Aero Engine Performance Monitoring Using Least Squares Regression and Spectral Clustering

Abstract

Threshold-based flight data recorder analysis techniques have been widely used across the aerospace industry for fault detection and accident prevention. These techniques can detect pre-programmed events but fail to capture unknown patterns in the dataset. This research proposes a machine learning (ML) algorithm to analyze and detect unusual aero engine performance of a turboshaft engine mounted on a single engine rotorcraft. The performance is first modelled from an FDR dataset consisting of hundred flights, using least squares regression (LSR). A technique to scale the model by adding flight data from subsequent flights is thereafter discussed. Spectral Clustering is used for testing and validating the hypothesis derived from the regression model, by employing synthetically generated FDR data for twenty-five flights.