Using Kolmogorov-Arnold Networks for an Interpretable and Continual Fault Detection of Helicopter Turbine Engines

In recent years, data-driven methods have become common in fault-detection systems, with deep learning architectures - particularly Multi-Layer Perceptrons (MLPs) - achieving state-of-the-art performance. However, MLPs suffer from two critical limitations in the context of safety-critical systems: lack of interpretability and vulnerability to catastrophic forgetting under continual learning scenarios. In this paper, we present a novel approach to predictive maintenance using Kolmogorov-Arnold Networks (KANs), which are based on a different mathematical foundation than MLPs. To evaluate the efficacy of KANs for engine-health monitoring, we benchmark their performance against conventional MLPs on the PHM North America 2024 Conference Data Challenge. Our results show that KANs significantly increase model transparency - allowing the predictions to be trusted - and exhibit superior resilience to forgetting, while still achieving high test scores. Our codes can be found at https://github.com/MrPio/PHM_North_America_2024_Challenge.