QE Skill Enhancement (14) | CUSUM and EWMA: Sensitive Detection of Minor and Gradual Offsets

1. Why Did No One Look at the Chart After Seven Days

A machining workshop in an automotive parts company once implemented a "minor offset monitoring" system: they added an EWMA chart for a key dimension sensitive to tool wear. The first three days were lively, with six alarms reported within a week. Each time the process engineer stopped the line to re-measure, the dimensions were found to be normal. Starting from the seventh day, the operators began to treat the alarms as background noise. Three months later, the chart was removed from the on-site bulletin board. Six months later, during the quarterly inventory, it was discovered that the process center of this dimension had drifted by about 0.8σ, and the scrap rate had more than doubled.

Another version of the same scenario occurred in an electronics company where the control limits were set too loosely. The chart did not trigger any alarms in a year, leading the engineers to believe that "this characteristic is very stable."

Both failures were not due to the choice of charts—neither the mechanism of EWMA nor CUSUM is flawed. The mistake lies in setting parameters based on intuition and not validating their actual effectiveness after deployment. A control chart is essentially a detector: it is crucial to specify the size of the offset to be detected and the acceptable number of false alarms before setting the parameters to meet these requirements. This article differs from the previous one in that it focuses not on whether to use these charts, but on how to set the parameters, how to validate them, and when a recalculation is necessary.

2. Key Principles: Sensitivity is a Curve, Not a Switch

The decision logic of conventional control charts is based on whether the current point exceeds the control limits. Their response to offsets is characterized by the Average Run Length (ARL): ARL₀ refers to the average number of points before a false alarm is triggered when the process is stable, and ARL₁ refers to the average number of points before an offset of δ is detected. The entire design work of these charts involves balancing ARL₀ and ARL₁—ARL₀ represents the inspection cost at the site, while ARL₁ reduces the risk of undetected offsets.

EWMA adds an exponential forgetting weight to historical data: Z equals λ times the current point value plus (1 - λ) times the previous cumulative value, where λ is the weight of the new data. The control limits for steady-state are calculated as L times σ times the square root of λ divided by (2 - λ). Two important inferences to remember are: the smaller the λ, the longer the memory (approximately 19 points for λ = 0.1, and 9 points for λ = 0.2), making it more sensitive to gradual drifts but slower to react to sudden jumps; the steady-state control limits are much tighter than those of a Shewhart chart, with about ±1σ for λ = 0.2 and L = 3. Therefore, the judgment of "point out of control" cannot be based on the intuition from a Shewhart chart, and a single point out of control must be re-measured.

CUSUM directly accounts for deviations: each point adds the part exceeding the target value by ±k to C⁺ and C⁻, respectively. If a cumulative value becomes negative, it is reset to zero, indicating that the evidence in that direction has been cleared. k is the reference value, determining sensitivity to the size of the offset; h is the control limit, determining the false alarm budget. These parameters are equivalent to a V-shaped template: the slope of the template's arms is k, and the distance from the vertex to the current point equals h divided by k. The unique advantage of CUSUM is its ability to trace back to the start of the offset—when the cumulative value begins to monotonically increase, it can directly align with events like "a new batch of materials was used" or "a new tool was installed." This time evidence is something EWMA cannot provide.

3. Practical Steps: Five Steps to "Design" Parameters

Step 1: Define the Requirements with Two Numbers. δ* is the smallest offset that must be detected (in units of σ, e.g., 0.8σ); T is the maximum time from the occurrence of the offset to the alarm, converted into the number of sampling points. T is derived from the acceptable number of outflows: if a workshop produces 200 pieces per hour and samples one piece every 30 minutes, and the requirement is that no more than 400 defective pieces should flow out, then T should not exceed 2 hours, or 4 points. Criterion: If δ* is greater than or equal to 2σ and T is relaxed (more than 10 points), conventional control charts with abnormal criteria are sufficient, and there is no need to use these two charts. The true value range for these charts is 0.5σ to 1.5σ.

Step 2: Back-Check Parameters Based on ARL Targets. First, set the false alarm budget and convert ARL₀ into the number of times the site can tolerate: with three shifts, each shift checking the chart once, the requirement is no more than one false alarm per week. With approximately 112 points sampled per week, ARL₀ should be at least 300 to 500. For CUSUM, k is set to half of δ*; if you are concerned about a 1σ offset, set k = 0.5σ. h is determined by back-checking ARL₀; with h = 4σ, ARL₀ is only about 170 points, leading to multiple false alarms per week, which is too aggressive; with h = 5σ, it is about 460 points, a commonly used balance point. Increasing h further reduces false alarms but extends ARL₁. For EWMA, λ is matched to δ*: use λ = 0.05 to 0.1 for a 0.5σ offset, and λ = 0.1 to 0.2 for a 1σ offset. L is commonly set to 2.7 to 3.0; with λ = 0.2 and L = 3, ARL₀ is about 560 points. Criterion: After setting the parameters, recalculate ARL₁ to ensure that the average number of points from the occurrence of δ* to the alarm does not exceed half of T. If it exceeds T, the parameters are invalid, and you should return to adjusting λ, k, or the sampling frequency.

Step 3: Back-Test and Validate with Three Hard Criteria. Use at least 6 months of real historical data with no fewer than 100 points (preferably 150 points or more). ① Real abnormal segments (tool changes, batch changes, mold repairs, major maintenance) recorded in the ledger should be detected within approximately ±3 points of the alarm, and the alarm should occur at least one sampling cycle before the out-of-spec or outflow. ② The theoretical and actual false alarm counts during stable segments should be of the same order of magnitude, with the actual count not exceeding twice the theoretical count. If it does, it indicates that the data is not independent or the σ measurement is incorrect. ③ The offset start point traced back from the alarm should align with traceable on-site events. Unexplained alarms must be recorded with their causes—unexplained alarms are a signal of parameter mismatch. If any of the three criteria are not met, adjust h or λ and retest; if two rounds of retesting still fail, it indicates that the sampling frequency or measurement system capability cannot support 0.5σ level monitoring, and you should return to Measurement System Analysis (MSA) and sampling strategy.

Step 4: Verify σ Measurement and Data Independence. σ must come from a controlled state: use the moving range for single-value data (average moving range divided by 1.128) or the overall standard deviation, with no fewer than 25 baseline points and no abnormal segments. Criterion: If the sampling interval is too dense, causing data autocorrelation, the absolute value of the lag-1 autocorrelation coefficient should not exceed 0.2. Otherwise, the theoretical ARL will be severely distorted, and the actual false alarm rate will be much higher than the design value. There are three ways to handle this: increase the sampling interval, use the mean of each time window (batch mean chart), or establish a process model and monitor the residuals. In scenarios with fully automated online inspection, where data is generated every few seconds, this issue is almost inevitable.

Step 5: Document Response Rules and Version Control. After an alarm, follow a three-step response: re-measure 2 to 3 pieces within 30 minutes to rule out measurement issues; verify material batches, parameters, tooling, and maintenance records within 4 hours; stop the line only if the cumulative value continues to increase on one side. The cumulative value can only be reset after three consecutive points are re-measured and confirmed to be normal. Parameters, baseline dates, and version numbers should be recorded in the control plan, with a re-evaluation triggered every 6 months or after material changes, major repairs, mold repairs, or parameter adjustments.

4. Five Common Misconceptions

First, Only Calculating ARL₀ Once and Ignoring Data Autocorrelation. Dense sampling makes adjacent data highly correlated, and the "cumulative evidence" on the chart is essentially the same event being counted repeatedly. A nominal ARL₀ of 500 points may result in multiple alarms per day in practice.

Second, Using the Wrong σ Measurement. Using the subgroup standard deviation instead of the overall standard deviation underestimates the true variation, effectively tightening the actual thresholds for h and L, doubling the false alarms; conversely, using the tolerance band to estimate σ will make the chart always quiet.

Third, Resetting the Chart After Re-Measuring Only One Piece. Minor offsets are systematic, and re-measuring a single piece can easily "return to normal." The correct reset condition is to check if the cumulative value continues to increase on one side and use the number of consecutive re-measured points as the basis for resetting.

Fourth, Reducing h from 5σ to 4σ to Increase Sensitivity Without Considering the False Alarm Cost. ARL₀ will drop from about 460 points to about 170 points, nearly doubling the workload at the inspection station, and the site will often learn to ignore alarms within a few weeks.

Fifth, Never Re-Testing After Deployment. After changing raw materials, repairing molds, or adjusting parameters, the old baseline becomes invalid. Using a baseline from three months ago to judge the current state is like changing the ruler without knowing it.

5. Self-Check List

• Have you clearly specified the smallest offset δ* (in σ units) and the allowable alarm delay T (in points)?
• Is the ARL₀ target consistent with the number of inspections the shift can tolerate (converted into weekly or monthly false alarm counts)?
• Have you recorded the σ measurement (moving range/overall standard deviation), the number of baseline points (no fewer than 25 points), and the purity of the baseline?
• Have you back-tested with no fewer than 100 real historical data points, and have all three validation criteria been met?
• Have you verified data independence (lag-1 autocorrelation not exceeding 0.2)? If not, have you switched to batch mean or residual monitoring?
• Have the parameters, baseline date, and version number been frozen along with the control plan, and have the conditions for the next re-evaluation been clearly stated?

Sensitivity is designed, not adjusted.

Knowledge code: 6.3.1

Version: v20260924

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.