Unsupervised learning and fusion for failure detection in wind turbines

Unforeseen failures of components of a wind turbine have a significant impact on the power generation capability. Anemometer, which measures the wind speed as seen by the turbine, can be used to predict within some degree of accuracy the amount of power the turbine should produce. It is a good indicator of incipient failures that the power production is consistently low. Unsupervised and automated monitoring of the dynamic behavior of a wind turbine and power produced can detect faults early. Thus, secondary defects and major breakdowns are avoided as well as minimizing the impact this failure has on the total production of the turbine. This paper presents a failure detection method relying on the wind speed differences between two turbines. We collect a week's worth of data and model the wind speed difference between two turbines using a Weibull distribution. Two original tests are proposed. The first test is based on the fitted Weibull distribution, where the abnormal weeks have higher values in both the scale and shape parameters. The second test evaluates the area under the weekly empirical cumulative density function. Both tests yield similar fault detection results, and the second test reveals the sequential progression. A fusion using the second test on multiple turbine pairs can clarify the ambiguity in the single turbine pair test.