SCADA Data Mining for Condition Monitoring of a Wind Turbine Gearbox

Abstract

Wind power is the fastest-growing energy source and the offshore wind industry is expected to grow by over 80 GW through 2024. There is a need to reduce all associated costs to be competitive in a market that might be fully subsidy-free soon. Adapting a predictive or condition-based maintenance strategy can help in reducing O&M costs, as it accounts for more than 30% of the lifetime costs in offshore turbines.This master thesis focuses on developing a novel gearbox condition monitoring technique using raw Supervisory Control And Data Acquisition (SCADA) data from an offshore wind farm. The developed approach attempts to leverage the use of physics of failure technique and artificial neural network models for continuous monitoring of gearbox condition. Utilizing SCADA data for monitoring and fault detection can help in eliminating retrofitting additional expensive sensors such as an online oil monitoring system, vibration monitoring systems and so on for the same purpose.Firstly, a physics of failure methodology is implemented to monitor turbines on a farm level. Using Failure Mode, Effects Analysis (FMEA), the operating regimes under which gears and bearings can potentially get damaged are ascertained. A correlation between the damage operating conditions and the root cause of failure is analyzed using boxplots. In particular, a correlation between operating conditions and, the unusual and unexplained White Etching Cracks or axial cracking failure that occur frequently in the high-speed shaft bearings is studied. Additionally, this methodology is also used to complement the Artificial Neural Network model developed for anomaly detection.Finally, an Artificial Neural Network (ANN) based on a Recurrent Neural Network (RNN) is developed for fault detection. The model is trained on data when the turbine was operating normally, to predict gear bearing and gearbox lubrication oil temperatures. A stacked Gated Recurrent Unit (GRU), a type of RNN model with two hidden layers resulted in predictions with the lowest mean squared error (0.14640퐶) and with good generalization. The difference/residual between the measured and ANN predicted temperatures are used to flag anomalies. The ANN-based condition monitoring method is validated through case studies using real data from wind turbines. The results from the case studies indicate that the RNNbased fault detection method can detect a failure in the wind turbine gearbox components as early as 3-6 months before it fails.