Abstract Summary
The current need for sustained renewable energy is pushing the boundaries of the size and availability of wind turbines. As a result, there is an urge to develop a well-functioning Structural Health Monitoring (SHM) systems to keep these systems online. This need is also shared with other pieces of critical infrastructure and structures in general. Among different issues, the effect Environmental and Operational Variability (EOV) in Damage Sensitivity Features (DSF) (and in the dynamics of the structure in general) is probably still one of the most relevant challenges facing SHM. While numerous methods have been developed to tackle this issue, these may be designated either as implicit or explicit. The first class of methods use characteristics of the features themselves, while in the latter, information from environmental and operational conditions is also introduced. In paper, explicit methods bear the advantage, as these directly correlate the information of the variables affecting the DSFs, and may potentially lead to increased damage diagnosis sensitivity. On the other hand, implicit methods do not make assumptions on cause-effect relations and will not introduce a bias from wrongly-assumed correlations. In this work, we postulate that the combined application of both types of methods can lead to a more robust approach that will also lessen their individual shortcomings. More precisely, we combine Principal Component Analysis (PCA) with Bayesian non-linear regression for EOV compensation in damage detection of a wind turbine blades. We evaluate the methodology on data from a laboratory-scale wind turbine blade in controlled temperature and humidity conditions, as presented in [1]. On each vibration test, the temperature was set to a fixed value within the range [-20◦C,+20◦C]. Furthermore, the humidity was increased for some of the samples by spraying water onto the blade. Tests were performed on a reference blade, and two blades with different damage severity. DSFs originate from Vector Auto-Regressive (VAR) models, and include the VAR model coefficients and the respective natural frequency estimates. The raw DSFs are initially PCA-transformed, and, following the PCA method for EOV compensation, PCs with largest eigenvalues would be removed, as these bear most of the EOV influence. In our approach, we use Bayesian non-linear regression on all the PCs and use the F-statistic to assess which PCs accept the regression model and which not. The PCs rejecting the regression model would be left as they are, while the remainder are adjusted with the help of the obtained regression model. We compare the performance of the combined method, with that obtained using only PCA, and using the regression model only, in terms of Receiver Operating Characteristic (ROC) analysis. Results point at the combined method providing a more robust performance among the considered variants. [1] A. Gómez-González and S.D. Fassois, “A supervised vibration-based statistical methodology for damage detection under varying environmental conditions & its laboratory assessment with a scale wind turbine blade”, Journal of Sound and Vibration, 366, pp. 484-500, 2016.
