Transmissibility-based monitoring and combination of damage feature decisions within a holistic structural health monitoring framework

Abstract

Over the past forty years, intensive research has been carried out in the field of structural health monitoring (SHM), since the identification of damage at an early stage contributes to avoiding structural failure and reducing maintenance costs. In particular, the monitoring of wind turbines has gained special interest, since there is an increasing number of installed wind turbines worldwide and a large number of wind turbines which have reached or will soon reach their design lifespan. This thesis focuses on vibration-based SHM methods, which observe features describing the dynamics of a structure. Moreover, this work is based on the conception that the consideration of SHM should not only involve the observation of damage-sensitive features, but should also address further aspects, such as the effect of environmental and operational conditions (EOCs) and the statistical pattern recognition approaches used for decision making. Wind turbines are complex structures which operate in a challenging environment. Most of the vibration-based approaches rely on assumptions which are violated, for example, during the operation of a wind turbine, raising doubts concerning their accuracy. Furthermore, there is a plethora of damage-sensitive features, alternatively called condition parameters (CPs), which can be used to assess the state of a structure. However, up to the present moment, little research has been conducted on the combination of damage feature selected and on the exploitation of decision making processes for improving the detection rates of damage when it exists. This work introduces a new vibration-based CP, which does not rely on any significant assumptions. The new CP is based on an output-only version of an autoregressive model with exogenous input (ARX), which is essentially a transmissibility function (TF) model. The poles of the model are therefore called TF poles. The proposed CP is based on the observation of TF pole migration due to structural changes. Several experimental datasets are used to explore the sensitivity of TF poles to damage, while the concept of implementing TF poles as a CP in unsupervised mode is described. The new CP is integrated into a three-tier SHM framework which performs data normalizaton (tier 1), extracts the CP for analysis (tier 2) and subsequently makes use of hypothesis testing (tier 3). This framework using TF poles is validated on the fatigue test data of a full-scale rotor blade. This work also proposes the implementation of adaptive boosting (AdaBoost) for the combination of decisions obtained from several damage features in order to attain a new and more accurate decision rule. The proposed concept is integrated into the aforementioned three-tier SHM framework and is used for combining the decisions of vibration-based damage features. However, the proposed concept can be implemented after any SHM process, even if other SHM approaches are employed. The concept of implementing Adaboost within the three-tier SHM framework is outlined and validated on the data of an operating 3~kW wind turbine. Finally, different damage features, including the proposed CP, are compared with respect to their sensivity to damage and sensitivity to EOC variability based on rotor blade fatigue tests.