Application of Natural Language Processing in Aviation Maintenance Management: Teaching machines to classify aircraft zones from maintenance reports using BERT

In the aircraft Maintenance, Repair and Overhaul (MRO) industry the high uncertainty of resources needed for a given operation is a very recurrent problem. This implies that being able to collect and analyze data from past events in order to predict for future operations provides a given company with more stability of outcomes and a competitive advantage in the industry. However, for most companies this is not the case. The heterogeneous manner in which the data is collected and stored causes it to be often disregarded or at least not used in its completeness. In the presented dissertation a NLP model is developed. This model analyses free text fields with the purpose of allowing MRO companies to exploit as much valuable information as possible from the maintenance reports generated throughout its activity. The model will focus on categorizing the zone of the aircraft where each maintenance task was performed. A BERT model was developed based on the existing maintenance reports of a Portuguese MRO. The research was performed with the support of the University of Bergamo, in which another work used the same dataset, but with different NLP models, in order to allow for comparison. An accuracy of 85.69% was obtained with BERT. A sensitivity analysis was also performed, showing that the most adequate combination of hyperparameters consisted in a longer warm-up period for the learning rate, along with low levels of both learning rate and dropout probability. The developed model is intended to be adaptable to other maintenance industries, other than aviation, or even to other economic sectors with due adaptations.