From raw data to monotonic and trendable features reflecting degradation trends in turbofan engines

The performance of prognostic models relies heavily on the form and trend of the extracted features. However, the raw data collected from physical systems are inherently noisy, large in volume, and exhibit significant variability, which makes them unsuitable for direct use in prognostics. These characteristics poorly reflect the degradation behavior of physical systems and contribute to the uncertainty of prognostic outcome. Hence, transforming this data into relevant features and carefully selecting them is crucial for meeting the specific needs of prognostic models. This paper aims to address data processing challenges by focusing on extraction and selection of high-quality monotonic features which clearly reflect the degradation and can reduce prognostics uncertainty. The proposed framework comprises three main stages: Data pre-processing, feature extraction, and feature selection. It includes a fitness analysis to evaluate the monotonicity and trendability of features supplemented by visual inspections to identify relevant features. Applied to the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) dataset from the NASA Ames Prognostics Data Repository, the framework reduces noise, improves feature monotonicity and trendability, and facilitates the selection of useful features - essential aspects for effective prognostic methods.