An Early Fault Detection Framework for Wind Turbines using Vibration Signals

Wind Turbines have been extensively adopted as primary renewable energy sources for over three decades. Many of these generators are now nearing the end of their expected operating time and have started to demand more intense maintenance. One alternative to extending their life is applying fault detection mechanisms to better schedule their maintenance. This work presents an early fault detection Framework based on Machine Learning applied to vibration data collected from Wind Turbines. The Framework includes mechanisms to filter less relevant data and decouple operational conditions from the signal. Feature extraction in the frequency domain and feature selection are applied to reduce data dimensionality and improve machine learning performance. Fault detection is achieved through a Deep Autoencoder model combined with a window-based reconstruction error analysis. The Framework is evaluated in a real case study, achieving promising results of detecting faults up to 76 days earlier than manual analysis.