PSI - Issue 64

Hendrik Holzmann et al. / Procedia Structural Integrity 64 (2024) 1303–1310 Hendrik Holzmann / Structural Integrity Procedia 00 (2019) 000 – 000

1309

7

Fig. 7: Localization accuracy of test data over data amount.

The recognition rate of the test data is used as the metric to describe the quality of the data model. To visualize the classification quality, a softmax calculation ( )= exp⁡( ) ∑ exp⁡( ) (1) is used, that maps the classification outputs of the neural network of a data sample to the individual classes with a value from 0 to 1 where all probabilities add up to 1. This value is assigned to the individual substructures using a colour scale and the actual defect position (red dot) is visualized at the same time as shown in Fig. 8.

Fig. 8: Example localization results with softmax visualisation of probability. Example of a correct classification (a); correct classification with distributed probability (b); and a wrong classification (c)

It can be seen, that in general, very good results for the detection of faults are achieved which is the primary goal in the realistic application. The localization can be seen as a supplement to this and works with an accuracy of over 80 %. The implemented workflow can be easily applied to experimental data. 4. Conclusion In this work, a simulation workflow for detection and localization of defects in sandwich panels was presented. It comprises a parametric finite element model, a simulation interface, feature engineering, a neural network and an optimizer for hyperparameter tuning. In conclusion, the results obtained from the data model demonstrate high recognition rates for detection and localization of defects in sandwich panels. Noteworthy is the achievement of a detection rate of up to 94.8 % and a localization rate of up to 81.3 %. These findings underscore the efficacy of the proposed approach in achieving robust fault detection, which is the primary objective in practical applications. Moreover, the localization aspect, though secondary, also exhibits high accuracy, exceeding 80 %. The implemented workflow can be directly used in practical