SAMPE Journal - November/December 2020 - 48

FEATURE / COMPLEX CFRP PARTS
volume data9. This architecture can be modified
easily to support practically any number of output
classes that are suitable for the respective application.
When using machine learning, it is often difficult to acquire a sufficient amount of training
data. We tackle this issue by using artificially generated data. To demonstrate the performance of
our network, we implemented the following classes: "background", "cuboid", and "cylinder". We implemented a randomized process that generates
volume data randomly containing cuboids and
cylinders. In order to make it difficult for the neural network to learn these shapes, random noise
was added.
Figure 5 shows the output of the network for
artificial data that includes a cuboid (a) and real
data with a cylindrical hole (b). Rows in Figure 5
represent individual slices of the volumes. The
columns show the input to the neural network
and output maps for the three different classes:
"background", "cuboid", and "cylinder". Brightness in the output maps represents the probability of individual voxels to belong to the respective
classes. For example, white pixels in the second
column represent a high probability for this voxel to belong to the background. In Figure 5(b) the
brightness of most pixels in the "cylinder" mask
is significantly higher than in the "cuboid" mask.
This indicates that the U-Net correctly classifies
the respective voxels to be part of a cylinder.
While our current implementation focuses on
differing between "cuboid"-shaped and "cylinder"-shaped defects, it is relatively straight-forward to adapt to other defect types. To do so, the
neural network only needs to be modified with respect to the number of output classes. The training process itself is very much the same, no matter which defects need to be detected. If available,
also real annotated data can be used for training.
This makes the method quite flexible and adaptable to varying requirements and scenarios.

we will evaluate more scan data to determine the
limits when using artificial training data for neural
network training.
ACKNOWLEDGMENTS
Work presented in this paper has received funding
from the Clean Sky 2 Joint Undertaking (JU) under
grant agreement No 831830. The JU receives support from the European Union's Horizon 2020 research and innovation programme and the Clean
Sky 2 JU members other than the Union. Work
presented in this paper has also received funding
by the European Union in cooperation with the
State of Upper Austria within the project "Investition in Wachstum und Beschäftigung" (IWB).
REFERENCES
1. M. Okulla, U. Düfert, A. Bulavinov, R. Pinchuk, Fortschrittliche Prüfmethoden zur Prüfung von CFK-Großkomponenten mit komplexer Geometrie, Jahrestagung der
Deutschen Gesselschaft für zerstörungsfreie Prüfung,
2017.
2. E. Galceran and M. Carreras, A survey on coverage
path planning for robotics, Robotics and Autonomous
Systems, pp. 1258-1276, 2013.
3. D. Kaljaca and B. A. Vroegindeweij, E. J. van Henten,
Coverage trajectory planning for a bush trimming robot
arm, J. Field Robotics, vol. 37, pp. 283-308, March 2020.
4. N. Atkar, H. Choset, A. Rizzi, E. Acar, Exact Cellular
Decomposition of Closed Orientable Surfaces Embedded in R3, Proceedings of the 2001 IEEE International
Conference on Robotics and Automation, May, 2001.
5. O. Ronneberger, P. Fischer, and T. Brox, U-Net: Convolutional networks for biomedical image segmentation,
International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234-241,
Springer, 2015.
6. M. Pedersen, S. Bengtson, R. Gade, N. Madsen, T.B.
Moeslund: Camera Calibration for Underwater 3D Reconstruction Based on Ray Tracing Using Snell's Law.
CVPR Workshops, pp. 1410-1417, 2018.

CONCLUSIONS AND FUTURE WORK
We outlined an approach for a highly automated
robotic workcell for ultrasonic inspection of complex CFRP parts. Our approach includes a novel
path planning algorithm. We use a deep neural
network for defect detection and classification of
ultrasonic test data of CFRP parts.
In future work we plan to improve the complete workflow by iterative refinement of the different algorithms involved. For coverage planning,
we will work with increasingly complex shapes.
With respect to defect detection and classification

SAMPE JOURNAL

7. G. Paul, N. Kwob, D. Liu: A novel surface segmentation
approach for robotic manipulator-based maintenance
operation planning, ScienceDirect 2013.
8. A. Sucan, Mark Moll, Lydia E. Kavraki, The Open Motion Planning Library, IEEE Robotics & Automation Magazine, 19(4):72-82, December 2012.
9. Ö. Cicek, A. Abdulkadir, S. Lienkamp, T. Brox, O. Ronneberger: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation, International Conference on Medical Image Computing and Computer
Assisted Intervention (MICCAI), 2016.

N OV E M B E R / D E C E M B E R 2 0 2 0

w w w. s a m p e . o r g

http://www.sampe.org

SAMPE Journal - November/December 2020

Table of Contents for the Digital Edition of SAMPE Journal - November/December 2020