Abstract Summary
The problem of random vibration response-based damage detection for a set of 30 composite aerostructures under varying operating conditions and uncertainty is experimentally investigated. The employed aerostructures are 14ply carbon fiber C-shaped laminates, 860 mm long and rectangular (25x82 mm) cross-section. 20 aerostructures are in pristine condition, while 6 are subjected to square-shaped delamination of three distinct sizes (20, 40, 50 mm) and another 4 to impact-induced damage under three distinct energy levels (12.9, 17.9, 23.1 Joules). The experimental set-up involves rigid mounting of each aerostructure at one end, and (non-measurable) random force excitation at two distinct locations with the acceleration vibration response being measured at distinct locations. The excitation forces are random and of varying (per experiment) spectral characteristics simulating those potentially acting in flight, while temperature varies within the 0 - 25 oC range. Preliminary analysis of the measured vibration response signals obtained from the healthy set of pristine structures within the range of 0 – 2 kHz indicates rich dynamical information with strong uncertainty among the structures due to variability in manufacturing (inter-structural uncertainty) and to variability in the experimental set-up (mounting assembly and small discrepancies in the sensor and excitation locations). In fact, the uncertainty induced on the dynamics is further compounded by those of varying temperature and excitation profiles and is so strong as to significantly “mask” the effects of the considered damages, thus leading to a highly challenging damage detection problem. Damage detection is tackled via two unsupervised and robust data-driven vibration-response-only methods which are based on the Multiple Input Multiple Output (MISO) Transmittance Function in order to properly compensate for the effects of the varying excitation: The first is based on the Multiple Model (MM) and the second on the HyperSphere (HS) based novel frameworks. Each one builds a distinct approximation of the healthy subspace within a properly selected feature space, which is hereby defined as the space spanned by Principal Component Analysis (PCA) reduced and transformed parameter vectors of stochastic MISO AutoRegressive with eXogenous (ARX) excitation representations of the Transmittance Function estimated from the measured vibration signals. The methods are trained using vibration signals from 16 healthy structures under various excitation profiles and 6 distinct temperatures (Baseline Phase), while the performance is assessed with signals from the remaining 4 healthy and 10 damaged structures under different, from those used in the training, excitation profiles and temperatures (Inspection Phase). Damage detection assessment is based on thousands of Inspection Phase experiments with the results presented in terms of Receiver Operating Characteristic (ROC) curves. The two methods are shown to be capable of overcoming the challenges of the varying operating conditions and uncertainty, achieving impressive detection performance, characterized by a 97.5% correct detection rate for 5% false alarm rate, with the MM based method exhibiting a slight edge over its HS counterpart.
