Modeling and Prediction of Gearbox Faults With Data-Mining Algorithms

Abstract

A data-driven approach for analyzing faults in wind turbine gearbox is developed and tested. More specifically, faults in a ring gear are predicted in advance. Time-domain statistical metrics, such as jerk, root mean square (RMS), crest factor (CF), and kurtosis, are investigated to identify faulty components of a wind turbine. The components identified are validated with the fast Fourier transformation (FFT) of vibration data. Fifty neural networks (NNs) with different parameter settings are trained to obtain the best performing model. Models based on original vibration data, and transformed jerk data are constructed. The jerk model based on multisensor data outperforms the other models and therefore is used for testing and validation of previously unseen data. Short-term predictions of up to 15 time intervals, each representing 0.1 s, are performed. The prediction accuracy varies from 91.68% to 94.78%.