LOC Dataset: Loss of Control flight trajectories

The primary kinematic dataset, collected from National Transportation Safety Board (NTSB) dockets (https://data.ntsb.gov/carol-main-public/query-builder), and aircraft performance data, sourced from FlightAware (https://www.flightaware.com/resources/registration) and the Base of Aircraft Data (BADA) (https://learningzone.eurocontrol.int/ilp/customs/ATCPFDB/default.aspx?).

Dataset and Flight Telemetry (ADS-B): The primary dataset consists of 50 historically recorded Loss of Control (LOC) flight trajectories, including extreme aerodynamic departures such as stall-spins and spiral dives. The kinematic trajectory data was sourced from Automatic Dependent Surveillance-Broadcast (ADS-B) telemetry. ADS-B provides high-fidelity, GPS-derived positional and velocity tracking, making it a standard real-world data source for aviation safety analysis.

To ensure temporal consistency across all flights, the raw ADS-B telemetry was interpolated to a strict 1 Hz sampling frequency (Δt = 1.0 second). Because raw ADS-B coordinates are provided in a geodetic frame of reference (Latitude, Longitude, Pressure Altitude), they are not directly suitable for modeling Euclidean spatial distances in a neural network. Therefore, the geodetic coordinates were projected into a local Cartesian North-East-Down (NED) reference frame, with the Z-axis inverted to represent positive altitude.

Selected Kinematic Features: To train the trajectory prediction models, a 6-dimensional state vector was constructed for every time step. The inputs combine physical location with instantaneous velocity to give the autoregressive decoder a complete kinematic profile of the aircraft:

Spatial Coordinates (x, y, z): The position of the aircraft converted to local Cartesian coordinates in meters. To ensure the model learned generalizable flight dynamics rather than absolute geographic locations, all spatial trajectories were zero-centered relative to the final known coordinate of the observation window.

Velocity Components (Vx, Vy, Vz): The directional velocities along the North, East, and vertical axes. These values, originally recorded in knots and feet per minute (ft/min), were converted to meters per second (m/s) for model training.

The sequences were structured with a 50-step historical observation window to predict a 25-step future trajectory horizon. All features were normalized using standard scaling prior to network training  to ensure stable gradient descent.

Aircraft Performance Data: The dataset is used for training a  Physics-Informed Transformer (PIT). A key contribution of the PIT framework is the integration of specific aircraft performance capabilities into the loss function. To facilitate this, each flight sequence was paired with static performance metadata mapped to the aircraft's specific registration number:

Mass (kg): The maximum takeoff weight of the aircraft, critical for calculating required lift and gravitational forces.

Maximum Power (W): The rated engine power, used by the physical model to bound the available specific excess power and thrust constraints.

Wing Area (m²): The aerodynamic reference area, required to dynamically compute parasite drag and induced drag during the prediction horizon.

By coupling the 6D kinematic state vector with these physical constants, the network was forced to generate future trajectories bounded by the actual mechanical capabilities of the specific airframe.