Abstract
The problem of random vibration based robust damage detection on operating wind turbine blade under variable natural excitation conditions is considered via multiple accelerometer sensors and three crack type damage scenarios. For this purpose four Hyper-Sphere Healthy Subspace type damage detection methods are postulated: Two based on Multiple-Input Single-Output Transmittance Function representations and referred to as Class I methods and two based on Multiple-Output representations and referred to as Class II ones. A Principal Component Analysis enhanced method and a non-enhanced counterpart are included in each class. The results, based on a systematic assessment via 3 000 Test Cases and training with only 100 healthy signal sets, indicate that four sensors are adequate for achieving excellent detection performance reaching almost 100% True Positive Rate (TPR) for just 1% False Positive Rate (FPR) for the 15 cm crack damage scenario and for 0% FPR for the 30 cm and 45 cm damage scenarios. The PCA-enhanced methods offer significantly improved performance over their non-enhanced counterparts, and the Class I methods surpass their Class II counterparts, with the PCA-enhanced Class I method being overall best. The drastic performance improvements achieved by over earlier studies are also illustrated.
