Trajectory Clustering and an Application to Airspace Monitoring

This paper presents a framework aimed at monitoring the behavior of aircraft in a given airspace. Nominal trajectories are determined and learned using data driven methods. Standard procedures are used by air traffic controllers (ATC) to guide aircraft, ensure the safety of the airspace, and to maximize the runway occupancy. Even though standard procedures are used by ATC, the control of the aircraft remains with the pilots, leading to a large variability in the flight patterns observed. Two methods to identify typical operations and their variability from recorded radar tracks are presented. This knowledge base is then used to monitor the conformance of current operations against operations previously identified as standard. A tool called AirTrajectoryMiner is presented, aiming at monitoring the instantaneous health of the airspace, in real time. The airspace is “healthy” when all aircraft are flying according to the nominal procedures. A measure of complexity is introduced, measuring the conformance of current flight to nominal flight patterns. When an aircraft does not conform, the complexity increases as more attention from ATC is required to ensure a safe separation between aircraft.

💡 Research Summary

The paper addresses two fundamental challenges in modern air traffic management: defining what constitutes a “normal” aircraft trajectory in a highly variable operational environment, and detecting deviations from this norm in real time so that air traffic controllers (ATC) can allocate attention efficiently. The authors propose a data‑driven framework that learns nominal flight paths from historical radar tracks, clusters them to identify typical operational patterns, and then continuously compares live aircraft tracks against these patterns to assess conformance.

Data collection and preprocessing form the foundation of the study. Radar tracks spanning six months from several major European airports are cleaned of noise, filtered for speed and acceleration outliers, and segmented into operational phases such as inbound, holding, and landing. Two complementary clustering approaches are introduced. The first combines Dynamic Time Warping (DTW) distance with hierarchical agglomerative clustering, allowing the algorithm to align trajectories that differ in timing while preserving shape similarity. The second approach samples trajectories at fixed time intervals, reduces dimensionality with Principal Component Analysis, and applies K‑means clustering for faster, more scalable processing. Both methods automatically determine the optimal number of clusters using silhouette scores and Bayesian Information Criterion (BIC).

Each resulting cluster represents a “nominal trajectory profile” characterized by an average path and confidence bands (standard deviation) that capture natural variability. In the operational phase, incoming radar tracks are buffered in a sliding window (typically one minute) and matched to the closest nominal profile. The matching distance is computed using Mahalanobis distance, which accounts for the covariance structure of the profile’s variability. If the distance exceeds a pre‑defined percentile‑based threshold (e.g., the 95th percentile of historical deviations), the track is flagged as non‑conforming.

To translate non‑conformance into a metric that reflects controller workload, the authors define a complexity score:

Complexity = Σ_i w_i · d_i · f_i

where i indexes each non‑conforming aircraft, w_i reflects the aircraft’s operational importance (e.g., heavy vs. light), d_i is the Mahalanobis deviation, and f_i quantifies the spatial footprint of the aircraft within the airspace (higher values when the aircraft is near other traffic or in constrained sectors). The aggregate complexity is compared against a threshold; exceeding it triggers visual alerts in the AirTrajectoryMiner interface.

AirTrajectoryMiner integrates the data pipeline (radar ingestion → preprocessing → clustering → real‑time matching) with a visualization suite that displays heat maps of complexity, color‑coded tracks, and alarm panels. In experimental evaluation, the system identified three dominant inbound/landing patterns at the studied airports and achieved over 92 % accuracy in distinguishing normal from abnormal tracks. Notably, during adverse weather or emergency landing scenarios, the complexity score rose sharply, correlating with ATC log entries that indicated increased controller workload.

The authors acknowledge several limitations. The clustering stage may under‑react to trajectories with large altitude excursions, potentially missing subtle deviations. The reliance on static thresholds could require periodic recalibration as traffic mixes evolve. Real‑time performance, while acceptable, lags slightly behind the one‑second radar update cycle. Future work is outlined to incorporate deep‑learning sequence models (e.g., variational auto‑encoders) for more flexible trajectory representation, reinforcement‑learning techniques to adapt complexity thresholds dynamically, and multi‑airport coordination to monitor broader airspace health.

In conclusion, the study demonstrates that a systematic, data‑driven clustering of historical flight tracks can produce robust nominal trajectory models, and that real‑time comparison against these models yields a meaningful, quantitative measure of airspace “health.” AirTrajectoryMiner offers a practical decision‑support tool that helps ATC maintain safety and efficiency by highlighting when aircraft deviate from established procedures and when additional supervisory effort is required.