228. DETECTION OF ANOMALIES IN AIRCRAFT USING MACHINE LEARNING ALGORITHMS (Детекција на аномалии кај воздухоплови со примена на алгоритми од машинско учење)

Melanija Gerasimovska; Dushko Stavrov; Gorjan Nadzinski; Vesna Ojleska Latkoska

doi:10.51466/JEEIT25101228019g

Melanija Gerasimovska Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University in Skopje, 18 Rugjer Bošković Str., 1000 Skopje, Republic of Macedonia
Dushko Stavrov Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University in Skopje, 18 Rugjer Bošković Str., 1000 Skopje, Republic of Macedonia
Gorjan Nadzinski Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University in Skopje, 18 Rugjer Bošković Str., 1000 Skopje, Republic of Macedonia
Vesna Ojleska Latkoska Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University in Skopje, 18 Rugjer Bošković Str., 1000 Skopje, Republic of Macedonia

DOI: https://doi.org/10.51466/JEEIT25101228019g

Keywords: anomaly detection, predictive maintenance, aircraft, machine learning, unsupervised learning

Abstract

Abstract: This paper investigates the application of machine learning (ML) algorithms for anomaly detection in aviation, focusing on predictive maintenance and improving safety through early fault identification. Time-series vibration sensor data from helicopters is used to evaluate four anomaly detection methods: Isolation Forest (IF), One-Class SVM, Local Outlier Factor (LOF), and Convolutional-Reconstruction Autoencoder (CRAE). Both supervised and unsupervised detection approaches are considered. One-Class SVM demonstrated the highest performance, achieving accuracy of 93.9% and F1-score of 93.6%, followed by LOF (91.9%) and Isolation Forest (86%). CRAE underperformed with F1-score of 66.5%, primarily due to minimal preprocessing. These results highlight the effectiveness of simpler ML models over complex deep learning architectures in environments with limited data and real-time constraints.