Machine Learning for Predictive Maintenance

Abstract

The newest generation of aircraft has seen a strong increase in sensor data generated on-board. The available data has the potential to indicate the health state of individual components based on which their maintenance requirements can be determined, a maintenance strategy called Condition Based Maintenance. Predictive Maintenance is a specific condition based maintenance strategy that aims to determine these requirements in advance by predicting failures from the sensor data. It has the potential to reduce costly unanticipated maintenance or unnecessarily conservative maintenance. In the short term it could add significant value for components that are currently subject to a reactive maintenance policy. In the long term it could potentially disrupt the traditional maintenance practice of periodic inspection. One of the main challenges in applying Predictive Maintenance in the aviation industry is translating the large amounts of sensor data into a reliable failure prediction, a process called prognostics.

In this study, state-of-the-art machine learning, and specifically deep learning models, have been investigated for their potential for prognostics. A case study has been performed at KLM Royal Dutch Airlines on the Boeing 747 Bleed Air Valves, traditionally some of the most challenging components from a maintenance perspective. It has been shown that fully self-learning algorithms can be used for prognostics, enabling the implementation of one of the first real-life predictive maintenance implementations.