PSI - Issue 64

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

1307

5

Fig. 4: Geometry of the foam (shown without metal sheets) with hemisphere fault geometry with randomized size and location.

The parametric model serves as the foundation for a simulation campaign, that generates simulation datasets including results for both fault-free models and models with faults. For this, a simulation loop is used, in which a Matlab ® Code iteratively changes an Ansys ® MAPDL Code and solves the model. In this way, material properties as well as geometric parameters can be changed in every iteration. After each iteration, the results of a modal analysis in the form of eigenfrequencies and the results of a harmonic analysis in the form of frequency response functions are used to create a new sample of the dataset. The addressed frequency range for the simulation data is chosen from 0 to 1000 Hz. In this range, the 30 first eigenfrequencies and eigenvectors are enough to describe the system dynamics using modal superposition, which is much more computationally efficient than simulating the full model. In the harmonic analysis, the FE model is excited at the corner of substructure one to three (to prevent excitation in a vibration node) with a unit force representing a modal hammer excitation at different points. The acceleration response is evaluated at all center points of each substructure and additionally at the three excitation points themselves. Datasets of different sizes are generated using the described simulation interface, whereby 200 samples for 18 transfer functions and the first 40 natural frequencies are visualized in Fig. 5. No fundamental differences between the data corresponding to the cases without and with defects can be observed, which underlines the need for a machine learning approach. Furthermore, labels are assigned to the simulation data.

(a)

(b)

Fig. 5: Frequency response functions (a) and eigenfrequencies (b) for a dataset of 200 samples without and with faults.

The labels for detection are 0 (no fault) and 1 (fault). From the perspective of the sandwich panel production, a division into localization classes in one spatial dimension appears to be suitable for practical application. Whole strips rather than cubes would be cut out of the panels in production if a defect occurs. Therefore, the structure is divided