HydroPrior: Offline LLM-Prior-Guided Residual Warning for Hydropower Bearing Vibration

Abstract

Hydropower vibration monitoring requires early warnings with manageable inspection costs despite scarce verified fault labels. HydroPrior combines offline large language model (LLM) priors with lightweight residual-based warning. The LLM converts channel descriptions and training-set statistics into frozen operating-context features, which enter a compact forecasting correction head. Residual scoring, calibrated thresholds, and contextual confirmation then produce budgeted alerts. Evaluation uses rapid vibration transitions as proxy events and the Balanced Warning Score (BWS) to summarize event coverage and alarm burden. On a public hydropower bearing-vibration dataset, HydroPrior obtains the lowest average forecasting error among the evaluated models. Under the proxy-warning protocol, its mean BWS is 0.7492 (standard deviation: 0.0405), with a transition-hit rate of 0.9106, 0.15 false alarms per 1000 non-transition points, and 0.0199 alert coverage. Comparisons using a common confirmation mechanism and coverage budget also favor HydroPrior. The findings support the feasibility of frozen operating priors for lightweight vibration forecasting and focused warning on the evaluated unit record.