Acoustic Anomaly Detection in Hydropower Plants

This thesis investigates sound-based anomaly detection as a tool for predictive maintenance in hydropower plants. At the pumped-storage hydropower plant Rodundwerk II in Austria, acoustic data was collected in the turbine house and used to evaluate three state-of-the-art unsupervised machine learning approaches: a dense autoencoder, a convolutional autoencoder, and a Self-Organizing Map. Each model was designed to identify anomalies in recorded sound signals that may indicate mechanical irregularities, evaluated by the area under the ROC curve on induced anomalies. Hyperparameter tuning led to a significant increase in classification accuracy. Among the evaluated models, the Self-Organizing Map achieved the best anomaly detection performance on both the audio recorded in the plant and the external MIMII dataset. Although the convolutional autoencoder required more computational resources during training, this was not justified by better accuracy. The work demonstrates the potential of acoustic analysis for early fault detection in critical hydropower infrastructure and provides a foundation for future real-time monitoring systems.