Maintenance management based on Machine Learning and nonlinear features in wind turbines

Abstract

Delamination is a common problem in wind turbine blades, creating stress concentration areas that can lead to the partial or complete rupture of the blade. This paper presents a novel delamination classification approach for reliability monitoring systems in wind turbine blades. It is based on the feature extraction of a nonlinear autoregressive with exogenous input system (NARX) and linear auto-regressive model (AR). A novelty in this paper is NARX as a Feature Extraction method for wind turbine blade delamination classification. Further, the NARX feature is demonstrated to be significantly better than linear AR feature for blade damage detection, and NARX can describe the inherent nonlinearity of blade delamination correctly. A real case study considers different levels of delamination employing ultrasonic guided waves that are sensitive to delamination. Firstly, the signals obtained are filtered and de-noised by wavelet transforms. Then, the features of the signal are extracted by NARX, and the number of features is selected considering the Neighbourhood Component Analysis as main novelties. Finally, six scenarios with different delamination sizes have been performed by supervised Machine Learning methods: Decision Trees, Discriminant Analysis, Quadratic Support Vector Machines, Nearest Neighbours and Ensemble Classification.