Posters

Abstract

A wind turbine generator is a complex rotating machine with multiple assembled critical subsystems. Since wind turbine generators placed in the same wind farms behave in a very similar way, it very often makes sense to compare the trend of a measured signal of a wind turbine to the trend of closer turbines or even to the whole fleet for the referred wind farm. However, for vibration analysis it may be interesting to compare the actual trend of vibration with the turbine's own historical data, once the vibration patterns for each machine may vary with its individual mechanical path. Gathering and analyzing vibration data is one of the most efficient ways to prevent long periods of unavailability by being noticed in advance of a failure and proceed with logistics. In this work, a process was developed to monitor the rotating subsystems in order to find any anomalous trends based on frequency domain analysis of vibration data. As a methodology, autoencoder models were trained in order to identify outliers points of vibration. After that, a statistical approach was applied considering moving averages to identify diverging trends and classify outliers prioritized by a linear regression model. Finally, a functionality was created in a proprietary system, developed in a Python web framework, to improve monitoring of maintenance execution and its effectiveness over the indicated variables. With the implementation of this tool and process, a better understanding of the wind turbine behavior allowed predictive detection of anomalies, enhanced maintenance planning optimization and maximized energy based availability, by carrying out interventions together with other preventive maintenance, or during low wind resource opportunities, as estimated by power forecast models.