Optimizing Automatic Flight Condition Recognition through a Multi-Strategy Machine-Learning Based Approach

Flight Condition Recognition (FCR) is essential in the usage monitoring of helicopters, as maneuver instances determine the usage spectrum, and thus the assessment of the original usage assumptions, adopted at design time for the definition of the retirement life of its components. Automated FCR capabilities, exploiting algorithms to detect the aircraft maneuvers by appropriate processing of on-board sensors measurements, allow us to reconstruct the usage spectrum, supporting the definition of improved maintenance manuals, with replacement times and inspection intervals tailored to the helicopters actual usage, thus enabling Condition-Based Maintenance (CBM) schemes. However, designing an efficient automatic FCR system is a challenging task, due to the complex machine dynamics characterizing the different flight regimes. In this work, we show how to optimize a machine-learning based approach to FCR design by exploiting a multi-strategy time-series segmentation framework, which combines two supervised learning approaches that leverage sliding windows and stacking ensembles to produce reliable estimates of the flown regimes. The approach is validated on an experimental dataset of nearly 500 labeled flights from two helicopter models, demonstrating its effectiveness in predicting the different maneuver types, and its improvement over a single-strategy approach.