HEALTH MONITORING FOR WIND TURBINES – DATASETS PROCESSING AND DEVELOPMENT OF RUL PROGNOSTICS

The increasing availability of condition-monitoring data for components/systems has incentivized, in the past years, the development of data-driven Remaining Useful Life (RUL) prognostics algorithms for assets such as Wind Turbines. This Thesis aims to support, through a critical analysis, the choice of the best Open-Source Datasets for Wind Turbines for Prognostic and Health monitoring (PHM) and for predictive maintenance planning, and to develop a data-driven Machine Leaning model for RUL prognostics. The first part of this Thesis is divided into three sections. The first section provides an overview of the main characteristics of Open-Source Datasets compared by time span, sampling rate, number and type of parameters and type of components. The second section reviews the current research publications for the Open Source and Not Open-Source Datasets discussing objective of the study, models developed and relative performance. Finally, the last section analyses, according to the information gathered in the previous sections, the possible application for PHM and predictive maintenance of each Open-Source Dataset, evaluating the most complete and most useful dataset. Once the most suitable Open-Source dataset for Prognostics applications has been identified, a Long Short-Term Memory (LSTM) model with Monte Carlo Dropout has been implemented for developing probabilistic RUL prognostics of Wind Turbines. After having cleaned and pre-processed the data available, the most significant features, directly linked to the systems correct behaviour, have been extracted and used to feed the LSTM model introduced. Firstly, a LSTM algorithm with regular Dropout has been used to make point-RUL predictions. After having implemented a LSTM model, able to well predict the true RUL, the features, considered in the training set, with the highest impact on the model’s output have been identified. In this sense, to analyse the features interactions and effect on the model's prediction a game theoretic approach, Shapley additive explanations (SHAP), have been implemented to understand which component, for our case study, has more impact of the RUL of the entire Wind Turbine. The SHAP method shows as the most significant component, for our case study, is the generator. In the past years, most studies have been focusing only on point RUL prognostics and therefore this Thesis aims to provide a probabilistic RUL prognostics for Wind Turbines in order to visualize the evolution of systems’ RUL estimating the uncertainty associated. For this thesis, 4 different case study have been considered to evaluate the sharpness and reliability of the data-driven algorithm developed.