Kalman Prototypical Networks Improve Few-Shot Fault Detection in Gas Turbines
June 26, 2026
Kalman Prototypical Networks model class prototype evolution as latent stochastic states in a dynamic system, reducing episodic variance in few-shot fault detection for combined-cycle gas turbines. Evaluated on Modelica-simulated progressive leak faults, KPN outperforms conventional prototypical network baselines.
HOW THIS AFFECTS YOU
●
researcherTreating few-shot prototypes as Kalman-filtered latent states is a principled way to handle non-stationary embeddings in industrial time-series — applicable beyond turbine fault detection to any scarce-label anomaly detection setting.