Sensitivity of model rankings to perturbation scale

Determine how the rankings of virtual-sensing models under the MuViS-C robustness evaluation change when the perturbation scale differs from the fixed value k=3.

Background

MuViS-C applies sensor-failure perturbations to z-score-normalized channels and fixes the maximum perturbation scale at k=3, motivated by the conventional 3-sigma threshold for anomalous observations. This choice determines the magnitude of additive, stochastic, drift, scaling, and hard-fault corruptions at maximum severity.

Because robustness rankings may depend on the intensity and relative realism of the injected failures, the paper identifies the effect of alternative perturbation scales as unresolved. Addressing this question would test whether the reported cross-model comparisons are stable under different corruption magnitudes.