Support vector regression guided multivariate EWMA monitoring with variable sample size

Multivariate EWMA control charts with a fixed smoothing constant and a fixed sample size face a persistent trade-off, as they cannot respond quickly to a large shift while also remaining efficient during normal operation or small shifts. Existing adaptive extensions typically address only one of these two fixed design choices at a time—adapting the smoothing constant but not the sample size, or vice versa. This paper addressed this gap with an SVR-based adaptive multivariate EWMA chart with variable sample size (SAMEWMA-VSS) that adapts both simultaneously within a single Mahalanobis-form charting statistic: An SVR model maps the estimated shift magnitude to the adaptive smoothing constant, while a variable-sample-size rule sets the subgroup size drawn at each inspection. Monte Carlo simulation showed that the proposed chart detects small-to-moderate shifts faster than existing adaptive multivariate EWMA charts, is invariant to the process correlation structure and to the shift direction by construction, and maintains stable performance across the investigated process dimensions. A real-data application to NASA bearing vibration monitoring demonstrates its practical use with physically interpretable features. This zero-state advantage, however, does not fully carry over to continuous, steady-state monitoring, a limitation discussed together with its design implications for practitioners choosing between the two charts.