Dynamic Mode Decomposition for Single-Frequency Ultrasonic Wrinkle Characterization in Aerospace Composites

Abstract

Ultrasonic non-destructive inspection (NDI) is the dominant method for examining the interior of carbon fiber reinforced polymer (CFRP) structures in aerospace manufacturing. While low-frequency ultrasound adequately characterizes most defect types, out-of-plane wrinkles remain a persistent challenge: lacking the strong acoustic-impedance contrast of other defects, they conventionally require a costly secondary scan at a higher frequency to resolve individual plyinterfaces. This thesis investigates the feasibility of characterizing wrinkle defect lengths directly from low-frequency (3.5 MHz) pulse-echo ultrasound, eliminating the secondary high-frequency pass. The proposed approach introduces dynamic mode decomposition (DMD) as a data-driven feature extractor for ultrasonic scan volumes. Through a multi-resolution DMD pipeline incorporating time-delay embedding, polar-coordinate eigenvalue filtering, and physics-informed K-means clustering, the dominant eigenvalue’s angle on the unit circle is shown to map directly to a wrinkle’s spatial wavelength through a simple geometric relation. This yields an interpretable, physics-based estimate of wrinkle geometry without resolving individual ply layers, in contrast to deep-learning approaches whose outputs are class labels or segmentation masks rather than interpretable physical quantities. The pipeline is validated on synthetic datasets of progressively increasing realism and on a limited experimental dataset of nine scrapped CFRP samples scanned at 3.5 MHz and 5 MHz with destructive caliper ground truth. The method produces wrinkle-length predictions thatcorrelate with measured values at both frequencies (R-squared 0.50 and 0.41), with comparable performance between the high- and low-frequency scans, suggesting that the secondary scan may be avoidable; the available high-frequency reference here was 5 MHz rather than the 7.5 MHz of the production secondary pass. These results are constrained by the small sample size and the manual tuning of DMD hyperparameters, and full in-process characterization would require a larger dataset and further refinement of the pipeline. The contribution of this work is therefore an application: an interpretable deployment of established DMD machinery on a costly, manual production-inspection problem, together with feasibility evidence that motivates the next phase of validation.