PSI - Issue 37

Tomasz Rogala et al. / Procedia Structural Integrity 37 (2022) 187–194 / Structural Integrity Procedia 00 (2019) 000 – 000

193

7

experiments were carried out employing MATLAB®R2019b – Deep neural network toolbox (MathWorks, Natick, MA, USA) installed on the personal computer station with Intel®Core™i7 -3930K CPU @ 3.20 GHz, 64 GB RAM, 512 GB SSD, 2 TB HDD, NVIDIA™RTX 2080 equipped with 8 G B RAM.

4.2. Results of classification Training, test and verification experiments were executed ten times in order to obtain statistical metrics. As it was mentioned above the additional set of verification patterns was applied in order to obtain quality indices such as true positive rate TPR (also called sensitivity or recall), positive predictive value PPV (also called precision) and accuracy ACC which were derived from a confusion matrix. The mean value and standard deviation for the selected measures were calculated. The outcomes were included in Table 1. It can be seen that the highest average accuracy with the best repeatability was obtained for the case in which the input pattern was based on 3D features. On the other hand, by using the whole set of features it was possible to minimize the dispersion of TPR and PPV results for crack and delamination damage, respectively. The lowest average accuracy with high value of the standard deviation was achieved based on 2D features. Such low value of accuracy could be mainly caused by the fact that normalized and transformed 2D features did not include relevant information needed to recognize delamination damage. Table 1: Summarized results of classification trials for the additional set of verification patterns

Damage class

TPR

PPV

ACC

2D features 3D features

Crack

0.96 ± 0.006 0.43 ± 0.045 0.98 ± 0.004 0.77 ± 0.032 0.98 ± 0.003 0.75 ± 0.056

0.80 ± 0.050 0.82 ± 0.037 0.95 ± 0.010 0.92 ± 0.021 0.94 ± 0.017 0.92 ± 0.018

0.80 ± 0.037

Delamination

Crack

0.94 ± 0.007

Delamination

2D and 3D features

Crack

0.94 ± 0.013

Delamination

The overall classification results presented in Table 1 show the high practical potential of the deep neural networks for automatic recognition of damage types in polymeric composites. 5. Conclusions This paper shows some partial results concerned with the classification problem of damage objects in XCT scans. The main objective was to explore the applicability of classifiers to recognize the type of damage in CT scans. The recognition was performed based on selected features of objects observed in tomographic scans. It was initially proved that crack objects due to fatigue dominated by the self-heating effect in computed tomography scans of polymeric composites can be automatically classified using feature-based descriptors and deep learning techniques with convolutional neural networks. An elaborated tool, in the form of the prepared network can be helpful as the decision supporting tool for NDT inspectors. A novel approach with the transformation of 3D volumetric objects into flat images was also proposed in this research. The results of classification of cracks indicate that these 2D features are also pretty informative and relevant- in this case for crack objects. The selected features are invariant to at least rotation, scale, or translation. References Bull, D.J., Helfen, L., Sinclair, I., Spearing, S.M., Baumbach, T., 2013. AComparison of Multi-Scale 3D X-ray Tomographic Inspection Techniques for Assessing Carbon Fibre Composite Impact Damage. Composites Science and Technology 75, 55. Cox M.,Budhu. M., 2008, A Practical Approach to Grain Shape Quantification. Engineering Geology 96, pp. 1–16 Fang, Q. Boas, D., 2009, Tetrahedral mesh generation from volumetric binary and gray-scale images, Proceedings of IEEE International Symposium on Biomedical Imaging pp. 1142-1145.