Advancing QE Skills (13) | Control Chart Selection Decision Tree: Continuous, Discrete, and Small Batch Applications

1. A One-Size-Fits-All Approach Leads to a Year of False Alarms

A certain electronic component company's SPC implementation plan was very "unified": all key characteristics were uniformly plotted on X̄-R charts, subgroups always consisted of 5 pieces, and all out-of-control rules were applied in full. After a year of operation, the only result was complaints.

The automated inspection line produced 200 pieces per batch, and the engineer selected 5 pieces from the finished products to form a subgroup. The control limits were calculated to be very narrow and sensitive, leading to daily alarms, and the site gradually learned to "look at the chart first before taking action." The injection molding line produced over a dozen different shell types, each with different nominal dimensions, but they were all mixed into the same chart for limit calculation, making the control limits so wide that even abnormalities could not be detected. The visual inspection station recorded the number of nonconforming products out of 200 sampled pieces per batch, but this data was forced into an X̄-R chart. The inspector had to temporarily fabricate 5 "dimension values" to meet the data entry requirements.

In all three scenarios, the same mistake was made—incorrect chart selection. Whether the data is variable or attribute, and whether the subgroup size is 1, 5, or 200, determines which chart to use. If the chart is selected incorrectly, all subsequent out-of-control judgments, capability indices, and improvement measures will be based on the wrong denominator.

2. The Fundamental Logic of Selection: Two Axes, Two Premises

Axis One: Data Type. Variable data (such as dimensions, weight, temperature, time, which can take continuous values) uses mean-based charts; attribute data (such as pass/fail, number of defects) uses count-based charts. The statistical distributions of these two types are different: variable data is approximately normally distributed, while attribute data follows a binomial or Poisson distribution. The control limit formulas and sample size requirements are different and cannot be interchanged.

Axis Two: Subgroup Structure. Single values (n=1) use I-MR charts; small subgroups (n=2～9) use X̄-R charts; large subgroups (n≥10) use X̄-S charts. The reason is that the range R only captures the maximum and minimum values within the subgroup, and the larger the n, the more information is wasted: when n=2, R is as efficient as the standard deviation S; when n=5, it is about 0.95 efficient; when n=10, it drops to around 0.85; and when n=15, it is less than 0.75. In engineering, n≥10 is the threshold for switching to S charts.

Premise One: Control limits must come from the process itself when it is in control. Tolerance limits, target standard deviations from technical agreements, or the standard deviation from a "best-performing month" are not valid sources for control limits. Using tolerances as control limits will make the chart always quiet—until the entire batch is scrapped.

Premise Two: Count-based charts have a normal approximation threshold. For p and np charts, the sample size for each point must satisfy np̄ ≥ 5 (ideally, the number of nonconformities should be above 5, as the approximation is poor when the sample size is small and p is low). For c and u charts, c̄ ≥ 5 is required. If the threshold is not met, the shape of the control limits is incorrect.

3. Five-Step Selection Decision Tree

Step One: Determine the Data Type. Ask, "Is this characteristic measured or counted?" Measured data is variable, and counted data is attribute. If this step is wrong, everything that follows will be wrong.

Step Two: Assign Variable Data Based on Subgroup Size. n=1 (single-piece flow, one result per batch, extremely high inspection cost) → I-MR, with at least 20～25 individual points for limit calculation. 2 ≤ n ≤ 9 → X̄-R, using d₂, D₃, and D₄ constants from tables. n ≥ 10 or varying subgroup sizes per batch → X̄-S, using B₃ and B₄ constants from tables. Regardless of the type, the subgroup must consist of continuous production pieces from the same equipment, mold, tooling, batch of material, shift, operator, and parameters.

Step Three: For Attribute Data, Identify the Monitoring Object and Check if the Sample Size is Constant. If the object is "the number or rate of nonconforming units" → np chart (constant n) or p chart (varying n); if the object is "the number of defects or defects per unit" → c chart (constant inspection area/unit count) or u chart (varying). Criterion: If the sample size fluctuates by more than ±25%, p or u charts must be used, and control limits must be calculated for each point, resulting in a stepped appearance. Using a single set of flat limits based on the average sample size is a common misuse of count-based charts. For p charts, the sample size for each point should generally be at least 50～100 units to maintain sensitivity.

Step Four: Determine if There is Enough Data to Build a Conventional Chart. The criterion is that there are at least 20～25 data points (approximately 100 or more individual points) under the same conditions. If this cannot be achieved, it is better to switch to a short-cycle approach: for products of the same family with only nominal size differences, use standardized Z charts or DNOM deviation charts. Entry Criteria include three conditions: the standard deviation ratio of different products in the same family should be between 0.5～2 (to ensure similar variability); MSA's %GRR ≤ 30%; and the measurement instrument's resolution ≤ 10% of the tolerance band. After accumulating 20～25 points across different products, limits can be established, and both sensitivity and comparability are restored.

Step Five: Establish Limits, Verify, and Set Out-of-Control Rules. Use 20～25 data points to establish limits, and recalculate after removing special points with known causes. Verification Criterion: All 25 points should fall within the limits, or no more than 1 point should fall outside the limits in 35 consecutive points. Out-of-control rules must match the chart type: X̄ charts can use the full set of rules (e.g., 7 consecutive points on one side, 6 consecutive points increasing or decreasing, 2 out of 3 points exceeding 2σ); I-MR and short-cycle charts, due to the autocorrelation of moving ranges/standardized data, should only use a few rules such as "point out of limits + clear trend," as applying the full set would significantly increase the false alarm rate.

4. Common Misconceptions

Misconception One: Using Tolerances as Control Limits. This mistake treats "customer requirements" as "inherent process variation." The result is that the chart never alarms, and SPC degrades into a mere formality; real issues often go undetected until they reach the assembly line.

Misconception Two: Changing Subgroup Size Without Adjusting Constants. Switching from n=5 to n=7 but continuing to use the d₂ and D₄ values for n=5 will systematically narrow or widen the control limits. If the subgroup size, sampling method, or measurement instrument changes, the control limits must be recalculated.

Misconception Three: Not Recalculating Limits for Count-Based Charts When Sample Size Varies. The limits for p and u charts depend on the sample size at each point. Using limits calculated from the average sample size for points with small sample sizes will lead to frequent false alarms, while using flat limits for points with large sample sizes will result in missed alarms.

Misconception Four: Using c Charts When c̄ < 5. At this point, the Poisson distribution is severely right-skewed, and the normal approximation fails, leading to negative lower 3σ limits, which are incorrect. Solutions include increasing the inspection area or sample size, combining multiple units and switching to u charts, or using tools like cumulative sum (CUSUM) or exponentially weighted moving average (EWMA) that are more tolerant of distribution.

Misconception Five: Mixing Different Specifications on One Chart. Combining data from different nominal sizes on one chart treats the differences between products as within-subgroup variation, leading to overly wide control limits that mask true abnormalities. The correct approach is to standardize the data (using Z charts or DNOM charts) before combining, not to mix and plot directly.

Misconception Six: Confusing "Units" and "Defects." c charts monitor the number of defects, while u charts monitor the number of defects per unit. Plotting c charts based on "defects per unit" when the inspection area or unit count varies is equivalent to losing the denominator, leading to incorrect conclusions.

5. Self-Check List

• Each monitored characteristic has a clear selection basis (data type + subgroup size), not a one-size-fits-all template for the production line.
• Subgroups are taken from continuous pieces under "six same" conditions, and the actual subgroup size matches the size used for limit calculation, with constants from the table checked according to the current n.
• For count-based charts, the monitoring object (nonconforming units/defects) and sample size constancy have been confirmed, and each point meets np̄ ≥ 5 or c̄ ≥ 5. If the sample size fluctuates by more than ±25%, limits are calculated for each point.
• The data for limit calculation consists of at least 20～25 groups (or accumulated across different products to meet the requirement), with special cause points removed and verified through "25 points all within limits" or "35 points with ≤ 1 point out of limits."
• Out-of-control rules match the chart type, and the full set of rules is not blindly applied to I-MR and short-cycle charts. After changing subgroups, measurement instruments, processes, or products, control limits have been reassessed.

If the chart is selected incorrectly, even the most accurate out-of-control rules are useless.

Knowledge code: 6.3.1

Version: v20260923

Author: QTank QTank is dedicated to providing systematic knowledge, methodologies, and practical tools for quality management professionals, helping companies continuously improve their quality capabilities.