Kalman Prototypical Networks for Few-shot Fault Detection in Combined Cycle Gas Turbines

In industrial settings like combined-cycle gas turbine (CCGT) operations, fault data is notoriously scarce and expensive to label. While few-shot learning (FSL) offers a way to train models with minimal examples, standard prototypical networks often suffer from high episodic variance—where the calculated class prototypes fluctuate significantly between training episodes. This paper explores whether modeling these prototypes as latent stochastic states in a dynamic system can stabilize the learning process and improve fault detection performance.

The authors introduce the Kalman Prototypical Network (KPN). In this framework, the prototype for each class is not just a simple mean of support samples; instead, it is treated as a latent state in a linear dynamical system. By applying a Kalman filter to these prototypes across training episodes, the model effectively "denoises" the representation of each class. This filtering process provides a more stable anchor in the embedding space, which helps the neural network learn more robust features for distinguishing between normal operations and various leak faults.

The researchers evaluated the KPN framework using synthetic data generated from a high-fidelity Modelica-based simulation of an offshore CCGT system. The results indicate that KPN consistently outperforms conventional FSL methods, including Matching Networks, Relation Networks, and MAML. The KPN approach demonstrated superior accuracy and stability under various support and query configurations, suggesting that the Kalman-based smoothing effectively mitigates the instability typically associated with few-shot learning in data-constrained industrial environments.

This study provides a practical solution for predictive maintenance in complex industrial systems where labeled fault data is rare. By stabilizing the representation of fault classes, the KPN framework allows for more reliable and consistent diagnostic performance. This is particularly valuable for safety-critical infrastructure where early detection of anomalies, such as tube leaks, is essential to prevent costly downtime and environmental hazards.