Resumen
Predictive maintenance has received considerable attention in the aviation industry where costs, system availability and reliability are major concerns. In spite of recent advances, effective health monitoring and prognostics for the scheduling of condition-based maintenance operations is still very challenging. The increasing availability of maintenance and operational data along with recent progress made in machine learning has boosted the development of data-driven prognostics and health management (PHM) models. In this paper, we describe the data workflow in place at an airline for the maintenance of an aircraft system and highlight the difficulties related to a proper labelling of the health status of such systems, resulting in a poor suitability of supervised learning techniques. We focus on investigating the feasibility and the potential of semi-supervised anomaly detection methods for the health monitoring of a real aircraft system. Proposed methods are evaluated on large volumes of real sensor data from a cooling unit system on a modern wide body aircraft from a major European airline. For the sake of confidentiality, data has been anonymized and only few technical and operational details about the system had been made available. We trained several deep neural network autoencoder architectures on nominal data and used the anomaly scores to calculate a health indicator. Results suggest that high anomaly scores are correlated with identified failures in the maintenance logs. Also, some situations see an increase in the anomaly score for several flights prior to the system?s failure, which paves a natural way for early fault identification.