ML-based sensor reduction for condition monitoring of wind turbine pitch bearings

Wind turbines often exhibit component failures before completion of their typical 20-year design life. Unexpected part failures increase the associated costs of their Operations & Maintenance (O&M). In turn, this raises the associated Levelized Cost of Electricity (LCOE), making this method of power generation less competitive compared to traditional methods. One of the components of particular concern is the slew or "pitch" bearing connecting the root of the blades to the rotor hub. The Drivetrain Reliability Collaborative (DRC) of the National Renewable Energy Lab (NREL) has begun investigations into pitch bearing reliability. One outcome of this is a collection campaign on the 1.5MW Wind Turbine at the NREL Flatirons Campus to observe variations in pitch bearing strains during real operation. This entailed outfitting the turbine with additional instrumentation such as strain gauges in the rotor hub. The present study intends to extend the applicability of the DRC1.5 field tests by relating the strain signals to standard operational output. Machine Learning (ML) techniques include supervised learning by Artificial Neural Networks (ANN) and Long-Short-Term Memory (LSTM), as well as Principal Component Analysis (PCA). The same DRC test data sets were applied to ANN and LSTM and their results are compared. Discussions of results describe which generalize best for the purpose of sensor reduction, and which of the operational signals are most indicative of bearing strain. The results showed that both ANN and LSTM predicted future (or nonfunctional) sensor signals well with slightly higher accuracy by LSTM. Post processing of time series predictions can then track the progression of fatigue damage without additional sensors. Examining the prediction results details the model performance and highlights the relevance for wind turbine condition monitoring. Incorporation of learning techniques is presented as a systematic approach that can be replicated to simplify and optimize real monitoring strategies.