Development of a Feature Extraction Methodology for Prognostic Tasks of Aerospace Structures and Systems

The performance of prognostic models used for prognostic health management (PHM) applications heavily depend on the quality of features extracted from raw sensor data. Traditionally, feature extraction criteria such as monotonicity, prognosability, and trendability are selected intuitively. However, this intuitive selection may not be optimal. This research introduces an innovative approach to transform raw data into 'high-scoring' data without the need for predefined feature extraction criteria. Our methodology involves generating a set of synthetic high-scoring time series. These synthetic time series, resembling the length and amplitude of target features, are created through Monte Carlo sampling (MCS) of a user-defined hidden semi-markov model (HSMM). These synthetic time series are paired with raw data/features from the signals and use them as targets to train a convolutional neural network (CNN). As a result, the trained CNN can convert input features into high-scoring ones, irrespective of their initial characteristics. So, this study provides the following contribution to PHM frameworks: it transforms raw data/features into high-scoring ones without relying on predefined criteria, rather on stochastically generated labels that resemble the nature of the degradation processes. It is worth noting, that the proposed FE technique is independent of the prognostic model that will be utilised, thus making it applicable to the established prognostic models. The effectiveness of this approach will be demonstrated and validated using acoustic emission (AE) sensor data for remaining useful life (RUL) estimation of open-hole CFRP specimens. The prognostic performance will be compared using cumulative AE features with their transformations via the proposed framework. The transformed features exhibit superior prognostic performance, underscoring the value of our innovative feature extraction framework.