Question 2 of 6: SPC Sampling, EPC vs. SPC, and Why EWMA/CUSUM Beat the Shewhart Chart

Question 2: SPC Sampling, EPC vs. SPC, and Why EWMA/CUSUM Beat the Shewhart Chart (20 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

(a) Rational sampling, run length, ARL/ATS, and where zone (run) rules apply

Samples for SPC should be drawn using the principle of rational subgrouping: each subgroup should be formed so that the variation within it is due only to common (chance) causes — items produced close together in time under the same conditions — while the variation between subgroups is given the best opportunity to reveal a special cause. In practice this means sampling consecutive or near-consecutive units rather than units spread across a shift, choosing a sampling interval short enough that a process shift is caught before too much bad product accumulates but long enough to be economical, and keeping subgroup size and interval consistent so the chart's statistical properties (control limits, run-length behaviour) stay valid over time. Samples must also be representative and unbiased — not cherry-picked — or the estimated control limits will not reflect the true process.

A run length is the number of samples plotted on a control chart before a point signals out-of-control. Because each sample's outcome is (for a Shewhart chart) an independent Bernoulli trial with probability $p$ of falling outside the limits, the run length follows a geometric distribution and the average run length is $ARL=1/p$: with the standard $3\sigma$ limits and a normally distributed, in-control statistic, $p=0.0027$ so $ARL_0\approx370$ samples between false alarms; after a real shift, $p$ rises and $ARL_1$ falls, giving the expected number of samples needed to detect that shift. The average time to signal converts this into calendar time: $ATS=ARL\times h$, where $h$ is the fixed sampling interval (or, more generally, the expected time between samples if the interval is itself variable).

Zone (Western Electric run) rules — e.g., 2 of 3 consecutive points beyond $2\sigma$, or 4 of 5 beyond $1\sigma$, on the same side of the centre line — are valid supplementary sensitizing rules only for charts whose successive plotted points are statistically independent and (at least approximately) normally distributed, which describes the ordinary Shewhart charts: $\bar X$, $R$, $s$, $p$, $np$, $c$, $u$. They are not appropriate for EWMA or CUSUM charts, because those statistics are deliberately built from a moving weighted history of previous samples ($z_t=\lambda x_t+(1-\lambda)z_{t-1}$ for EWMA; a cumulative sum for CUSUM) — consecutive plotted points are therefore strongly positively autocorrelated by construction. Applying independence-based zone rules to an autocorrelated statistic produces a false-alarm rate far higher than the nominal one, because several consecutive points near a zone boundary are not independent evidence of a shift, only evidence that the smoothing filter has not yet forgotten the last one.

(b) Engineering process control (EPC) vs. statistical process control (SPC)

SPC treats a deviation from target as a signal to investigate and permanently remove its assignable cause — it does not touch the process; it only monitors it and calls for human intervention when the output looks statistically unusual. EPC (also called automatic process control or feedback control) instead treats every deviation from target as a disturbance to be compensated automatically and continuously by adjusting a manipulable process input (e.g., a valve position, a feed rate), using a control algorithm such as a proportional-integral (PI) controller or a minimum-mean-square-error predictor — it keeps the output close to target without ever asking why the disturbance occurred.

In a real production setting the two are complementary rather than competing: EPC is used to regulate a continuous or semi-continuous process that experiences frequent, small, often unavoidable drift (e.g., feedstock composition variation in a chemical process), keeping output variance low without operator intervention; SPC is then applied on top of the EPC-compensated output (or on the compensating adjustments themselves) to detect when the disturbance has grown, changed character, or become something the controller can no longer economically absorb (e.g., a sensor failure, a step change in raw-material quality) — a signal that calls for the kind of root-cause investigation and permanent fix that automatic compensation alone can never provide, and that lets the plant tell the difference between "the controller is doing its job" and "something is now genuinely broken."

EWMA is usable for both roles because it is, mathematically, the same exponential filter used in engineering process control: $z_t=\lambda x_t+(1-\lambda)z_{t-1}$ is a first-order recursive (IIR) filter that is exactly the minimum-mean-square-error one-step-ahead predictor for an integrated-moving-average disturbance, which is precisely the model EPC feedback controllers are designed around — so EWMA can serve directly as the forecast term inside a feedback control law. At the same time, EWMA has a well-defined statistical run-length distribution when its output is compared against control limits, making it a legitimate SPC monitoring chart in its own right, particularly effective (unlike the memoryless $\bar X$ chart) at detecting the small, sustained shifts that EPC is meant to compensate but that SPC still needs to be able to flag when compensation is no longer adequate.

(c) Why $\bar X$ struggles with small shifts, and the EWMA paradox

The $\bar X$ chart's plotted statistic uses only the current sample — it is memoryless. For a shift of size $\delta\sigma$ in the mean, the probability that any one sample's average falls outside the $3\sigma$ limits depends only on that shift and the current sample size; for a small shift, this probability is barely above the in-control false-alarm rate, so the chart must wait, on average, for a very large number of independent samples ($ARL_1=1/p$, which stays close to $ARL_0\approx370$ for small $\delta$) before a point happens, by chance, to land beyond the limit. Nothing about a small, persistent shift accumulates evidence from one sample to the next on a Shewhart chart — each sample restarts the detection problem from zero.

EWMA and CUSUM are better for small sustained shifts precisely because they are not memoryless: EWMA's $z_t$ carries forward a geometrically weighted average of all previous samples, and CUSUM's statistic is a running cumulative sum of deviations from target, so a small but persistent shift steadily pushes the plotted statistic away from the centre line, sample after sample, rather than requiring one lucky extreme observation. This lets both charts achieve a much smaller $ARL_1$ for small shifts than the $\bar X$ chart, at a comparable in-control $ARL_0$.

The apparent paradox resolves once "detecting on the very next sample" and "having a small average run length" are recognized as different questions. Immediately after a shift occurs, EWMA's statistic $z_t$ is still dominated by its pre-shift history (the weight $(1-\lambda)$ on the old, unshifted average is large unless $\lambda$ is close to 1), so the chance that $z_t$ alone jumps outside the control limits on that very first post-shift sample is genuinely small — smaller, in fact, than the $\bar X$ chart's one-sample detection probability, because $\bar X$ reacts fully and immediately to whatever the current sample shows while EWMA is deliberately slow to react to any single point. But EWMA does not need to detect on sample one: because each subsequent sample keeps nudging $z_t$ further toward the new (shifted) mean, and because these nudges accumulate rather than reset, the expected number of samples to signal (the ARL, not the one-shot probability) ends up far smaller than the $\bar X$ chart's, whose memoryless per-sample probability never improves as more post-shift samples accumulate. In short: EWMA trades a low probability of an immediate, one-shot detection for a much lower expected time-to-detection overall, because it is the only one of the two statistics that actually remembers the shift is still there.