Chesterman, X. (2024). Condition monitoring and failure diagnosis for wind turbines in the context of rare failing and slowly progressing damage. [PhD Thesis, Vrije Universiteit Brussel].

@phdthesis{4cd666346c054efc9ea67bc42373f546,
title = "Condition monitoring and failure diagnosis for wind turbines in the context of rare failing and slowly progressing damage",
abstract = "It is a well-known fact that the current generation of operational wind turbines suffers from premature component failures. These result often in long downtimes. If it were possible to identify these failures sufficiently in advance, it would be possible to replace or repair the failing components during regular maintenance, which would result in significantly lower maintenance costs and shorter downtimes.Failure prediction and diagnosis for wind turbines is currently an unsolved question. A useful system should be able to detect many different failure types well in advance. This means it should not just be able to identify the moment a component starts behaving out of the ordinary, but it should also be able to interpret the abnormal behavior.The improved availability of data has been a blessing and a curse. On the one hand, it has made it possible to analyze the behavior of the wind turbines thoroughly. On theother hand, the large quantity of data has made the task of analyzing and understanding it challenging for experts. An automated approach would solve this problem. Developing such a methodology is currently the topic of state-of-the-art research.The main goal of the research presented in this thesis is the development of an automated failure detection and diagnosis framework for wind turbine drivetrains. This is done using data that is readily available from wind farms, e.g. 10-minute Supervisory Control And Data Acquisition (SCADA) and status log data. The framework must be able to predict failures of wind turbine drivetrains, e.g. gearbox, generator, ..., well in advance by analyzing component temperatures. It must also be able to determine the failure mode by analyzing the patterns in the identified abnormal behavior.The framework is a 4-step pipeline consisting of several artificial intelligence (AI) techniques that are used sequentially. The first step uses machine learning (ML) techniques to model the normal component temperature behavior of wind turbines given their operational state. In the second step, statistical models, i.e. CUSUM, MAD-IOD are used to identify anomalies in the difference between the observed and predicted normal component temperatures. This information is combined in the third step with information from the status logs. Pattern and association rule mining techniques are used to identify patterns or rules that can be associated with drivetrain errors or failures. The patterns or rules are after postprocessing used for failure diagnosis.The validation, which is done on data from three operational wind farms, shows that the best-performing failure prediction methodology can detect failures accurately and well in advance. It is now used periodically to monitor an operational wind farm at the specific request of the wind turbine operator. The best failure diagnosis methodology succeeds in identifying patterns that are related to certain failure modes. These patterns can be used to make an accurate failure diagnosis.The thesis is structured as follows. It starts with an introduction that describes the general problem, the problem statement, and research questions. The first chapter ofthe thesis gives a description of the data characteristics, the failure properties, and two wind turbine ontologies. The second chapter discusses the failure prediction methodology. Several different methodologies are presented and compared. The third chapter focuses on the failure diagnosis. Two methodologies are discussed. The thesis ends with a general conclusion.",
author = "Xavier Chesterman",
year = "2024",
language = "English",
school = "Vrije Universiteit Brussel",
}