The Impact of Data Representation on Predicting Aircraft Component Removals

In the field of aeronautics, predictive maintenance plays a crucial role in maximizing aircraft reliability, operational efficiency, and cost-effectiveness by proactively identifying and preventing potential failures. During flight, aircrafts continuously collect various metrics (such as flight parameters) and generate diagnostic messages (like fault codes). Certain fault codes trigger in-depth investigations, demanding time, expertise, and, sometimes, the removal and replacement of specific components. This work explores how different data representation approaches can significantly influence the effectiveness of predictive maintenance models. We demonstrate that, while traditional feature engineering based on domain expertise can yield strong results, learned representations can outperform these experience-based techniques. Specifically, the aircraft’s diagnostic messages can be interpreted as a type of “language” from which a transformer encoder neural network can learn a robust high-quality representation. Experiments on real-world data confirm the effectiveness of this approach, underscoring its potential in enhancing predictive maintenance systems.