Early UAV Motor-Fault Prediction Using Classical Machine Learning on RflyMAD

Abstract

Propulsion-fault anticipation is vital for multirotor-UAV safety. We formulate motor-fault prediction as a 3s earlywarning binary task and evaluate three lightweight classifiers–Logistic Regression (LR), Random Forest (RF) and Gradient Boosting (GB)–on a curated subset of the public RflyMAD corpus (34252 sliding windows, 8.1 % faults). Each window is summarised by 64 statistical features (mean, standard deviation, minimum, maximum) derived from 16 telemetry channels and standardised on the training split only. A leak-free GroupKFold (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k=3$</tex>) protocol ensures flight-wise separation; class imbalance is corrected with RandomOverSampler, and a 15 % validation slice sets the F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf>-optimal threshold. On the outer test folds RF delivers the best compromise, achieving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{F}_{1}=0.850 \pm 0.003$</tex>, ROC-AUC <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=0.994 \pm 0.000$</tex> and a false-alarm rate (FAR) <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=21 \pm 14 \mathbf{h}^{-\mathbf{1}}$</tex>, while issuing alerts <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.67 \pm 2.88 ~\mathrm{s}$</tex> before failure. GB follows closely <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\left(F_{1}=0.832; F A R=25 ~\mathrm{h}^{-1}\right)$</tex>, whereas LR trails (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$F_{1} =0.649; \text{FAR}=64 \mathbf{h}^{-\mathbf{1}}$</tex>). All three models satisfy the 3 -s warning criterion, but tree-based ensembles offer a superior recall-to-false-alarm balance and sub-millisecond inference, meeting edge-deployment constraints without deep-learning complexity. These findings support the use of certifiable classical models for on-board UAV fault prognostics and motivate future validation on real flights and multi-fault scenarios.