QAR flight-data feature analysis and model training

Problem

QAR data is dozens of high-frequency parameters; staring at curves finds nothing. The goal was to extract comparable features from a landing and retrieve historically similar cases.

Approach

Windows are cut per flight phase and features extracted (vertical acceleration at touchdown, pitch rate, throttle movement). After normalisation a similarity index returns the closest historical segments for a given flight.

Details

Features are stored as Parquet; retrieval uses vector distance plus key-parameter filters; training and inference run as separate processes, with inference loading weights only.

• Python
• feature engineering
• retrieval
• models

Open the demo →

Download source ↓ (free with an account)

Compliance note: demo data is de-identified or synthetic, with no real flight operations records; real data would require authorisation from the operator.