Abstract
The problem of vibration-based damage detection under varying environmental condi- tions and uncertainty is considered, and a novel, supervised, PCA-type statistical meth- odology is postulated. The methodology employs vibration data records from the healthy and damaged states of a structure under various environmental conditions.
Unlike stan- dard PCA-type methods in which a feature vector corresponding to the least important eigenvalues is formed in a single step, the postulated methodology uses supervised learning in which damaged-state data records are employed to sequentially form a feature vector by appending a transformed scalar element at a time under the condition that it optimally, among all remaining elements, improves damage detectability.
This leads to the formulation of feature vectors with optimized sensitivity to damage, and thus high damage detectability. Within this methodology three particular methods, two non- parametric and one parametric, are formulated. These are validated and comparatively assessed via a laboratory case study focusing on damage detection on a scale wind turbine blade under varying temperature and the potential presence of sprayed water.
Damage detection performance is shown to be excellent based on a single vibration response sensor and a limited frequency bandwidth.
Keywords:
Random vibration based SHM, Varying environmental conditions, Principal Component Analysis, Time series methods, Supervised learning, Wind turbine blades
Introduction
Structural damage diagnosis is of primary importance in mechanical, aerospace, and civil engineering. Vibration-based methods are based on the premise that damage causes changes in the structural dynamics which may be detected through measured vibration data records and inverse-type techniques.
This is accomplished by monitoring a characteristic quantity (or else feature vector), selected to best embed the dynamical information pertinent to damage, which is estimated based on the measured vibration data records.
Yet, an important limitation of vibration-based methods is that changes in the dynamics may be also caused by varying environmental conditions and uncertainty, thus potentially limiting the performance characteristics and reliability of damage diagnosis.
This is also true for typical statistical methods, which although accounting for certain types of uncertainty, they still fail to account for environmental conditions and uncertainties not employed during their training (baseline) phase. The problem of effectively coping with varying environmental conditions and uncertainty is thus important, and a current technology application barrier.
The general approach for overcoming it involves, in broad terms, proper “pre-pro-cessing” of the vibration data records or the selected characteristic quantity (feature vector), aiming at the “removal” of the effects of uncertainty. Such pre-processing may be distinguished into model-based or non-model based, and is also referred to as “data normalization” (see for an overview).
Preview Journal
Related Posts
Mau donasi lewat mana?
Donate with PaypalGopay-