Pitch system failure identification using a combination of subject matter expert knowledge of offshore wind turbines and machine learning techniques.
,2, Mark Spring1, Peter Davies1, Jonathan Shek2, Philipp Thies2, Erkan Oterkus2
1Lloyd's Register, Aberdeen, UK, 2Industrial Doctorate Centre for Offshore Renewable Energy Systems, IDCORE, Edinburgh, UK
Reducing the levelised cost of energy (LCoE) by 40% will be tough to achieve for the offshore wind energy industry (ref [ 1]). A key proportion of LCoE are the costs of operation and Maintenance (O&M) activities, at up to 30% (ref ). Automatic and intelligent systems are needed to minimise human intervention during the operating life.
A comprehensive Failure Modes and Effects Analysis (FMEA) of a generic offshore wind turbine, including all assemblies and sub-assemblies has been performed (see ref ). The pitch system has been identified as one of the most critical assemblies in term of turbine operation. For the pitch system the limited number of signals available through SCADA hampers the identification of failure causes, the development of physics-based approaches to quantify degradation, estimate risk and hence schedule maintenance tasks. It is clear however that health of the pitch system may be discerned from the available data. The challenge is to identify how to combine together existing signals. Therefore, machine learning and data mining methodologies have been used to understand the pitch system normal behaviour. Observed deviations in SCADA data can be categorised as positive or negative in terms of the deduced risk profile. Consequently, critical modes of failure of the pitch system have been anticipated in advance.
The proposed method involves an unsupervised Support Vector Machine (SVM), used for novelty detection. Given a set of samples, the SVM detects the soft boundary of that set. In two dimensions, the soft boundary can be displayed as a contour. A decision tree is used to take into account expert knowledge of offshore wind turbine operation. The classification technique used is the K-Nearest Neighbour (KNN). KNN enables new SCADA data observations to be categorised and an operation status to be assigned. The model searches for a number of observations from the training data and then calculates the numerical distance between the unknown "status" of the new observation and the training data and, selects known status.
The main result is the identification of outliers that might represent an incipient failure in the pitch system. The normal behaviour and frontier are defined using offshore wind turbine parameters such as blade position, wind speed, wind turbulence, power output, hydraulic pressure of the pitch system. The alarm is a green, yellow or red flag for every new observation status.
This methodology is less computationally demanding and will allow pitch system anomalies to be identified. More SCADA database variables can then be included in the analysis of anomalies to diagnose the failure mode and cause. Thus the most efficient O&M tasks can be selected to reduce turbine downtime and associated costs.