240 / 2025-06-15 16:20:19

A Clustering-Based Damage Identification and Localization Method for Composite Structures Using Multi-Physical Features of Ultrasonic Guided Waves

Structural Health Monitoring,Ultrasonic Guided Waves,Clustering Analysis,Nondestructive Testing,Multi-Physical Features

Special Sessions > SSL 11 New Developments in Guided Waves (GW) and Nondestructive Testing (NDT): Propagation, Design and Applications

Structural Health Monitoring (SHM) has become a vital technique for ensuring the safe operation of critical engineering structures. Among various SHM approaches, guided wave–based damage detection has attracted widespread attention due to its ability to capture early structural anomalies over long distances by extracting signal-based features from wave propagation paths. However, existing methods often rely on single damage-sensitive features, which are susceptible to noise and insufficient for describing complex signal variations. Meanwhile, deep learning–based approaches,although effective in feature learning, typically require large labeled datasets and suffer from limited interpretability in practical engineering applications.To address these limitations, this study proposes a two-stage clustering-based unsupervised damage identification method using multiple physically interpretable features.

Three types of damage-sensitive indicators are extracted from each guided wave propagation path: correlation coefficient, envelope energy ratio, and time of maximum difference. In the first stage, K-means clustering is employed on a path level to distinguish damaged from undamaged paths based on the correlation-energy feature space. In the second stage, a 36-dimensional time-of-flight vector is constructed using the identified damaged paths for each damage case. A hierarchical clustering algorithm is then applied to classify 28 damage cases into seven structural zones.

The proposed method achieves an identification accuracy of 85.7%, demonstrating strong performance in both localization and interpretability. Compared to traditional single-feature-based or deep learning approaches, the proposed technique offers advantages in requiring no labeled data, being computationally efficient, and providing robust physical interpretability, thus showing great potential for real-world SHM scenarios with limited data and high reliability requirements.