Ground Monitoring Console with Causal Link-Quality Prediction

Ground monitoring becomes more useful when aircraft state, communication condition, time, and location are interpreted together rather than as independent streams. This paper presents an integrated ground-monitoring workflow that combines historical flight-telemetry visualization with a causal multi-horizon machine-learning pipeline for predicting future wireless-link change. The main prediction target is future change in signal-to-interference-plus-noise ratio (ΔSINR) at horizons of 1, 2, 5, 10, 15, 20, and 30 s. The quantitative study uses three physical flight records drawn from two public AERPAW measurement sources: two Ericsson 5G NSA traces at yaw orientations of 45° and 315°, and one Lake Wheeler Android-based 4G LTE/5G NR trace. The records contain LTE/NR radio indicators, throughput, position, altitude, cell information, and motion/orientation fields, with 2376 raw observations across the three physical flights. The processing uses a causal 1-s grid, flight-isolated temporal features, chronological evaluation with a horizon-sized purge, training only mutual-information feature selection, and a persistence baseline. Three classical regressors—HistGradientBoosting (HGB), Random Forest, and Extra Trees—are evaluated. An engineering link-quality index (LQI) in the range 0–100 summarizes 12 available cellular and throughput indicators. At the predeclared 30-s operating point, HGB obtains a ΔSINR MAE of 3.110 dB versus 3.669 dB for persistence, a 15.23% reduction. A separate NASA Tail-654 ground-monitoring implementation replays a historical MATLAB flight record, presents map-based trajectory and primary telemetry, and applies phase-aware statistical anomaly monitoring using robust envelopes, eight-second persistence, and recovery hysteresis. The NASA component is kept experimentally separate from the wireless-link prediction metrics. The resulting system connects telemetry observability with forward-looking communication-state information while preserving a strict separation between information available at prediction time and future observations.