Anomaly Detection for Large Hydrogenerators Using the Variational Autoencoder Based on Vibration Signals

This paper presents the potential and the sensitivity of the Variational Autoencoder (VAE) technique in early detection of a failure mode of large hydrogenerator. Real vibration data collected in situ, acquired from the site for a healthy machine, are used for the training and validation of this technique. Synthetic faulty signals (rotor inter-turn short-circuits faults), in which statistical parameters are calculated, have been used to test the sensitivity of the model in detecting the occurrence of this fault. The faulty signals were created by injecting the fault, based on its frequency pattern, to a set of healthy vibration signals collected in situ. The obtained results prove the prevalence of this technique and its sensitivity in detecting anomalies at early stages based on the variation of the reconstruction error