DETECTION OF UNSTABLE APPROACHES IN FLIGHT TRACK WITH RECURRENT NEURAL NETWORK

• Main Author: HANIFA, AINI
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/25225
• Institution: Institut Teknologi Bandung

Due to the increasing of air traffic density is needed to improve flight safety level <br /> <br /> especially in approach and landing phase. A potentially risky approach phase of <br /> <br /> flight is called an unstable approach. There are related studies in detecting these <br /> <br /> conditions, including anomaly detection of flight tracks for all flight phases. <br /> <br /> Compare with Multiple Kernel and Clustering methods, Recurrent Neural Network <br /> <br /> (RNN) has advantages in term of accuracy, sensitivity towards short term anomaly, <br /> <br /> and does not require dimensional reduction to identify the flight track anomaly <br /> <br /> pattern. However, RNN require experts to identify the risk event characteristics that <br /> <br /> cause anomalies. While heuristic methods can identify anomaly patterns <br /> <br /> specifically in the approach phase or called unstable approach with rule-based for <br /> <br /> each risk event. Therefore, this study combines the two methods approach to <br /> <br /> identify the unstable approach pattern. The main focus of this research is to prepare <br /> <br /> the data until ready to do the process of model formation by preprocessing <br /> <br /> technique, then done data modelling using RNN method with architecture stacked <br /> <br /> Long Short Term Memory (LSTM), and identify the type of risk event that influence <br /> <br /> unstable approach with heuristic method. In the modelling experiment performed <br /> <br /> by tuning the optimum value for each input parameter. The optimum value for batch <br /> <br /> size = 128, epoch = 150, and optimizer is rmsprop, with accuracy value equal to <br /> <br /> 90.12%, recall value equal to 59,44%, and precision value equal to 100%. That <br /> <br /> results show that the established model has been able to classify unstable and stable <br /> <br /> approaches, with high precision value, but the recall value is still low. Therefore <br /> <br /> for future work, the more amount of unstable approach data can be used to improve <br /> <br /> performance of data training.