Abstract
The maintenance of wind turbines is essential to reduce wind energy levelized costs. Earlier detection of potential faults in the rotating subcomponents, such as the drivetrain, helps to plan maintenance actions. Several vibration processing methods, e.g., short-time Fourier analysis, are available in the literature to detect faults, however, they require domain expertise. Moreover, many researchers are focusing on machine learning methods to complement such techniques. This paper combines deep learning methods, more specifically auto-encoders, with more than 400 indicators based on advanced signal processing techniques. It is impractical to train a deep-learning model for each indicator, and significant manual effort is required to investigate all indicators. This paper focuses on a multivariate deep learning model to explore the potential to learn the underlying relationships among the signal processing indicators for healthy datasets and compress them into one health status. The reconstruction error of this model is then used to identify changes in the condition of the system. The output directly illustrates the high-level health state of the system for early fault detection. This proposed method is demonstrated on real-life offshore wind turbine data from several years. Further analysis can be done to pinpoint specific fault types using frequency spectrum analysis.
