A large, full‐scale (6 m tall, 2.5 m wide) full‐composite ship hull section was subjected to three consecutive impacts with an impact energy level (∼20 kJ) mimicking a realistic heavy crash. The structure was continuously monitored with acoustic emission (AE) sensors during the impacts, which allowed the possible degradation of the composite hull to be assessed. Unsupervised learning was applied to AE data features to enable the categorization of the damage accumulated during the consecutive impacts. The implemented unsupervised learning routine was a combination of automatic Laplacian data feature selection followed by density‐based spatial clustering of applications with noise (DBSCAN), for which the hyperparameters were automatically optimized with a silhouette‐driven approach. Three predominant damage mechanisms (sandwich‐core crushing/cracking, skin‐core debonding, and matrix cracking) were identified through the clustering solution (with 0.9 mean silhouette and 6% outliers) and verified by AE feature bands found in the literature. The AE‐based damage categorization was validated with impact slow‐motion videos and postimpact digital laser shearography inspections. The study highlights the capability of the developed categorization methodology to be deployed for online monitoring, where SHM algorithms are retrained during ship operation to keep improving diagnosis accuracy. The results show that in the online training scenario, where Impacts 1 and 2 were used for training and Impact 3 was used for testing, the categorization performance was comparable to when data from all three impacts had been used for training, with a mean silhouette of 0.884 and only 2% outliers. Altogether, the damage categorization routine demonstrated reliability and stability for handling realistic AE data variability and different data availability scenarios. This study is an important step toward a complete condition diagnosis (comprising detection, localization, categorization, and quantification) of full‐scale composite ship structures, which is crucial for estimating their remaining useful life in real time and thereby for enabling their condition‐based maintenance.
Ochôa et al. (Thu,) studied this question.