Operational Variables for Improving Industrial Wind Turbine Yaw Misalignment Early Fault Detection Capabilities Using Data-Driven Techniques

Offshore wind turbines are complex pieces of engineering and are, generally, exposed to harsh environmental conditions that are making them to susceptible unexpected and potentially catastrophic damage. This results in significant downtime and high maintenance costs. Therefore, early detection of major failures is important to improve availability, boost power production, and reduce maintenance costs. This article proposes a supervisory control and data acquisition (SCADA) data-based Gaussian process (GP) (a data-driven, machine learning approach) fault detection algorithm where additional model inputs called operational variables (pitch angle and rotor speed) are used. First, comparative studies of these operational variables are carried out to establish whether the parameter leads to improved early fault detection capability; it is then used to construct an improved GP fault detection algorithm. The developed model is then validated against existing methods in terms of capability to detect in advance (and by how much) signs of failure with a low false positive rate. Failure due to yaw misalignment results in significant downtime and a reduction in power production was found to be a useful case study to demonstrate the effectiveness of the proposed algorithms. Historical SCADA 10-min data obtained from pitch-regulated turbines were used for models training and validation purposes. Results show that: 1) additional model inputs were able to improve the accuracy of GP power curve models with rotor speed responsible for a significant improvement in performance and 2) inclusion of rotor speed enhanced early failure detection without any false positives, in contrast to the other methods investigated.