Sensor Choice Reverses Under Operating Condition Shift for Robotic Condition Classification

Abstract

Vibration and acoustic sensing are established nondestructive evaluation modalities for machine state estimation, and their fusion is now standard practice in condition monitoring. What this practice does not establish is what each modality independently contributes, or whether its in-distribution advantage survives an operating-condition shift, a question with direct cost consequences. This paper treats sensing modality as the experimental variable, evaluating vibration-only, acoustic-only, and fused configurations from a delta robot against one fixed classifier, scored on in-distribution and out-of-distribution accuracy under a held-out payload, combined with a cost model into a Pareto comparison. Vibration-only outperforms acoustic-only in-distribution (98.5% vs. 90.0% balanced accuracy), but under payload shift the two become statistically indistinguishable (54.2% vs. 54.7%), despite acoustic costing half as much, dominating vibration on the frontier. Fusion achieves the highest accuracy in both regimes but the lowest accuracy-per-dollar yield. An eight-classifier comparison shows the same reversal: the strongest in-distribution classifier is not the most shift-robust. We further prove a fused configuration’s population risk cannot exceed any single modality it contains, turning an observed single-modality advantage into a predictable finite-sample effect. These results show evaluating modality and classifier choice in-distribution alone yields a different, more expensive recommendation than a shift-aware evaluation supports.