From SCADA Pressure to Actionable Alarms: Calibration-Aware Monitoring of Kaplan Turbine Runner-Blade Actuation Friction

Abstract

Runner-blade friction and lubrication wear in a Kaplan turbine stay hard to observe directly between costly hub inspections. We introduce KFM-Cal, a unit-adaptive pressure-to-alarm algorithm that converts installed one-minute close/open servo-pressure SCADA into persistent friction and lubrication-risk alarms.  The algorithm keeps three stages distinct: mechanism-mapped extrac-tion of pressure evidence, target-normal adaptation of a frozen source ranker’s operating point, and causal proportion-voting alarm formation.  Target adaptation draws on a short lubricated reference segment and holds target fault labels out of training.  Our tests pair two public long-duration Kaplan records with 365 days of field SCADA from Sichuan Shawan. The T1 same-unit route reaches 0.987 MDF balanced accuracy at a 0.018 false-alarm rate, and T2-to-T1 transfer reaches 0.782 balanced accuracy at a 0.005 false-alarm rate. On the field record the frozen monitor catches all three engineering events, and MDF lowers routine false alarms from 0.030872 to 0.000441 with a mean delay of 30.96 h. Route-wise ablation, grouped SHAP, block permutation, and deep-model audits single out pressure and trend as the main evidence and quantify route-specific adaptation. The results establish a calibratable SCADA route from conditional net-actuation evidence to maintenance-facing alarms.