Answers are grounded in this paper and cite their sources.
Tactical vehicle terrain classification often relies on statistical features extracted from raw multisensor signals, although their relationship with physical vehicle and occupant responses is not always clear.This study develops a leakage-safe machine-learning approach using internally validated physical parameters for fine-grained terrain classification.Eleven features represented vibration and shock, seat-to-body transmissibility, dynamic stability, translational and rotational motion, and segmental mechanical-load response.Data were segmented into five-second windows with 50 percent overlap, producing 930 windows from 92 independent runs.Run identity was preserved throughout nested cross-validation.Five classifiers were evaluated.Gradient Boosting achieved the highest numerical seven-class Macro-F1 of 0.786, followed by Random Forest at 0.758.Leave-one-domain-out analysis showed that removing segmental mechanical-load features produced the largest and most consistent performance loss.Physical-11 also outperformed Conventional-9 and HFW-GB5 across all classifiers.Three-class gains were driven mainly by label aggregation, while stronger chronological grouping reduced the mean Macro-F1 of the corresponding run-level Physical-11 reference from 0.730 under standard Run_ID grouping to 0.652.
No takes yet. Share an insight, caveat, or question.
Mutiara et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: