flow:
∆x (1)
where vxis the 1D flow velocity, kxrepresents the permeability value of the
textile (corresponding to how “easily” fluids can permeate it), ηdenotes the fluid
viscosity, and ∆P
∆x expresses the pressure drop along a specific flow length [3].
PAM-RTM includes other features as well, including draping simulation, the
meshing of CAD parts, and distortion during curing which makes it a stan-
dard tool in composite manufacturing. For this paper, we only needed the fluid
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 5
Fig. 2. Simulated composite plate with a full sensor grid and central injection point -
1140 sensors located at distance 1 cm in xand y. On the right, a fiber volume content
(FVC) map is given with local perturbations (rectangular and circular) .
simulation part modeling resin injection. Getting the flow front prediction and,
subsequently, dry spot classification right is a high-impact quality goal.
Setting up a particular RTM simulation consists of defining the geometrical
3D-model of the part to be constructed, the viscosity and permeability param-
eters of the resin and textile, respectively, as well as temperature and pressure
of the injection. A single simulation run then represents a whole injection that
continues to pump resin until the mold cavity is filled entirely.
As our running example, we choose a simple rectangular composite plate
with dimensions of 38 ×30 cm. Adding a simulated pressure sensor for every
centimeter in xand ydirection yields a total of 1140 sensors, as Figure 2 shows.
The sensors can be placed at any location in the simulated composite plate,
independent of the mesh grid of the plate. To ensure that the chosen parameters
are realistic, this configuration follows the setup for experimental permeability
characterization available at the Processing of Composites Group of Montanuni-
versit¨at Leoben as presented by Gr¨ossing et al. [3]. For the textile, the setup
assumes a natural fiber fabric with a fiber volume content (FVC) of 0.268 and
a 3.5 mm thickness. Finally, the resin used in the model experiment in Leoben
was a plant oil with a viscosity of 0.065 N·s
In reality, dry spots emerge from local variations in the textiles (thicker or
thinner areas) that lead to regions of low or high fluid permeability, respectively.
To recreate that effect in our simulation and obtain varied training data, we
altered the fiber volume content in certain areas, as shown in Figure 2. These
6 S. Stieber et al.
Table 1. Data set sizes – train/val/test-split
Data set # Runs # Samples % (Samples)
Training 29,663 5,067,352 81.11
Validation 756 131,072 2.09
Test 6,244 1,048,576 16.78
All 36,663 6,247,000 100
perturbations are responsible for the distribution our training and test RTM runs
are drawn from. For simplicity, every perturbation corresponds to one circle with
a random diameter and one rectangle with random side lengths – with varying
strengths of FVC perturbation. All of these perturbations were automatically
created for 40k simulation setups that were run in parallel on 10 hosts with 32
cores each. We repeated the simulations with different random strengths of FVC
perturbations: setting them to 0.1 - 0.8 covers a wide range of cases (including
mild cases that will eventually not produce dry spots), increasing them to 0.2 -
0.8 or even 0.3 - 0.8 provokes more severe perturbations leading to dry spots.
For training our models, Table 1 shows how we split the data obtained from
those FVC-perturbed simulation runs. Although we consider individual frames
(corresponding to time steps of runs), the splits do not break up runs, i.e., a run
is fully contained in either train, validation, or test set. That allows for reusing
the data when considering sequence models in the future.
When working with simulations, it is imperative to double-check the time
series resolution. We noted that PAM-RTM distinguishes between multi-state
results for the simulated pressure sensors and single-state results for the simu-
lation results at every node of the mesh of the plate such as filling status and
pressure. While multi-state results are present for every simulated time step
(including the respective simulation times), single state results are saved every
k-th second in simulation time or every i-th simulation step. To keep the sim-
ulation efforts tractable and produce enough runs, we selected an approximate
time resolution of 0.5 seconds and dispose of all pressure sensor results that do
not correspond to time steps with available filling status. Finally, before feeding
the pressure sensors’ values to the ML-models, we divide them by 105to obtain
numerically well-behaved training dynamics in the early layers of the neural
2.2 Dry Spot Label Creation
Varying the FVC locally in the input textiles and recording the simulated runs
provides us with pressure value time series and flow front developments as “im-
ages” where every pixel corresponds to the filling level at a given point in time.
To classify dry spots in a supervised manner, we also need labels indicating
whether a flow front image contains a dry spot (e.g., Figure 3a) or not (e.g.,
Figure 3c). And we need those labels for all 6 million frames from 36k runs.
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 7
(a) Interior dry spot (b) FVC map corre-
sponding to 3a
(c) No dry spot at
the flow front
(d) FVC map corre-
sponding to 3c
Fig. 3. Exemplary flow fronts resulting from locally perturbed FVC-maps extracted
from specific time steps
An area counts as a dry spot if it is a non-filled area that is enclosed by
resin. Comparing Figure 3a and Figure 3b, we might be tempted to directly
derive binary labels from the modified FVC maps that served as simulation
input. However, the matter is more complicated. First, an observed dry spot
tends to be smaller than the original perturbation since the outskirt still gets
permeated by resin. Second, an FVC perturbation would be the same for all
frames of a run even though the dry spot only becomes apparent after the flow
front reached it geometrically. Third, a regular flow front might be jagged such
as the one in Figure 3c. Such areas should not be counted (yet) as dry spots since
they tend to eventually be filled. Thus, we devised a heuristic that combines the
perturbed FVC-maps with computer vision techniques on the flow front images
to output label maps indicating if a pixel belongs to a dry spot or not.3
First, we extract the already filled pixels from each frame, e.g., the light
(yellow) pixels in Figure 3a. The negative of that image only contains the dry
parts. Since dry spots cannot lie within the filled regions, we focus on the latter
for finding contours. We use OpenCV4to find contours in those negative images.
The contours are constrained to be smaller than the whole flow front but at least
larger than the central injection point. However, not all of the contours identified
as dry spot candidates must be dry spots. Instead, they could emerge from jagged
flow fronts as apparent in Figure 3c) that are not enclosed by resin. To distinguish
between these two cases, we enhance the pure computer vision operations (such
as contour or hole detection) with the perturbed FVC maps. Since a high FVC
corresponds to low permeability and vice versa, we overlap the dry spot candidate
contours with the FVC map. For every contour in the dry area, its probability of
being a dry spot is estimated proportionally to the percentage of the overlap with
3Note that we did not use the “air entrapment” feature in PAM-RTM since that
would prematurely end simulation runs, produces lagging information, and cause
the experimental setup to diverge from the model setup in Leoben.
8 S. Stieber et al.
Fig. 4. FlowFrontNet: The first part (dashed, green) is a Deconv/Conv network that
maps from sensors to images. The second part (solid, orange) is the DrySpotNet consist-
ing of five convolutional/max-pooling layers and two fully-connected layers to perform
the final classification task. Numbers denote the resulting feature maps.
an FVC perturbation. To reduce the number of incorrectly labeled dry spots, we
look at runs, i.e., sequences of frames, and discard candidates that only occur
in a single frame. This is justified since the flow front is expanding continuously
and, thus, the probability of a dry spot has to be similar for multiple frames in
a run. Some edge cases (e.g., simulation errors or dry spots in areas other than
the FVC perturbations) lead to unexpected jumps in dry spot probabilities over
consecutive frames. To maintain high-quality data, we excluded whole runs with
such phenomena – approximately 9.16% of all runs which leaves us with 36,663
valid runs (cf. Table 1).
3 Approach - Model and Training
After generating data for different FVC contents in a sufficient amount and
labeling them in an automated manner, the next step is to present FlowFrontNet,
the main model of our approach, which is shown in Figure 4. Before going into
detail, we present the key points of this network.
The overall FlowFrontNet maps pressure sensor inputs to classification de-
cisions. It consists of the generative part that upsamples from sensor grids to
flow front images (called Deconv/Conv) and the binary classification part (called
DrySpotNet). By learning to produce a flow front from sensor input, the network
learns a representation of fluid dynamics. The Deconv/Conv part itself is useful
for other use cases down the road, say, exact dry spot localization or pixel-wise
flow front detection. The later DrySpotNet performs a binary classification on
the generated images (see Figures 5b and 5c for example inputs to that part).
By adding spatial fluid dynamics knowledge, we aim to surpass the performance
of a conventional feed-forward classifier achieves based on the same sensor input.
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 9
We present the version of FlowFrontNet that is used for 1140 sensors. The
other sensor resolutions (see Table 2) only require slight changes in kernel size
and layer count. Those are necessary since the smaller input sizes lead to smaller
outputs with the hyperparameters used for the 1140-sensors Deconv/Conv net-
work. Due to the resulting poor image resolution, the results would not be ad-
equate for detecting dry spots. In the following, we describe the intricacies of
each network and their training processes.
Deconv/Conv Network: Sensor Data to Flow Front Images The first part of
FlowFrontNet, Deconv/Conv, is a fully convolutional neural network [14] (see
Figure 4, left part) that receives the low-resolution sensor grid values and out-
puts a high-resolution flow front image. The first four layers are deconvolutional
to extract features from the sensor array. As opposed to convolutional layers,
deconvolutional layers increase the spatial dimensions of an image, known from
image segmentation or superresolution [11,15].
Our approach is similar in this regard: the pressure sensor grid has a low
resolution and constitutes a compressed representation of the flow front. We ap-
ply deconvolutional layers to increase the resolution of the sensor array while
simultaneously filling the spaces between the sensors, i.e., interpolating miss-
ing values. Afterward, we utilized five standard convolutional layers (providing
the required amount of non-linearity) to create and shape the final flow front
image (e.g., with spatial dimensions 117 ×149).
The first step is to pre-train the Deconv/Conv network to produce images
of the flow front from sensors, as shown in the left part of Figure 4. As men-
tioned before, this pre-training needs simulated data since flow fronts are hidden
in real-world closed molds. All hidden layers are followed by the rectified lin-
ear unit (ReLU) activation, while the output layer is activated by the sigmoid
function to make sure the generated flow front images lie within the range [0,1].
DrySpotNet: Flow Front Image to Binary Dryspot Classification The second
part of FlowFrontNet, DrySpotNet, receives the generated images and classi-
fies them concerning dry spots. The dry spot labels are obtained as described
in Section 2.2. The architecture for this classification follows a standard con-
volutional classification network [6] (see the latter part of Figure 4). The final
output yields soft classification scores using the sigmoid function (commonly
interpreted as class probabilities) that are eventually thresholded to achieve a
hard classification.
FlowFrontNet: Training End-to-End Sensor Data to Dryspot After pretraining
the Deconv/Conv net using flow front images, we append the untrained DrySpot-
Net for the final classification. To avoid changing the already trained weights of
the generative Deconv/Conv layers, the pre-trained layers are “frozen”, mean-
ing that their weights are not updated during training. This is identical to the
practice of freezing convolution layers in fine-tuning for special purpose image
classification [16]. After training the newly appended output layer, it can be
useful to also “unfreeze” the pre-trained weights of early hidden layers during
10 S. Stieber et al.
Table 2. Sensor layouts
Sensors Layout Sensor distance - x and y
1140 38 ×30 1 cm
80 10 ×8 4 cm
20 5 ×4 8 cm
backpropagation. The data is exchanged and parts of the original network are
used as the backbone for a new image processing task. Here, our approach is
different: we change the objective and targets from generating images to that of
doing dry spot binary classification, leaving the early layers fixed.
When combining the generative part of the Deconv/Conv net with the DrySpot-
Net, low sensor input numbers may lead to comparably “blurry” flow front im-
ages (e.g., Figures 5b to 5d) before handing them to the classification part. To
obtain higher-contrast flow front images and, consequently, better classification
results, we add a non-differentiable hard pixel-threshold to the forward pass of
the network after the Deconv/Conv net. This operation sets all values below the
threshold to 0 and all above to 1. Since the flow-front-generating layers of the
networks that lie before the pixel-threshold are frozen and need no gradients, we
can easily incorporate this operation.
Feed-forward Network: Baseline Classifier To judge the merits of FlowFrontNet,
we design a basic end-to-end feed-forward network as a baseline classifier. It
consists of two ReLU-activated, fully-connected hidden layers and a sigmoid
output for the dry-spot probabilities. As input, it receives the sensor values
described in Section 2 without performing upsampling to the flow front image.
For training all models, we utilized an Nvidia DGX-1 with 8 Tesla V100
GPUs. This machine is able to train with a batch size of 2048 for both steps of the
training process. Especially the Deconv/Conv network is very resource-intensive
in terms of its parameters, with the highest consumption of computation power
when using 1140 sensor values as input. For this training, the DGX-1 reached its
maximum load with batch size 2048, for training runs with smaller sensor grids,
the same batch size yielded the best performance compared to larger batch sizes.
4 Experimental Evaluation
To put FlowFrontNet to the test, we devised three central evaluation hypotheses.
The first is if we can predict a dry spot from simulated pressure sensors at all.
The second and central hypothesis is, whether the intermediate step of flow
front image generation can improve dry spot classification. This also involves
testing the pixel thresholding introduced to obtain sharper internal flow front
representations. As a third point that is interesting to evaluate, we investigate
the number of sensors that is necessary to classify dry spots sufficiently well,
with and without the intermediate flow front generation. We investigate models
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 11
Table 3. Accuracy values on three different sensor resolutions for feed-forward base-
lines and FlowFrontNet
Feed-forward FlowFrontNet
# Sensors Threshold Accuracy Threshold Accuracy
1140 0.54 82.74% 0.49 91.68%
80 0.52 79.57% 0.57 83.69%
20 0.49 74.68% 0.51 75.22%
for 1140, 80, and 20 sensors, to estimate the prediction quality achievable by a
reduced number of sensors, see also Table 2.
These particular numbers emerge from taking every 4-th or every 8-th sensor
of the full 1140 grid. Taking 1140 sensors corresponds to a sensor distance of 1
cm, which is an unrealistically high number. To come closer to reality, we focused
on a sensor distance of 4 cm for the 80 sensors because that is within range of
physical feasibility and set a baseline with even fewer sensors at 8 cm distance.
4.1 Results
Can machine learning predict dry spots based on sensor inputs? To start with the
first question, we obtained a baseline from the feed-forward network performing
the classification task directly from sensor inputs. When optimizing this archi-
tecture, we found that larger batch sizes positively affected the evolution of the
validation loss, with a sweet spot found at 32,768 = 215 training instances per
batch. The results from experiments suggest the following baselines: the best
feed-forward network with two hidden layers achieves an accuracy of 82.73%
with 1140 sensors, an accuracy of 79.51% with 80 sensors, and an accuracy of
74.6% with 20 sensors, see Table 3. Unsurprisingly, the more sensors we use, the
better the accuracy gets for a network that does not know anything about the
underlying fluid dynamics.
(a) Label (b) 1140 sensors (c) 80 sensors (d) 20 sensors
Fig. 5. Exemplary flow front predictions in the Deconv/Conv part of FlowFrontNet
based on the different sensor counts.
12 S. Stieber et al.
Does the Deconv/Conv net improve the classification accuracy? For the second
question – the improvement of the classification by generating flow front im-
ages using deconvolutional and convolutional layers – the answer is a clear yes,
with only a slight limitation. Figure 5 allows us to visually inspect exemplary
results from the learned sensor-to-flow-front mapping on the test set. Even for
20 sensors, a rough idea of the underlying flow front is obtained, for 80 and
more sensors, a fairly accurate image can be reconstructed. Based on that in-
ternal representation of the flow front, the accuracy of FlowFrontNet with 1140
sensors as input increases to 91.68%, which is a 9% advantage over the pure
Feed-forward Network. For 80 sensors (with pixel-thresholding), the accuracy
can be enhanced from 79.51% to 83.69%, a margin of 4%. Furthermore, the 80
sensor FlowFrontNet performs better than the feed-forward network with 1140
sensors which shows that sensor investments could be reduced in favor of encoded
simulation knowledge.
Alas, the improvement over the feed-forward network decreases even more
with 20 sensors: 75.22%, which is less than 1% of accuracy boost. The spatial
information density in a 5 ×4 input sensor grid turned out too low to still get
a useful representation, which can also be observed in Figure 5d. The image of
the flow front is not clear at all and it appears as though there is no dry spot
enclosed by resin but rather a jagged flow front. Therefore, we only focus on the
models equipped with 80 or 1140 sensor inputs.
To get a more comprehensive performance overview, Table 4 shows the con-
fusion matrices for the 1140 and 80 sensors input, respectively. The models are
rather balanced regarding their false positive and false negative shares, but these
values are – as always – object to modification by the classification threshold
applied to the output probability score. Especially in industrial processes, one
type of error can be more favorable than the other, depending on whether false
positives (e.g., causing disruptions in the process) or false negatives (e.g., leading
to undetected errors in the produced parts) are more acceptable. For the behav-
ior of the classifiers under various classification thresholds, consider the ROC
curve in Figure 6. The confusion matrices here are given for the classification
threshold value that gives the highest validation accuracy, as per Table 3.
Is the pixel-thresholding step in FlowFrontNet useful? By experimenting with
possible pixel thresholds (see Section 3) of 0.2, 0.5, and 0.8, we found that 0.8
yielded the best results and is used for the 80-sensor results (also in the already
presented results). The accuracy increased from 81.48 to 83.69%. We observed
that the training loss decreases steeper and farther without pixel-thresholds but
the validation loss increases to a greater extent. By contrast, the validation loss
for the pixel-threshold model declined steadily – indicating that the thresholding
of flow front images regularizes.
Moreover, training more than one epoch without pixel-thresholds produced
heavy overfits. Even with an exponential learning rate scheduler and a very low
initial learning rate, the training dynamics did not improve. While the valida-
tion loss was decreasing, it was at a higher base level than before. With pixel-
thresholds in place, training got easier, even without scheduling the learning
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 13
Table 4. Confusion matrices for 1140 and 80 sensors: Feed-forward vs. FlowFrontNet
For 1140 sensors
¬Dry spot Dry spot
Pred. ¬Dry spot 40.44% 8.59%
Dry spot 8.66% 42.31%
¬Dry spot Dry spot
Pred. ¬Dry spot 44.60% 4.43%
Dry spot 3.89% 46.98%
For 80 sensors
¬Dry spot Dry spot
Pred. ¬Dry spot 38.97% 10.06%
Dry spot 10.37% 40.06%
¬Dry spot Dry spot
Pred. ¬Dry spot 41.60% 7.44%
Dry spot 8.87% 42.10%
Fig. 6. ROC Curves for different models and sensor inputs. * pixel-threshold at .8
rate but using a fixed value of 104with AdamW [10]. Alas, for the smallest 20
sensor grid, pixel-thresholding did not give better results.
The ROC curves and the corresponding area under the curve (AUC) values
are shown in Figure 6. The AUC for the .8-pixel-thresholded model is the best of
all 80 sensor models, but only by a small margin. The other curves do not hold
any surprises, with the 1140 sensor model outperforming all and all other models
with similar AUCs. Only the 80 sensor feed-forward net is underperforming.
4.2 Discussion - Metrics on Run Level
Our previous evaluations exclusively considered metrics per frame that are each
drawn from many independent injection runs. Such a dry spot classification for
every frame and, thus, point in time is desirable to closely monitor the RTM
process (in a “digital twin”-manner) to intervene at process execution time by
adjusting process parameters.
Additionally, practitioners care about a judgment concerning the process
quality of a single run, e.g., a single produced composite plate. This is similar to
lifting single-frame classification to video classifications, e.g., how many dry spot
14 S. Stieber et al.
(a) (b)
Fig. 7. Different classifications of two runs by three different models
frames make a dry spot run? The first difficulty is to decide when a run counts
as a failure and doing so automatically for both the label maps and predictions
for all 36,663 runs.
The naive approach would take the last few frames to determine if a run
counts as having produced a dry-spot. However, we would miss dry spots that
only occur in the middle of a run. These might hint at problems in reality
and only “close” due to simulation artifacts (e.g., unrealistically increasing the
pressure). Another possible criterion counts all dry spot frames and prescribes
a minimal amount to mark a run as failed. Alternatively, we could require the
dry spot sequences to be contiguous, to avoid listing too many runs as failed if
there are single dry spots in the label maps or predictions which could also be
artifacts from the label generation process described in Section 2.2. Figures 7a
and 7b offer some insight into how different run classifiers would behave for two
exemplary runs, given the frame-wise ground truth and predictions.
Here, it becomes apparent that the models perform very differently and also
confirms that the models taking 80 sensors as input cannot be compared to the
1140 sensor input models. The 1140-sensor-models adhere closely to the label
most of the time whereas both the 80-sensor-models produce more noise, in
different ways.
We anticipate further experiments as future work, especially combining the
single-frame predictions with another sequence model to classify whole runs –
provided that we can obtain or define meaningful run labels.
5 Conclusion and Future Work
We presented FlowFrontNet, a deconvolutional/convolutional neural network
suitable for detecting dry spots in simulated RTM processes, i.e., improved
process monitoring. In doing so, we showed that it is possible to learn the in-
termediate representation of flow fronts from sensor data by upsampling via a
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 15
deconvolutional network. This enabled us to reliably classify dry spots on indi-
vidual frames substantially better than with a feed-forward network using the
same sensor input. The classifier makes it possible to intervene during the man-
ufacturing of a single composite plate. We also investigated that the prediction
quality decreases with the number of sensors in use, and found that acceptable
accuracy requires at least 80 sensors, a sensor grid of 4 cm distance. That is a
realistic order of magnitude for real mold sensor layouts whereas 1140 sensors
cannot be placed as closely due to cost and wiring issues.
Future work can be divided into short and long term goals that focus on use
cases with simulated and actual data, respectively. In the short run, we plan
to use the flow front image generated from sensors for other classification or
object detection tasks and use sequence models for predicting full runs, with
their temporal information. Our long term goal is to use models pre-trained on
simulated data for real RTM process data [18], much like sim-to-real applications
are being used in reinforcement learning and robotics [13]. In addition to the
costly process of actually producing composite plates in a sufficient number for
training, reality confronts us with heterogeneous and noisy sensors other than
pressure alone. Eventually, we want to generate feedback for process control to
avoid rejects. The first step towards this goal, learning the flow front from sensor
data and applying it to dry-spot classification, has been successfully achieved.
Acknowledgments This research is funded by the Bavarian Ministry of Economic
Affairs, Regional Development and Energy in the project CosiMo. We thank Ewald
Fauster from Montanuniversit¨at Leoben for his expert advice on the RTM process and
Frederic Masseria from ESI for supporting our RTM-simulations.
1. Babb, D.A., Richey, W.F., Clement, K., Peterson, E.R., Kennedy, A.P., Jezic, Z.,
Bratton, L.D., Lan, E., Perettie, D.J.: Resin transfer molding process for compos-
ites (1998)
2. Darcy, H.P.G.: Les Fontaines publiques de la ville de Dijon. V. Dalamont (1856)
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 (9 2016).
4. Heuer, H., Schulze, M., Pooch, M., G¨abler, S., Nocke, A., Bardl, G., Cherif,
C., Klein, M., Kupke, R., Vetter, R., Lenz, F., Kliem, M., B¨ulow, C., Goy-
vaerts, J., Mayer, T., Petrenz, S.: Review on quality assurance along the CFRP
value chain - Non-destructive testing of fabrics, preforms and CFRP by HF
radio wave techniques. Composites Part B: Engineering 77, 494–501 (2015).
5. Heywood Don MacKenzie, J., Bonde Akerlind Parisa Bastani Irene Berry Kandarp
Bhatt Alice Chao Eric Chow Valerie Karplus David Keith Michael Khusid Eriko
Nishimura Stephen Zoepf, I.: On the Road toward 2050: Report Massachusetts
Institute of Technology Potential for Substantial Reductions in Light-Duty Vehicle
Energy Use and Greenhouse Gas Emissions. Massachusetts Institute of Technology
16 S. Stieber et al.
6. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to
document recognition. Proceedings of the IEEE 86(11), 2278–2324 (1998)
7. Lee, D.H., Lee, W.I., Kang, M.K.: Analysis and minimization of void formation
during resin transfer molding process. Composites Science and Technology 66(16),
3281–3289 (2006)
8. Lee, J., Bagheri, B., Kao, H.A.: A cyber-physical systems architecture for industry
4.0-based manufacturing systems. Manufacturing letters 3, 18–23 (2015)
9. Liu, B., Bickerton, S., Advani, S.G.: Modelling and simulation of resin trans-
fer moulding (RTM) - Gate control, venting and dry spot prediction. Com-
posites Part A: Applied Science and Manufacturing 27(2), 135–141 (1996).
10. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: 7th Interna-
tional Conference on Learning Representations, ICLR 2019 (11 2019)
11. Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmen-
tation. In: Proceedings of the IEEE International Conference on Computer Vision.
vol. 2015 Inter, pp. 1520–1528 (5 2015).
12. Pantelesis, N., Bistekos, E.: Process monitoring and control for the production of
CFRP components. SAMPE Conference pp. 5–9 (2012)
13. Peng, X.B., Andrychowicz, M., Zaremba, W., Abbeel, P.: Sim-to-real transfer of
robotic control with dynamics randomization. In: 2018 IEEE international confer-
ence on robotics and automation (ICRA). pp. 1–8. IEEE (2018)
14. Shelhamer, E., Long, J., Darrell, T.: Fully Convolutional Networks for Semantic
Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence
39, 640–651 (4 2017).
15. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D.,
Wang, Z.: Real-Time Single Image and Video Super-Resolution Using an Efficient
Sub-Pixel Convolutional Neural Network. In: Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition. vol. 2016-Decem,
pp. 1874–1883 (9 2016).
16. Shin, H.C., Roth, H.R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mol-
lura, D., Summers, R.M.: Deep Convolutional Neural Networks for Computer-
Aided Detection: CNN Architectures, Dataset Characteristics and Transfer
Learning. IEEE Transactions on Medical Imaging 35(5), 1285–1298 (5 2016).
17. Sorg, C.: Data Mining als Methode zur Industrialisierung und Qualifizierung neuer
Fertigungsprozesse f¨ur CFK-Bauteile in automobiler Großserienproduktion. Ph.D.
thesis, Technische Universit¨at M¨unchen (2014)
18. Stieber, S.: Transfer Learning for Optimization of Carbon Fiber Reinforced Poly-
mer Production. Organic Computing: Doctoral Dissertation Colloquium 2018 pp.
1–12 (2018)
19. Stieber, S.: FlowFrontNet Checkpoints (2020).,
FlowFrontNet Checkpoints/12102714
20. Stieber, S.: FlowFrontNet Data: Sensor to Flowfront / Dryspot (2020).,
FlowFrontNet Data Sensor to Flowfront Dryspot/12063480
21. Xu, L., Ren, J.S., Liu, C., Jia, J.: Deep convolutional neural network for image
deconvolution. Advances in Neural Information Processing Systems 2(January),
1790–1798 (2014)
FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs 17
22. 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).
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
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 to draw samples of produced components statistically and perform a destructive test 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 Resin Transfer Molding press.
Full-text available
Remarkable progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets and the revival of deep CNN. CNNs enable learning data-driven, highly representative, layered hierarchical image features from sufficient training data. However, obtaining datasets as comprehensively annotated as ImageNet in the medical imaging domain remains a challenge. There are currently three major techniques that successfully employ CNNs to medical image classification: training the CNN from scratch, using off-the-shelf pre-trained CNN features, and conducting unsupervised CNN pre-training with supervised fine-tuning. Another effective method is transfer learning, i.e., fine-tuning CNN models pre-trained from natural image dataset to medical image tasks. In this paper, we exploit three important, but previously understudied factors of employing deep convolutional neural networks to computer-aided detection problems. We first explore and evaluate different CNN architectures. The studied models contain 5 thousand to 160 million parameters, and vary in numbers of layers. We then evaluate the influence of dataset scale and spatial image context on performance. Finally, we examine when and why transfer learning from pre-trained ImageNet (via fine-tuning) can be useful. We study two specific computer-aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, with 85% sensitivity at 3 false positive per patient, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks.
Full-text available
Eddy current testing is well established for non-destructive testing of electrical conductive materials [1]. The development of radio frequency (RF) eddy current technology with frequency ranges up to 100 MHz made it possible to extend the classical fields of application even towards less conductive materials like CFRP [2][3](Table 2). It turns out that RF eddy current technology on CFRP generates a growing number of valuable information for comprehensive material diagnostic. Both permittivity and conductivity of CFRP influence the complex impedance measured with RF eddy current devices. The electrical conductivity contains information about fiber texture like orientations, gaps or undulations in a multilayered material. The permittivity characterization influenced by dielectric properties allows the determination of local curing defects on CFRP e.g. hot spots, thermal impacts or polymer degradation. An explanation for that effect is seen in the measurement frequency range and the capacitive structure of the carbon rovings. Using radio wave frequencies for testing, the effect of displacement currents cannot be neglected anymore. The capacitive structures formed by the carbon rovings is supposed to further strengthen the dielectric influences on eddy current measurement signal [3]. This report gives an overview of several realized applications and should be understood as a general introduction of CFRP testing by HF Radio Wave techniques.
Full-text available
Recent advances in manufacturing industry has paved way for a systematical deployment of Cyber-Physical Systems (CPS), within which information from all related perspectives is closely monitored and synchronized between the physical factory floor and the cyber computational space. Moreover, by utilizing advanced information analytics, networked machines will be able to perform more efficiently, collaboratively and resiliently. Such trend is transforming manufacturing industry to the next generation, namely Industry 4.0. At this early development phase, there is an urgent need for a clear definition of CPS. In this paper, a unified 5-level architecture is proposed as a guideline for implementation of CPS.
Conference Paper
Recently, several models based on deep neural networks have achieved great success in terms of both reconstruction accuracy and computational performance for single image super-resolution. In these methods, the low resolution (LR) input image is upscaled to the high resolution (HR) space using a single filter, commonly bicubic interpolation, before reconstruction. This means that the super-resolution (SR) operation is performed in HR space. We demonstrate that this is sub-optimal and adds computational complexity. In this paper, we present the first convolutional neural network (CNN) capable of real-time SR of 1080p videos on a single K2 GPU. To achieve this, we propose a novel CNN architecture where the feature maps are extracted in the LR space. In addition, we introduce an efficient sub-pixel convolution layer which learns an array of upscaling filters to upscale the final LR feature maps into the HR output. By doing so, we effectively replace the handcrafted bicubic filter in the SR pipeline with more complex upscaling filters specifically trained for each feature map, whilst also reducing the computational complexity of the overall SR operation. We evaluate the proposed approach using images and videos from publicly available datasets and show that it performs significantly better (+0.15dB on Images and +0.39dB on Videos) and is an order of magnitude faster than previous CNN-based methods.
Convolutional networks are powerful visual models that yield hierarchies of features. We show that convolutional networks by themselves, trained end-to-end, pixels-to-pixels, improve on the previous best result in semantic segmentation. Our key insight is to build "fully convolutional" networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning. We define and detail the space of fully convolutional networks, explain their application to spatially dense prediction tasks, and draw connections to prior models. We adapt contemporary classification networks (AlexNet, the VGG net, and GoogLeNet) into fully convolutional networks and transfer their learned representations by fine-tuning to the segmentation task. We then define a skip architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations. Our fully convolutional network achieves improved segmentation of PASCAL VOC (30% relative improvement to 67.2% mean IU on 2012), NYUDv2, SIFT Flow, and PASCAL-Context, while inference takes one tenth of a second for a typical image.
Many fundamental image-related problems involve deconvolution operators. Real blur degradation seldom complies with an ideal linear convolution model due to camera noise, saturation, image compression, to name a few. Instead of perfectly modeling outliers, which is rather challenging from a generative model perspective, we develop a deep convolutional neural network to capture the characteristics of degradation. We note directly applying existing deep neural networks does not produce reasonable results. Our solution is to establish the connection between traditional optimization-based schemes and a neural network architecture where a novel, separable structure is introduced as a reliable support for robust deconvolution against artifacts. Our network contains two submodules, both trained in a supervised manner with proper initialization. They yield decent performance on non-blind image deconvolution compared to previous generative-model based methods.
We propose a novel semantic segmentation algorithm by learning a deconvolution network. We learn the network on top of the convolutional layers adopted from VGG 16-layer net. The deconvolution network is composed of deconvolution and unpooling layers, which identify pixel-wise class labels and predict segmentation masks. We apply the trained network to each proposal in an input image, and construct the final semantic segmentation map by combining the results from all proposals in a simple manner. The proposed algorithm mitigates the limitations of the existing methods based on fully convolutional networks by integrating deep deconvolution network and proposal-wise prediction; our segmentation method typically identifies detailed structures and handles objects in multiple scales naturally. Our network demonstrates outstanding performance in PASCAL VOC 2012 dataset, and we achieve the best accuracy (72.5%) among the methods trained with no external data through ensemble with the fully convolutional network.
Liquid composite molding (LCM) techniques are innovative manufacturing processes for processing fiber reinforced polymer parts used e.g. for aerospace structures. Thereby the reinforcing material is placed in a mold and infiltrated with a low viscosity polymer matrix. Increasing production rates as well as part complexity lead to high production risks such as air inclusions or incomplete mold filling. Numerical mold filling simulations are promising tools enabling the composite manufacturing engineer to detect dry spots in the mold and find the optimal positions of the resin entry and ventilation system at an early process development stage. Today, different numerical models and software packages are available for modeling the flow through the reinforcing structure for visualization of the flow behavior. The goal of this study is the systematic comparison of two different software packages, namely PAM-RTM® and OpenFOAM. Both software tools are operated as they are commonly foreseen. Real world experiments under real process conditions are the basis for the assessment of the numerical predictions. The resin transfer molding (RTM) experiments are executed in a tool with a transparent upper mold half in order to see the flow front advancement. POLYM. COMPOS., 2015. © 2015 Society of Plastics Engineers