Physics-consistent Bayesian early warning of wind turbine blade stall using an auxiliary regularized particle filter

Early warning of wind turbine blade stall is difficult because early aerodynamic degradation is weak and easily masked by operating-condition variability and non-Gaussian disturbances in SCADA data. This study proposes a physics-consistent Bayesian sequential early-warning framework based on an auxiliary regularized particle filter. A physically constrained nonlinear state-space model with a latent health factor represents aerodynamic degradation while preserving the coupling among wind speed, rotor speed, and power. For stable inference under turbulence and model mismatch, the likelihood-enhanced filter adds a bounded operating-condition likelihood correction, stall-risk-gated noise adaptation, and kernel-density-regularized resampling. A decision layer with dynamic thresholds, persistence constraints, and a recovery-time criterion converts pointwise anomalies into event-level warnings. The framework is validated on historical SCADA data from two wind farms, covering 8 fault cases and 60 healthy runs. Across the eight fault runs, ARPF achieves a mean two-hour pre-shutdown coverage (Coverage@2h) of 96.7%, exceeding RPF by 8.5% and EKF/UKF by 81.3%. On the two stall cases at wind farm A, ARPF detects and warns on both, whereas four standard data-driven baselines (random forest, XGBoost, LSTM, and CNN) detect at most one. These results show that the proposed framework provides earlier and more sustained blade-stall early warning from SCADA data than the Gaussian-filter and power-curve-residual baselines, while maintaining reliability under varying operating conditions.