Hybrid Statistical Process Control — Combined SPC

The terms are used interchangeably in the literature, but 'hybrid' typically implies that components from different methodological traditions are joined — for example, a statistical chart augmented with a machine-learning classifier or a fuzzy-logic rule engine — rather than two charts from the same family (e.g., Shewhart plus CUSUM, which some authors call 'combined' rather than 'hybrid'). The defining feature is that the augmenting component brings a qualitatively different capability, such as pattern diagnosis or adaptive parameter updating.

Start from the shift sizes you need to detect and the acceptable in-control ARL. If you need to detect both large shifts quickly and small persistent drifts, pair a Shewhart chart (sensitive to large shifts) with a CUSUM or EWMA (sensitive to small shifts). If the process data are multivariate, consider T-squared paired with MEWMA. If operators need pattern diagnosis, add a neural-network classifier trained on the six standard control-chart pattern types.

Often yes. Closed-form ARL expressions exist for some combined Shewhart-CUSUM and Shewhart-EWMA schemes, but once a neural-network or fuzzy layer is added, the joint distribution becomes analytically intractable. Monte Carlo simulation is the standard approach for estimating ARL0 and ARL1 of complex hybrid schemes.

Yes. Healthcare quality control, financial transaction fraud monitoring, and service-level management all use SPC variants. Hybrid designs that handle non-normal data (via transformation or nonparametric charts) and autocorrelation (via residuals from a time-series model) are especially relevant in service contexts where the independence assumption of classical SPC is routinely violated.

As a practical guideline, Phase I should include at least 100–200 individual observations (or 20–25 rational subgroups) to obtain stable estimates of the mean and variance. For multivariate or adaptive components, more data are typically required. Deploying with fewer observations leads to underestimated control limits and an inflated false-alarm rate in Phase II.