Edge-Deployed Anomaly Indicators for UAV Telemetry with Context-Aware Advisory Decision Support
TL;DR
This paper presents a companion-computer chain that converts Pixhawk-class MAVLink telemetry into operator advice on an NVIDIA Jetson Nano, and labelled evaluation uses synthetic injection, not concurrent live-load flight scoring.
Abstract
This paper presents a companion-computer chain that converts Pixhawk-class MAVLink telemetry into operator advice on an NVIDIA Jetson Nano. A twelve-channel sliding window (L=50, 10 s at 5 Hz) feeds a reference long short-term memory autoencoder and an edge-evaluated dense TensorFlow Lite model, which were trained with a session-level split of eight field sorties (train 14,429/validation 7281/test 6976 windows). Reconstruction mean squared error is mapped by validation-calibrated thresholds and a safe-inaction-first decision engine with K=3 action persistence. On a per-sortie fault-injection benchmark, One-Class SVM attains the highest pooled F1 score (0.713); the edge-evaluated dense TensorFlow Lite model reaches F1 0.524 (detector false-alarm rate 5.8%) at a 1.57 ms median on-device cost. Decision replay reports a 0.68% escalation false-alarm rate with the on-board gyro-x plausibility gate enabled, and 2.72% with the gate off. The study is a replay-profiled prototype on a single hexacopter; labelled evaluation uses synthetic injection, not concurrent live-load flight scoring.