This case study was presented at the 32nd International Conference on Artificial Neural Networks in Heraklion Greece September 2023 by Peter John Mayhew
A copy of the conference paper was accepted by the the ICANN (International Conference on Artificial Neural Networks) and published in Springer Link
Citation
Mayhew, P. J., Ihshaish, H., Deza, I., & Del Amo, A. (2023). Maintenance automation using deep learning methods: A case study from the aerospace industry. In Artificial Neural Networks and Machine Learning – ICANN 2023 (295-307). https://doi.org/10.1007/978-3-031-44204-9_25
Abstract
In this study, state-of-the-art AI models are employed to classify aerospace maintenance records into categories based on the fault descriptions of avionic components. The classification is performed using short natural language text descriptions provided by specialised repair engineers. The primary goal is to conduct a more comprehensive analysis of a complex and lengthy maintenance dataset, with the objective of determining the likelihood of failures in non-critical components of airplanes. Various methodologies are used, including two vectorisation models to natural language representation, as well as several machine learning algorithms such as BiGRU and BiLSTM, to identify repair and replacement event likelihood from the provided corpora. The resulting performance of the deployed models provide a very high F1 score overall, indicating models’ ability to learn repair patterns from the, typically complex, engineering description of components with high confidence. Two case studies are conducted. The first for a binary classification, with several models achieving an average F1 score of around \(95\%\). In the second case, a multi-class classification is performed for four different classes, with the BiLSTM model achieving the highest performance, accurately predicting the validation set with a \(95.2\%\) F1 score. The misclassified samples were manually inspected, and it was found that in many cases, the relevant information was simply missing from the text due to errors or omissions by description authors. Only \(12\%\) of the misclassified samples were found to be due to errors made by the model, resulting in an effective accuracy rate of \(\sim 99.4\%\).
Recent Comments