Multimodal Sensing for Acoustic–Vibration Based Fault Localization in Hydropower

Pumped-storage hydropower underpins grid-scale flexibility for renewable generation, but its frequent mode changes expose a weakness of acoustic fault diagnostics: when the operating regime shifts, the baseline signature moves with it and unimodal detectors raise synchronized false alarms, a domain shift confirmed on the industrial Rodundwerk II (ROW II) machine. This thesis asks a sharper question than the usual fuse-and-condition remedy: at realistic data scale, when does jointly trained multimodal fusion earn its complexity over simpler unimodal, late-fusion, or classical alternatives? It answers with a leakage-controlled comparison across anomaly detection and source localization. Because ROW II recordings could not be released in time, the study uses a 3Dprinted, reduced-scale turbine prototype recorded over five campaigns.