A clustering-based damage segmentation for ultrasonic C-Scans of CFRP plates

Despite their desirable mechanical properties, damage propagation in carbon fiber-reinforced polymers (CFRP) due to manufacturing flaws and continued use may particularly be hard to assess. In this work, damage maps are generated to identify the health state of a CFRP plate from ultrasonic signals obtained under C-Scan mode. This configuration allows us to visually inspect the effective state of the plate through the thickness. Firstly, signals are processed using an all-pole model with a sparse set of coefficients, which retains the most relevant information of each signal. Then, model coefficients are transformed to the cepstral domain in order to apply a unsupervised clustering procedure. From the resulting signal classification a visual map of the damage is generated. Five different clustering techniques are selected to this end and compared. As a result, clear and consistent maps of the damage pattern can be achieved when a underlying sparse model is exploited along with hierarchical and density-based clustering techniques.