Douglas C Montgomery Statistical Quality Control

Setting Up Control Charts That Actually Work

I spent three years fighting with control charts before I stopped trying to make them look pretty and started treating them like diagnostic tools. The textbook version of Statistical Quality Control is clean. Real production lines are not. What follows is the stuff they don't always emphasize in the chapters. Start with the method, not the theory. When you're pulling data from a machine that runs 24/7, you collect subgroups, calculate the centerline and control limits, plot the points, and interpret. That's the surface routine. The part that matters is how you define those subgroups. A subgroup isn't just a convenient batch of data. It represents a slice of common-cause variation, and if you construct it wrong, the chart tells you nothing useful. The textbook says rational subgroups should minimize between-subgroup variation while maximizing within-subgroup variation. In practice, this means your subgroup should span a short, consistent time window where the process is running under the same conditions. If your CNC mill changes tools every 45 minutes and you group samples taken across a tool change, your R-chart will be inflated and your X-bar chart will be blind to real shifts.

Douglas C Montgomery Statistical Quality Control

Montgomery's framework, as laid out in his widely used text, covers X-bar and R charts, X-bar and S charts, individual and moving-range charts, CUSUM and EWMA approaches, attribute charts, and process capability analysis. The structure is systematic. You identify the quality characteristic, determine sampling strategy, estimate parameters, set control limits at three standard deviations, monitor over time, and respond to out-of-control signals with root-cause investigation. Here is what beginners miss. The three-sigma limits are not a universal rule. They are a convention derived from the assumption of normality and the desire for a false alarm rate of roughly 0.27 percent per point when the process is in control. That convention breaks down in several scenarios I have encountered firsthand. Pitfall one: assuming symmetry. If your data are skewed, which happens constantly with things like defect counts, cycle times, or contamination measurements, the upper and lower control limits will not be equidistant from the centerline in a meaningful way. Montgomery addresses this with log-transformations and non-normal base distributions, but practitioners often skip that step and end up with a chart where the lower limit is negative, which is physically impossible for many characteristics.

Pitfall two: treating a single shift as a signal of trouble. A run of eight points on one side of the centerline is a standard Western Electric rule. But in a process with deliberate drift or seasonal cycling, that rule fires constantly and you spend your time investigating ghosts. I worked on a film-coating line where the coating weight drifted upward over a six-hour shift cycle due to solvent evaporation rates. The X-bar chart flagged the drift as out of control every single shift. The fix was not adjustment. It was recognizing the drift as assignable-cycle variation and building it into the control strategy rather than fighting it. Let me walk through a real situation. I was monitoring a injection molding process where we tracked part dimensions on a Cpk baseline study that read 1.67, which looks healthy until you look at the control chart over time. The X-bar chart showed occasional excursions past the upper limit, maybe two per month, but the process average had been slowly trending upward over a three-week period between those spikes. The R-chart was stable. This is a classic case where the Shewhart chart detects the break but misses the drift. The workaround was switching to a CUSUM chart for the same dataset. The CUSUM accumulated the small deviations and flagged the trend about ten days earlier than the X-bar chart would have caught it. That meant we adjusted the mold temperature before we started producing out-of-spec parts instead of after. So here is the practical breakdown of how to actually implement this.

First, select your quality characteristic and measurement system. If your gauge repeatability and reproducibility study shows a %GRR above ten percent of the tolerance band, your control charts are measuring your gauges, not your process. This is more common than people admit. I once saw an entire SPC program thrown out because the measurement system was contributing forty percent of the observed variation. The fix was upgrading to a laser micrometer and re-running the control chart, which immediately revealed patterns that were previously hidden by noise. Second, determine subgroup size and sampling frequency. For variables data, subgroup sizes of four to five are standard. Smaller and you lose sensitivity. Larger and you dilute the impact of real shifts. Sampling frequency depends on process dynamics. If your process stabilizes within minutes of a change, sampling every hour is wasteful. If it takes hours to reach steady state, sampling every ten minutes gives you redundant data and a false sense of precision. Third, calculate initial control limits from a phase one dataset. This dataset should contain at least twenty to twenty-five subgroups. Fewer and your estimates of sigma are unstable. Montgomery recommends using the standard estimators, either R-bar divided by d2 or S-bar divided by c4, depending on subgroup size. For subgroups of ten or more, S-chart based limits are preferred because R loses efficiency as subgroup size increases.

Fourth, review the plotted chart for non-random patterns before declaring the process in control. Run tests for trends, cycles, mixtures, and stratification, not just the standard Western Electric or Nelson rules. A histogram of your subgroup averages can reveal bimodality that a control chart alone might mask if the two modes happen to fall within the control limits. Fifth, move to phase two monitoring with periodic limit revision. Control limits are not permanent. When you implement a process improvement, recalculate the limits. Leaving old limits in place after a genuine improvement makes the chart less sensitive to future degradation. I see this all the time. A team reduces process variation by twenty percent, gets a new Cpk target, and never updates the control chart. Six months later they are baffled because the chart looks overly reactive. It is not reactive. It is just calibrated to the wrong distribution. Process capability analysis is where most organizations go wrong. Cpk and Ppk are not interchangeable. Cpk uses within-subgroup variation and answers the question of what the process can achieve under controlled conditions. Ppk uses overall variation and answers what the process has actually delivered. The gap between them tells you about stability. If Cpk is 1.8 and Ppk is 1.1, you have assignable-cause variation inflating your long-term performance. Fixing that gap is almost always more valuable than chasing a marginal Cpk improvement.

Attribute charts come up when you are dealing with defectives or defects rather than measurements. The p-chart, np-chart, c-chart, and u-chart each have assumptions that are routinely violated. The p-chart assumes a constant sample size or uses variable-width limits. The c-chart assumes a constant inspection area. I once ran a c-chart on defect counts from a woven fabric line where the roll width varied by fifteen percent between products. The chart was completely useless. Switching to a u-chart normalized for area fixed it immediately. There are limitations worth being honest about. Montgomery's framework assumes stationarity for valid inference. Many modern processes are non-stationary by design. Semiconductor lithography, chemical batch reactors, and additive manufacturing all have inherent drift and transition phases. SPC in those environments requires either stratified analysis, time-weighted charts, or alternative modeling approaches. CUSUM and EWMA help with small shifts but are slower to detect large ones than Shewhart charts. They also require more judgment in setting the k parameter and decision interval. Get those wrong and you are either chasing noise or missing signals. Another blunt reality: control charts do not replace process knowledge. I have seen organizations treat SPC as a compliance exercise, collecting data and plotting points without any operator understanding of what the signals mean. The chart becomes decoration. The most effective SPC programs I have encountered are the ones where operators can look at a chart and explain what happened, not just flag a point and call engineering.

If you are starting from scratch and need a reference, Montgomery's Statistical Quality Control remains the standard textbook used in both academic courses and industrial training programs. The material is comprehensive, covering everything from basic Shewhart charts to advanced multivariate control charts and acceptance sampling plans. You will find it through academic publishers, technical book retailers, and various online platforms. The bottom line is straightforward. Control charts are tools, not solutions. They reveal variation. They do not eliminate it. The value comes from what you do after the chart tells you something is wrong, not from the act of charting itself. Start with a clear understanding of your process, build rational subgroups, choose the right chart type for your data structure, and be willing to abandon the textbook approach when the process demands it.