Application of an unsupervised deep learning framework for acoustic emission-based characterization of delamination process in composite laminates
Artykuł w czasopiśmie
MNiSW
200
Lista 2024
| Status: | |
| Autorzy: | Rzeczkowski Jakub |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 276 |
| Numer artykułu: | 111507 |
| Strony: | 1 - 18 |
| Impact Factor: | 9,8 |
| Web of Science® Times Cited: | 0 |
| Scopus® Cytowania: | 0 |
| Bazy: | Web of Science | Scopus |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | NIE |
| Abstrakty: | angielski |
| This study presents an advanced unsupervised deep learning framework for characterization of delamination processes in composite laminates based on acoustic emission (AE) measurements. The AE signal descriptors acquired during a double cantilever beam test were processed through a multi-stage analytical pipeline integrating stacked autoencoder for nonlinear feature extraction, uniform manifold approximation and projection (UMAP) for low-dimensional embedding and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) for unsupervised acoustic emission signals classification. This comprehensive approach enabled effective segregation of heterogeneous AE events into distinct clusters associated with specific damage mechanisms occurring during delamination process. The clustering outcomes were further validated through complementary time-frequency analysis by using continuous wavelet transform (CWT). In addition, a scanning electron microscopy fractographic observations of real delamination surfaces were also conducted. The proposed framework facilitated the differentiation of AE signal groups that may be associated with typical damage mechanisms, including matrix cracking, interfacial debonding with fiber pull-out and delamination. By removing the reliance on manual feature engineering and labeled datasets, this methodology provides a fully data-driven tool for interpretation of complex acoustic emission data. Furthermore, a prototype software application was developed to enable real-time processing, clustering and visualization of AE signals during experimental testing. The originality of this work lies in the integration of deep representation learning, nonlinear manifold embedding and density-based clustering into a coherent unsupervised analytical framework enabling efficient clustering of nonlinear acoustic emission data acquired during experimental testing of composite laminates. |
