Dynamic Behaviour Classification in Multi-Domain Operations Using ADS-B and Air Traffic Data

Abstract

The increasing complexity of air traffic in multi-domain operations (MDO) demands real-time monitoring and classification of aircraft behaviour to support safety, efficiency, and situational awareness. Conventional approaches to conformance monitoring and behaviour classification are largely batch-oriented, relying on full-trajectory analysis and neglecting operational context, which limits their applicability during evolving flight operations. This dissertation proposes a dual-pipeline, AI-based framework that integrates unsupervised anomaly detection for conformance monitoring and supervised behaviour classification using Automatic Dependent Surveillance-Broadcast (ADS-B) and air traffic control (ATC) context. An LSTM-Autoencoder (LSTM-AE), incorporated with Notice to Airmen (NOTAMs) constraint, models nominal kinematics and flags deviations via reconstruction error, while a Bi-directional LSTM (Bi-LSTM) classifies behaviour at each time instant. To enable real-time inference, both models operate on overlapping sliding windows, providing early decisions while preserving short-term temporal dependencies. Experiments on ADS-B trajectories from the OpenSky Network show that the proposed framework achieves strong classification performance and the LSTM-AE reliably filters non-nominal patterns via elevated anomaly scores, by incorporating NOTAM-derived constraints improves operational relevance. These results demonstrate the feasibility of real-time, context-aware trajectory monitoring for MDO environments.