PSI - Issue 80

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

404

2

engineering structures, particularly those composed of advanced composite materials such as Carbon Fiber Reinforced Polymers (CFRPs), which are extensively used in aerospace, automotive, and civil engineering applications (Güemes et al., 2020; Hassani et al., 2021). Among the various SHM techniques, Ultrasonic Guided Waves (UGWs) have been recognized for their capability to inspect large structural areas with relatively few sensors, offering a cost-effective and minimally intrusive solution for early damage detection (Du et al., 2022; Wu et al., 2021). UGWs propagate along structural components and exhibit sensitivity to defects such as delamination, impact damage, and fibre breakage, making them suitable for continuous monitoring of composite panels (A. De Luca et al., 2023; Perfetto et al., 2024). Despite these advantages, the interpretation of UGW signals remains challenging, especially in complex anisotropic composite structures where wave modes exhibit high dispersion and are influenced by noise and environmental variability (A. D. De Luca et al., 2018; Yang et al., 2023). Traditional signal processing methods often rely on expert knowledge and manual feature extraction, which limits scalability and robustness in practical applications (Ferri Aliabadi & Khodaei, 2017). Consequently, data-driven approaches leveraging machine learning have gained increasing attention to automate damage detection and classification tasks (Luo et al., 2024; Perfetto et al., 2023; Rezazadeh et al., 2024; Spencer et al., 2025). Unsupervised clustering methods have emerged as effective tools in SHM, as they do not require extensive labelled datasets, which are often difficult and costly to obtain in structural monitoring scenarios (Boratto et al., 2025; Lomazzi et al., 2024). These methods enable the identification of intrinsic patterns and the separation of healthy and damaged states based on statistical properties of extracted features, facilitating anomaly detection and condition assessment without prior knowledge of damage types or locations (Nerlikar et al., 2024; Sattarifar & Nestorović, 2022) . However, the success of unsupervised learning critically depends on the selection of informative features and the clustering algorithms’ ability to handle high -dimensional and noisy data typical of UGW signals (García-Macías & Ubertini, 2022; Liu et al., 2024). Recent investigations have applied unsupervised clustering to UGW-based SHM with encouraging results but also reveal several limitations. For example, k-means clustering applied to Lamb wave features demonstrated impact damage detection in CFRP panels, yet required extensive feature engineering and showed sensitivity to environmental variations (Ikotun et al., 2023; Zhu et al., 2025). Hierarchical clustering combined with principal component analysis (PCA) has been used for damage classification, though consistent separation of subtle damage states remains challenging due to feature overlap (El Mountassir et al., 2021; Yue & Aliabadi, 2021). Moreover, many studies focus on single-feature or pairwise feature analyses, potentially overlooking complex interactions among multiple features that could enhance classification performance (Qiu et al., 2019; Yuan et al., 2019). These observations indicate the need for systematic exploration of feature combinations and rigorous evaluation of clustering performance using robust metrics such as silhouette score, purity, and cluster balance. Additionally, lightweight and interpretable methodologies are essential for practical SHM systems requiring real-time processing and seamless integration with existing sensor networks. The present study proposes a data-driven methodology combining UGWs with unsupervised clustering to classify low-velocity impact damage in CFRP panels. A comprehensive set of scalar features is extracted from UGW signals acquired in both pristine and damaged conditions, and all possible pairwise and triplet feature combinations are explored to identify those that maximize damage separability. This approach aims to provide a scalable, interpretable, and label-free diagnostic tool that advances the state of the art in SHM of composite structures.

Nomenclature CFRP Carbon Fiber Reinforced Polymer GMM Gaussian Mixture Model SHM Structural Health Monitoring PZT Lead Zirconate Titanite UGW Ultrasonic Guided Waves

2. Methodology The experimental campaign was carried out on ten CFRP panels, illustrated in Figure 1a. Each panel measured