Question 2 of 6: SPC Concepts, Control-Chart Sensitizing Rules, and an Individuals/Moving-Range Chart

Question 2: SPC Concepts, Control-Chart Sensitizing Rules, and an Individuals/Moving-Range 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) Engineering vs. statistical process control; off-line vs. on-line QC; SPC tools; FMEA/FTA

Engineering process control (EPC) reacts to every observation of a process (typically one that drifts continuously, such as a chemical or temperature process) by making a compensating adjustment (e.g. feedback/feed-forward control) to keep the output on target — it treats the process output as inherently non-stationary and actively cancels disturbances. Statistical process control (SPC), by contrast, assumes a well-designed process should run in a stable, statistically predictable state, and it monitors the process with control charts to detect when an ASSIGNABLE cause has disturbed that stable state, triggering investigation and correction of the root cause rather than a routine compensating move; SPC deliberately does NOT adjust the process in response to ordinary common-cause variation, since doing so (over-adjustment/tampering) actually increases output variance. The two approaches are complementary rather than competing: EPC and SPC are often combined on the same process (EPC to hold a naturally drifting mean on target, SPC monitoring the EPC-adjusted output/adjustments for the onset of an assignable cause the controller cannot compensate for).

Off-line quality control refers to quality-improvement activity that happens before full-scale production — product and process design, designed experiments, robust parameter design, supplier and process qualification — where changes are cheap because nothing is yet running at volume. On-line quality control refers to the monitoring and control activity that happens during actual production — control charts, acceptance sampling, in-process adjustment — where the process is already committed and changes are comparatively expensive and time-critical. Effective quality management uses off-line methods to design a process that is inherently capable, so that on-line control only has to detect rare, genuine upsets rather than fight a poorly designed process every shift.

Two basic SPC problem-solving tools: (1) the cause-and-effect (Ishikawa/fishbone) diagram, which organizes candidate causes of a defect into major branches (commonly the "6 M's": Man, Machine, Method, Material, Measurement, Environment/Mother-Nature) to structure a brainstorming/root-cause investigation; and (2) the Pareto chart, a bar chart of defect/cause frequencies sorted in descending order with a cumulative-percentage line overlaid, applying the 80/20 principle to focus improvement effort on the "vital few" causes that account for most of the loss rather than spreading effort evenly across the "trivial many."

FMEA (Failure Modes and Effects Analysis) is a systematic, bottom-up, tabular procedure that walks through each component/step of a design or process, lists its potential failure modes, the effects of each failure on the system and the customer, its likely causes, and current controls, then scores each failure mode by severity, occurrence and detection (multiplied into a Risk Priority Number) to prioritize corrective action before the failure occurs in the field. FTA (Fault Tree Analysis) is the complementary top-down technique: starting from a single defined undesired "top event" (e.g. a specific catastrophic failure), it works backward through AND/OR logic gates to identify every combination of lower-level component failures or human errors that could cause that top event, which is especially suited to analyzing rare, high-consequence system failures with multiple interacting causes that FMEA's component-by-component sweep can miss.

(b) Control charts; R chart vs. S chart

A control chart is a time-ordered plot of a sample statistic (e.g. $\bar X$, $R$, $S$, $p$) against sample number, with a centre line at the statistic's expected in-control value and upper/lower control limits (conventionally $\pm3$ standard deviations of the plotted statistic) beyond which a plotted point signals that the process has very likely been disturbed by an assignable cause. It is used to distinguish common-cause (inherent, chance) variation, which the process should simply be left to run under, from special/assignable-cause variation, which should trigger investigation and correction — and, once a process is shown to be in a state of statistical control, to estimate its stable parameters (mean, standard deviation) for capability analysis and for designing future monitoring/sampling plans.

Both the R chart (range, $R=X_{max}-X_{min}$ within a subgroup) and the S chart (sample standard deviation within a subgroup) monitor process variability; they differ in statistical efficiency and how they scale with subgroup size. The range uses only the two extreme observations in a subgroup and ignores the interior data, so its relative efficiency as an estimator of $\sigma$ falls off rapidly as subgroup size $n$ grows past about 8–10; the sample standard deviation uses every observation and remains an efficient estimator at any $n$. The S chart is generally preferable, especially for $n>8$–10 or whenever the extra computation is not a burden (routine with modern data-collection software), because it makes fuller use of the sample information and gives a more statistically efficient (lower-variance) estimate of the true process standard deviation for the same sample size. The R chart remains attractive mainly for very small, manually-recorded subgroups ($n\le5$, e.g. hand-charted shop-floor data) where the range is trivial to compute by eye and its slightly lower efficiency is an acceptable trade for simplicity.

(c) Effect of combined zone/sensitizing rules on the ARL

Sensitizing (zone) rules supplement the basic 3-sigma out-of-limits test by dividing the region between the centre line and each control limit into three "zones" (each one standard deviation of the plotted statistic wide: A, B, C moving outward) and flagging additional non-random patterns — e.g. 2 of 3 consecutive points in Zone A or beyond, 4 of 5 consecutive points in Zone B or beyond, 8 consecutive points on one side of the centre line, or a run of points steadily trending in one direction. Because each additional rule gives the chart another independent (or nearly independent) chance to signal on any given sample, using several zone rules simultaneously makes the chart much more sensitive to small, sustained shifts that a lone-points-beyond-3-sigma test would take many samples to detect — the out-of-control ARL drops substantially (fewer samples needed, on average, to catch a real disturbance) compared with the basic rule alone.

The cost of that added sensitivity is on the in-control side: each rule, run continuously, carries its own (small) false-alarm probability, and because the rules are evaluated simultaneously every sample, their false-alarm probabilities essentially add (the rules are not mutually exclusive events on any one sample), so the overall probability of a false signal per sample rises and the in-control ARL falls from the basic chart's roughly 370 (for standard 3-sigma limits, $\alpha=0.0027$) to something considerably lower — commonly cited combined-rule packages fall in the 91–150 range for in-control ARL, depending on exactly which rules are stacked. This is the fundamental sensitivity/false-alarm trade-off of adding sensitizing rules: they buy faster detection of real, sustained process shifts at the cost of more frequent unnecessary investigations of a process that has not actually changed, so a practitioner should add only as many rules as the investigation cost of a false alarm can tolerate, and should be aware the rules are not statistically independent of each other or of the basic 3-sigma test.

(d) Individuals and moving-range charts for extrusion-sheet weight

Given. One weight measurement is recorded every half hour (a single cross-section per sampling instant, i.e. rational subgroup size $n=1$), so the appropriate chart pair is the individuals ($X$) and moving-range ($MR$) chart, not an $\bar X$/$R$ or $\bar X$/$S$ chart (those require $n\ge2$ per subgroup). The 20 recorded weights, in pounds, are given in the table below.

Find. Trial and (if needed, revised) control limits for the individuals and moving-range charts, and the resulting in-control estimates of the process mean $\hat\mu$ and standard deviation $\hat\sigma$.

Approach. Compute $\bar X$ and the average moving range $\overline{MR}$ (of the 19 successive absolute differences), form trial 3-sigma limits using the individuals-chart constants ($d_2=1.128$, $D_3=0$, $D_4=3.267$ for a moving range of span 2), plot both charts, and drop/re-estimate only if a point signals out of control.

1. Centre line and moving ranges. $\bar X=\dfrac{\sum X_i}{20}=\dfrac{3512}{20}=175.6$ lb. The 19 successive moving ranges are $MR_i=|X_i-X_{i-1}|$: 5, 5, 9, 1, 4, 2, 14, 11, 5, 23, 32, 2, 14, 5, 13, 18, 34, 28, 14, giving $\overline{MR}=\dfrac{239}{19}=12.58$ lb.
2. Trial limits. $$UCL_{MR}=D_4\overline{MR}=3.267(12.58)=\boxed{41.10\ \text{lb}},\quad LCL_{MR}=D_3\overline{MR}=0.$$ $$UCL_X=\bar X+3\dfrac{\overline{MR}}{d_2}=175.6+3\left(\dfrac{12.58}{1.128}\right)=\boxed{209.05\ \text{lb}},\quad LCL_X=175.6-3\left(\dfrac{12.58}{1.128}\right)=\boxed{142.15\ \text{lb}}.$$
3. Check for out-of-control points. All 20 individual weights (min 148, max 202) fall strictly between $LCL_X=142.15$ and $UCL_X=209.05$, and all 19 moving ranges (max 34) fall below $UCL_{MR}=41.10$. No point signals on either chart and no non-random pattern is evident, so no revision is needed — the trial limits stand as the final control limits.
4. In-control process-parameter estimates. Since both charts are in control as drawn, $$\hat\mu=\bar X=\boxed{175.6\ \text{lb}},\qquad \hat\sigma=\dfrac{\overline{MR}}{d_2}=\dfrac{12.58}{1.128}=\boxed{11.15\ \text{lb}}.$$