Posters

Abstract

Main component failures are, usually, the most critical failures in wind turbine generators. Use Condition Monitoring System (CMS) data is the best way to anticipate a failure and strategically plan component replacement minimizing revenue losses. The two main challenges of using vibration data are: handling a large amount of data on an ongoing basis and defining a methodology to systematically identify faults accurately. This work proposes a methodology to solve these challenges using neural network models implemented in BigqueryML from GCP in the following the steps: (1) for each measurement position a dataset was created with the amplitude values at the main vibration levels; (2) then statistical filters were applied to remove outliers comparing the different wind turbines; (3) for each filtered dataset an autoencoder model was trained to reconstruct the normal operational data; (4) in a hold out dataset the thresholds to identify anomalies were determined; (5) finally, the model was evaluated on a temporally separated data for testing and the performance metrics were compared with the previous strategy (for one year of data in 156 wind turbines). These models are in production making predictions on a daily basis for a portfolio of 1.5 GW.