Gas-path measurements used to assess the health condition of an engine are corrupted by noise. Generally, a data cleaning step occurs before proceeding with fault detection and isolation. Classical linear filters such as the exponentially weighted moving average filter are traditionally used for noise removal. Unfortunately, these low-pass filters distort trend shifts indicative of faults, which increases the detection delay.
The present paper investigates two new approaches to non-linear filtering of time series. On one hand, the synthesis approach reconstructs the signal as a combination of elementary signals chosen from a pre-defined library. On the other hand, the analysis approach imposes a constraint on the shape of the signal (e.g., piecewise constant). Both approaches incorporate prior information about the signal in a different way, but they lead to trend filters that are very capable at noise removal while preserving at the same time sharp edges in the signal. This is highlighted through the comparison with a classical linear filter on a batch of synthetic data representative of typical engine fault profiles.