Kalman Prototypical Networks Enhance Turbine Fault Detection.

Combined-cycle gas turbines (CCGTs) are crucial for modern power generation due to their high efficiency and reduced environmental impact. However, their intricate thermo-fluid and mechanical interactions make fault detection challenging, especially when there's a scarcity of labeled fault data. Traditional methods often struggle in such few-shot learning scenarios. This research introduces the Kalman Prototypical Network (KPN), a novel metric-based few-shot learning (FSL) framework specifically designed for CCGT fault diagnosis. KPN enhances robustness and reduces episodic variance by modeling the evolution of class prototypes as latent stochastic states within a dynamic system, thereby improving embedding representation. Using synthetic datasets generated from a high-fidelity Modelica-based simulation of an offshore CCGT, the framework was tested on detecting progressive leak faults under transient conditions. KPN demonstrated superior performance in both accuracy and stability compared to conventional FSL methods like Matching Networks, Relation Networks, and MAML, across various support and query configurations. This stabilization of class representations leads to improved training convergence and generalization, making KPN highly suitable for real-world CCGT fault detection where labeled data is inherently limited.