Aircraft engine remaining useful life prediction: a comparison study of Kernel adaptive filtering architectures

Abstract

Predicting the Remaining Useful Life (RUL) of mechanical systems poses significant challenges in Prognostics and Health Management (PHM), impacting safety and maintenance strategies. This study evaluates Kernel Adaptive Filtering (KAF) architectures for predicting the RUL of aircraft engines, using NASA's C-MAPSS dataset for an in-depth intra-comparison. We investigate the effectiveness of KAF algorithms, focusing on their performance dynamics in RUL prediction. By examining their behavior across different pre-processing scenarios and metrics, we aim to pinpoint the most reliable and efficient KAF models for aircraft engine prognostics. Further, our study extends to an inter-comparison with approximately 60 neural network approaches, revealing that KAFs outperform more than half of these models, highlighting the potential and viability of KAFs in scenarios where computational efficiency and fewer trainable parameters are both crucial. Although KAFs do not always surpass the most advanced neural networks in performance metrics, they demonstrate resilience and efficiency, particularly underscored by the ANS-QKRLS algorithm. This evaluation study offers valuable insights into KAFs for RUL prediction, highlighting their operational behavior, setting a foundation for future machine learning innovations. It also paves the way for research into hybrid models and deep-learning-inspired KAF structures, potentially enhancing prognostic tools in mechanical systems.