Hybrid Quantum-Classical Machine Learning for UAV Engine Failure Severity Classification: A Multi-Flight Benchmark on ALFA Telemetry Data
Atul Rawal, Charles A. Milligan
This paper compares classical and hybrid quantum-classical machine learning methods for classifying the severity of UAV engine failures from telemetry data. Using 23 engine failure flights from the ALFA dataset, we create a severity label from per-flight z-scored latent variables and validate it against ground-truth fault onset timestamps. Classical models are tested with leave-one-flight-out (LOFO) cross-validation. The results show that strong within-flight F1 scores overestimate performance on unseen flights by 0.19-0.37 points for complex models. We also run a quantum pilot study on an identified flight, comparing a Variational Quantum Classifier (VQC) and a Quantum Kernel SVM with classical baselines using the same features and sample budgets. Under identical PCA and sample-budget conditions, the Quantum Kernel SVM exceeds all matched classical baselines. A qubit width sweep, together with per-layer gradient variance measurements, provides empirical evidence of barren plateaus where the VQC performance gets worse as width increases, while the kernel SVM improves with the richer representation. We also test noise robustness using consistent train/test splits and evaluate the quantum model across three flights. Overall, the results are presented as an preliminary/exploratory benchmark, with cross-flight generalization remaining the main open challenge.
Proceedings of the AAAI Symposium Series•Nov 6, 2026•0 citations