Development of CNN and LSTM fusion model for early fault warning in hydro turbine units.

Abstract

Early fault detection is important for the safe and stable operation of large hydro turbine units under complex and variable operating conditions. This study develops an early fault warning framework for high dimensional multisource monitoring data by integrating Kendall rank correlation-based parameter grouping, systematic data governance, sparse principal component analysis, and a CNN LSTM prediction model. Fifty two operating parameters are organized into six groups and transformed into six sparse state features. The CNN LSTM model captures local correlations and temporal evolution, achieving R² values of 0.947, 0.955, 0.848, 0.701, 0.687, and 0.853 for SPCA1 to SPCA6. Prediction residuals combined with 2σ and 3σ thresholds identify early degradation before the recorded fault, with warning lead times of up to 33 time steps. The framework provides a practical approach for condition monitoring and hierarchical early warning of large hydro turbine units.