QE Capability Advancement (19) | Robust Design: Making Processes Insensitive to Variations

1. Parameters Tuned to "Best," but Mass Production Can't Maintain

A home appliance company was working on improving the warpage of injection-molded parts. The QE used full factorial experiments to optimize the mold temperature, holding pressure, and cooling time, reducing the average warpage from 0.80mm to 0.35mm and increasing the capability index Cpk from 0.9 to 1.45. The report was approved smoothly. However, three months later, the yield rate dropped back to the pre-improvement level. Upon review, it was found that the experiments were conducted using the same batch of raw materials, the same drying machine, and the same environmental humidity for a week. In mass production, the raw material batch changes every two weeks, and the workshop humidity varies by nearly 25%RH between the rainy and non-rainy seasons. The optimal point happened to be on a "steep slope" — a 3°C change in mold temperature caused a 0.15mm jump in warpage. They optimized the "mean performance," but did not address the more critical question: How sensitive are these parameters to variations?

Robust design aims to solve not just "how well the mean is tuned," but "how much performance drops when variations occur." These are two independent mathematical goals that must be optimized separately.

2. Principle: The Interaction Between Control Factors and Noise Factors Can Be Utilized

Factors affecting quality characteristics can be divided into two categories. Control factors are those that can be freely chosen in design (temperature, pressure, time, formula ratio); noise factors are those that are difficult or inappropriate to control in reality (raw material batch, environmental temperature and humidity, equipment wear, operator, aging). Traditional parameter tuning only adjusts control factors, assuming that noise remains constant.

The core discovery of robust design is that there is an interaction between control factors and noise factors. At one level of a control factor, noise variations can cause significant fluctuations in the response; at another level, the response curve is nearly flat in the direction of noise. Therefore, control factors can be used to "hedge" against noise — engineering efforts do not need to eliminate noise sources (which is often extremely expensive), but rather to make the process insensitive to them.

The quantification tool is the signal-to-noise ratio (S/N): it compresses the responses under multiple noise conditions at the same experimental point into a single metric, reflecting both "how far from the target" and "how large the variation." The selection of the formula depends on the quality characteristic — for target characteristics, use S/N = 10·lg(μ²/σ²); for smaller characteristics, use -10·lg(Σy²/n); for larger characteristics, use -10·lg[Σ(1/y²)/n]. The interpretation rule is simple: the higher the S/N, the better, and it is independent of whether the mean meets the target.

This leads to the most important analytical discipline in robust design: the two-step method. The first step is to use S/N as the response to find the control factor levels that maximize S/N, thereby reducing variation — the combination selected in this step may not have the mean on target. The second step is to choose a factor that "has no significant interaction with noise factors, has a minor impact on S/N, but a significant impact on the mean," and use it to pull the mean back to the target. Selecting the wrong adjustment factor can undermine the robustness achieved in the first step.

3. Five Practical Steps (Criteria for Each Step)

Step 1: Determine if Robust Design is Worthwhile. Criteria: noise-induced batch-to-batch or time-to-time variations account for more than 30% of the total process variation, or customer complaints/defects are concentrated after environmental changes, material changes, or equipment changes. If the process is dominated by a single control factor and noise effects are negligible, direct capability improvement or tolerance optimization should be performed without robust design. The measurement system must also be qualified: GR&R P/TV ≤ 10%, otherwise, measurement noise will be mistaken for process variation.

Step 2: Factor Classification and Level Setting. Select 3 to 5 control factors (each with 2 to 3 levels; use 2 levels in the screening stage, then use 3 levels for fine-tuning after determining the robust zone); select 2 to 4 noise factors. The hard criteria for noise factor levels: they must be set according to the actual variation range — raw materials should be set to ±3σ of the actual batch-to-batch differences, temperature and humidity should be set to the extreme values of the workshop throughout the year, and equipment wear should be set to the worst acceptable state. Laboratory conditions should not be used. Insufficient noise levels are the primary reason for the failure of robust design.

Step 3: Inner and Outer Array Design and Scale Calculation. Use an orthogonal array to arrange the control factors into an inner array, and use L4 or L8 to arrange the levels of noise factors into an outer array attached to each row of the inner array. Total number of experiments = number of inner array rows × number of outer array combinations × number of repetitions per point. Three quantitative thresholds: each inner array point must have at least 3 repetitions (otherwise, the degrees of freedom for S/N are insufficient, and the estimated value is dominated by noise); the number of outer array combinations must be ≥ 4 (L4), and there must be ≥ 2 noise factors to form interaction estimates; the total number of experiments should be controlled within 100 to 150 — exceeding this indicates that the factors are too broadly screened, and a fractional factorial (resolution ≥ Ⅳ) should be used for screening before proceeding to robust design. Randomize the experimental sequence, and record the raw material batches and environmental conditions row by row.

Step 4: Draw Conclusions Based on "Contribution Rate + Gain + Interaction" Criteria.

• Robust factors: perform ANOVA with S/N as the response; a factor with a sum of squares contribution rate ≥ 15% is considered a significant robust factor; stop pursuing causes when the cumulative contribution rate reaches 80%.
• Gain significance: the S/N gain after robust optimization must be ≥ 2dB to be meaningful for mass production; 1 to 2dB is marginal and requires more repetitions for verification; <1dB is considered noise and should not be reported as a result.
• Adjustment factors: the candidate factor must meet "contribution rate to S/N < 5%" and "no significant interaction with noise factors (p > 0.05)," while also "significantly affecting the mean (p < 0.05)." If any of these criteria are not met, it cannot be used to adjust the mean.
• Mean on target: after the second step of optimization, the target characteristic must fall within ±10% of the specification center, and the Cpk must be ≥ 1.33.

Step 5: Confirmation Experiments and "Parameter or Tolerance" Decision. Re-run the confirmation with the optimized combination: independent raw material batches ≥ 3 (safety/critical characteristics ≥ 5), the difference between the measured S/N and the model prediction ≤ 2dB, and the mean within the predicted interval. If any of these criteria are not met, return to Step 3 to check the number of repetitions and noise levels. Then make a decision: first adjust parameters, then tighten tolerances — the marginal cost of parameter adjustment is close to zero, while tightening tolerances means more stringent supplier material changes, inspections, and increased scrap. Only when parameter adjustments are exhausted and the capacity calculation shows that tightening tolerances will increase Cpk by ≥ 0.33 and the quality benefits outweigh the cost increases, should the tolerance route be taken. After stabilization, write the robust parameters into the control plan and establish monitoring points and rollback rules for key noise factors (raw material batches, environment).

4. Five Common Misconceptions

Misconception 1: Incorrect S/N Formula Selection. Using the S/N formula for larger or smaller characteristics for target characteristics changes the goal from "the closer to the target, the better" to "the larger, the better," often resulting in a "high-stability deviation" solution where the mean is severely off-target. Before selecting the formula, clearly specify the type of quality characteristic and have it confirmed in writing by both the process and the customer.

Misconception 2: Noise Factors Set to Fixed Levels. Conducting all experiments with a single batch of material or during a single time period degrades the outer array to a single point. At this point, no matter how the inner array is arranged, it is just a regular orthogonal experiment, and the "robustness" obtained is purely illusory. The outer array must realistically simulate variations, which is the fundamental difference from ordinary factor screening.

Misconception 3: Directly Taking the S/N Maximum as the Final Solution. The combination selected by robust factors usually does not have the mean on target. Without the second step of adjustment factors, the result will be "small variation but overall deviation"; conversely, if a factor with significant interaction with noise is used to adjust the mean, the variation reduced in the first step will be amplified again.

Misconception 4: Insufficient Repetitions or Repetitions Within the Same Batch. Each experimental point in the inner and outer arrays is only run once, making the S/N estimation purely a matter of chance; all repetitions are taken from the same piece or batch, underestimating within-batch variation and overestimating robustness. Criteria: each point must have ≥ 3 repetitions, and repetitions should cover both within-batch and between-batch levels.

Misconception 5: Believing Robust Design Can Replace Control Charts and Capability Improvement. Robust design reduces the process's sensitivity to noise, but it does not eliminate the process's inherent drift and sudden changes. After implementing robust parameters, control charts are still needed for monitoring; otherwise, systematic drifts such as tool wear and catalyst degradation will still reduce yield rates.

5. Self-Check List

• There is quantitative evidence that noise variations account for ≥ 30% (or complaints are concentrated after material changes/environmental changes), and GR&R P/TV ≤ 10%
• 3 to 5 control factors, ≥ 2 noise factors, and levels set according to actual variation ranges (batch-to-batch ±3σ, annual environmental extremes), not laboratory mild conditions
• Both inner and outer arrays are complete, each point has ≥ 3 repetitions, the outer array has ≥ 4 combinations, the total number of experiments is ≤ 150, the sequence is randomized, and batches and environments are recorded row by row
• Conclusions are based on three criteria: robust factor contribution rate ≥ 15%, S/N gain ≥ 2dB, adjustment factor contribution to S/N < 5% and no significant interaction with noise
• ≥ 3 independent batches are confirmed (the difference between predicted and measured S/N ≤ 2dB), the mean meets the target, and Cpk is calculated to be ≥ 1.33; the decision to adjust parameters or tighten tolerances is based on economic considerations

First, make the process insensitive, then adjust the mean back to the target.

Knowledge code: 6.4.2

Version: v20260929

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