Measuring the Effectiveness of Descriptors to Indicate Faults in Wind Turbines

Over the years, numerous descriptors have been proposed to indicate the health and various failure modes of wind turbine components. Description and mathematical interpretation of descriptors, such as kurtosis, crest factor, peak to peak, mesh, and residual RMS, have been shown in several works to detect bearing and gear faults. However, there are only a few works that evaluate the effectiveness of various descriptors in detecting faults. This paper proposes Descriptor Ranking, a novel machine learning application that measures descriptors’ effectiveness to indicate different failure modes accurately. The proposed application uses different feature selection algorithms to rank the importance of the descriptors. Results show that Descriptor Ranking can correctly rank the importance of any calculable descriptors. Also, we evaluate the performance of different feature selection algorithms, such as random forest regressor and logistic regression. Given a balanced set of vibration data with apparent faults and no faults, Descriptor Ranking visualizes different descriptors’ effectifeness ranks, including newly proposed descriptors, in detecting different failure modes. Knowing the effectiveness of a descriptor in indicating specific failure modes is essential to optimize the condition monitoring strategy of wind turbines or any other machinery. Having multiple descriptors to identify the same failure mode can be avoided. Hence, we can reduce false alarms, and faults can be detected more accurately. With fewer false alarms, monitoring the condition of many wind turbines is going to be more efficient.