PSI - Issue 80

Alessandro De Luca et al. / Procedia Structural Integrity 80 (2026) 403–410 D Alessandro De Luca / Structural Integrity Procedia 00 (2019) 000 – 000

405

3

100 × 150 mm 2 in- plane, with a thickness of 2.7 mm, and a stacking sequence of [45/0/−45/90] s . A network of six Lead Zirconate Titanate (PZT) piezoelectric transducers was surface-bonded in a 50 × 60 mm 2 rectangular configuration (Figure 1) with transducers 1, 3, 4 and 6 positioned at the corners, and transducers 2 and 5 at the midpoints of the shorter sides. This arrangement maximized wave propagation coverage and directional diversity, enabling multiple actuator-receiver paths with varying propagation distances and orientations. Signal generation and acquisition were performed using two parallel-connected TiePie Engineering oscilloscopes with integrated arbitrary waveform generator. A five-cycle Hanning-windowed tone burst served as excitation signal, with four central frequencies tested: 150, 200, 250, and 300 kHz. Data acquisition occurred at 100 MHz sampling frequency within a 200 µs time window. For each panel, in both pristine and post-impact conditions, every transducer served sequentially as an actuator while the remaining three transducers positioned on the opposite side of the network functioned as receivers. This yielded 18 unique actuator – receiver combinations per panel condition (i.e., undamaged, damaged). All actuator-receiver pairs were then categorized into four path types based on propagation direction and length (Figure 1b). Each path category includes both directions of wave propagation (each transducer acts as both actuator and receiver for a given path): • W path (Blue): diagonal paths bypassing the panel centre (1-5, 3-5, 2-4, 2-6) • X path (Green): diagonal paths intersecting the panel centre (1-6, 3-4) • I path (Red): vertical path crossing the panel centre (2-5) • H path (Purple): vertical paths not crossing the panel centre (1-4, 3-6) All panels underwent a controlled 15J low-velocity impact at their geometric centre using a drop-weight impact tower, in accordance with ASTM D7136 (ASTM D7136/D7136M-15, 2015).

a

b

Figure 1. (a) CFRP panels used in the experimental study; (b) sensor path categories.

A set of 32 scalar features, listed in Table 1, was extracted from each recorded signal to capture its main characteristics in a compact and interpretable form. To evaluate the effectiveness of feature combinations for damage classification, all unique pairs of features were considered as input for unsupervised clustering using a Gaussian Mixture Model (GMM) algorithm. Since the order of features in a pair does not affect the clustering input (i.e., the pair {feature A, feature B} is equivalent to {feature B, feature A}), duplicate pairs differing only by order were excluded.