Wind turbine condition monitoring feature extraction using the empirical wavelet transform
,2, Paul McKeever
2, Hyunjoo Lee2, Wenxian Yang1
1Newcastle University, Newcastle, UK, 2ORE Catapult, Blyth, UK
Reliable wind turbine (WT) condition monitoring (CM) relies heavily on the effectiveness of fault-related feature extraction from CM signals. However, due to slow rotating speeds, constantly varying loads and unpredictable operation conditions, fault-related feature extraction from lengthy, non-linear, non-stationary WT CM signals is extremely difficult. This makes WT CM one of the most challenging tasks in wind power asset maintenance, despite much effort that has been spent in this field of research and innovation to date. The recently developed Empirical Wavelet Transform (EWT) provides a potential breakthrough in dealing with non-linear, non-stationary CM signals. However, the conventional EWT adopts a default method to pre-define values for both the mode number and mode boundaries in the spectrum. It is not adaptive to the signals being inspected. As a consequence, it leads to inaccurate feature extraction and thus, unreliable WT CM results. For this reason, an improved EWT method is studied in this research to precisely extract features and overcome this shortcoming.
In the improved EWT, boundaries are initially determined by a ‘local minima of spectrum’ envelope with a pre-defined threshold. With detected boundaries, the signal is decomposed into a set of mono-components carrying key feature information. Time-frequency analysis is then applied to these components to extract modulated information.
The experiments have shown that thanks to the use of an optimisation algorithm, the fault-related features buried in WT CM signals have been extracted out successfully. With the aid of the proposed optimisation algorithm, the accuracy of the EWT has been significantly improved. For example, the impact features contained in the bearing vibration signals have been explicitly extracted out despite the considerable noise in the signals and the lack of any pre-knowledge about the CM signals.
The main contribution of this research focuses on the development of data-driven adaptive EWT and its application in WT CM feature extraction. The early results are very promising and warrant further research and analysis with more real wind turbine data.
This research will educate the audience and relevant stakeholders on how to use advanced signal processing techniques to perform feature extraction from WT CM signals. This feature extraction can be adopted to increase the reliability of WT CM prognoses used to determine the health of key WT components. This intelligence can then be used to optimise Operations & Maintenance strategies for current and future wind farms.