Diagnostic of erosive cavitation in hydraulic turbines from indirect measurements using SCADA data with uncertainty quantification

This paper proposes a data-driven framework for monitoring and diagnosing erosive cavitation in hydraulic turbines using supervisory control and data acquisition data only, with the objective of replacing costly and sparsely deployed condition monitoring systems. A deep-learning-based indirect measurement approach is developed to estimate the instantaneous cavitation rate from operational variables and is validated on multi-year real industrial data from two Francis turbine units, while ensuring model transferability across machines. Beyond point prediction, the framework explicitly addresses uncertainty quantification to support reliable decision-making, and several methods, including heteroscedastic regression, Bayesian neural networks, split conformal prediction, and conformalized deep ensembles, are benchmarked using calibration metrics and evaluated under distribution shifts. Diagnostics is performed by aggregating instantaneous predictions into an uncertainty-bounded cumulative degradation trajectory, enabling an uncertainty-aware assessment of cavitation progression and supporting reliability- and safety-informed maintenance decisions. Results demonstrate that operational data can effectively replace condition monitoring systems for cavitation monitoring and that conformalized deep ensembles provide the most reliable uncertainty estimates under realistic industrial conditions.