Active Learning in Flight Anomaly Detection

Mural, PC and Gn, R and Bhola, V (2024) Active Learning in Flight Anomaly Detection. In: AIAA SCITECH 2024 Forum (AIAA 2024-1379); Session: Human-Machine Teaming: Models, Anomalies, and Evaluation, 8-12 January 2024, Orlando, FL, USA.

PDF AIA_Sci_Tec_ For_Exp_2024_ Jan_310199.pdf - Published Version Restricted to Registered users only Download (759kB) | Request a copy

Abstract

Flight anomaly detection is critical to ensuring aviation safety and operational efficiency. However, traditional supervised learning algorithms, which randomly select labelled samples and train, face challenges when dealing with flight anomaly datasets. These datasets often have imbalanced characteristics due to pilots� intention to fly without anomalies, resulting in a scarcity of anomalous instances. Consequently, randomly selecting samples may not adequately represent the minority classes in the dataset. We propose using Active Learning Techniques to overcome this challenge and enhance the collaboration between humans and machines in flight anomaly detection. Active Learning is an approach that carefully selects and labels samples from an unlabeled dataset based on their information content (like near the decision boundary). Selecting informative samples helps to improve the accuracy and efficiency of flight anomaly detection with less labelled data, thereby promoting effective Human-Machine Teaming in the aviation industry. This research compares various active learning algorithms on NASA Dashlink�s Flight Data Recorder (FDR) dataset comprising 99,837 flight landings with four classes (three different anomaly types and one nominal class). We compare the accuracy and F1 score of three uncertainty-based active learning algorithms (Least Confidence Method, Margin Sampling, and Entropy Sampling) and three Committee Disagreement-based Active Learning Algorithms (Vote Entropy, Max Disagreement, and KL Max Disagreement). We evaluate these algorithms for labelled sample sizes ranging from 0.1 to 1 of the dataset. The findings of this research reveal that the Vote Entropy active learning algorithm outperforms other algorithms with an impressive accuracy rate of 95.5. Notably, this high accuracy is achieved with minimal labelling effort, using only 1 of the available data. The successful application of active learning algorithms highlights their role in fostering successful collaboration between humans and machines, promoting the concept of Human-Machine Teaming in the aviation industry. © 2024 by Prashant Channappa Mural, Rathna GN and Virat Bhola.

Item Type:	Conference Paper
Publication:	AIAA SciTech Forum and Exposition, 2024
Publisher:	American Institute of Aeronautics and Astronautics Inc, AIAA
Additional Information:	The copyright for this article belongs to the authors.
Keywords:	Efficiency; Entropy; Learning algorithms; Learning systems; NASA; Supervised learning, Active Learning; Active-learning algorithm; Anomaly detection; Artificial intelligence; Flight anomalies; Human-machine; Human-machine teaming; In-flight anomalies; Machine learning; Machine-learning, Anomaly detection
Department/Centre:	Division of Electrical Sciences > Electrical Engineering
Date Deposited:	29 Jul 2024 05:20
Last Modified:	29 Jul 2024 05:20
URI:	http://eprints.iisc.ac.in/id/eprint/85197

Actions (login required)

View Item