The EWMA Control Chart: Catching Small Process Shifts Early

An EWMA control chart (exponentially weighted moving average chart) plots a weighted average of all measurements to date, weighting recent readings most heavily, so that small sustained shifts of 0.5 to 1.5 sigma in the process mean signal long before any single point breaks a limit. Introduced by S. W. Roberts in 1959, it closes the biggest blind spot of the classic Shewhart chart: slow drifts from tool wear, nozzle fouling, temperature creep, or a slipping calibration. Each plotted point carries the memory of everything before it, accumulating weak evidence no single measurement could provide.

Why small shifts slip past Shewhart charts

A Shewhart individuals chart judges each measurement in isolation against limits three standard deviations from the center line, which catches large abrupt failures well but is nearly blind to small persistent ones. If the mean drifts by one sigma, a Shewhart chart takes about 44 samples on average to signal; for a 0.5 sigma shift, over 150. A well tuned EWMA chart flags the same shifts in roughly 10 and 30 samples respectively.

Supplementary run rules such as the Nelson rules narrow the gap, but every added rule inflates the false alarm rate. The EWMA chart, part of the wider statistical process control toolkit, attacks the problem directly: instead of asking whether one point is unusual, it asks whether the recent weighted history is.

How the EWMA statistic is calculated

Each plotted value blends the latest reading with the previous plotted value: new EWMA = lambda times the new measurement, plus (1 - lambda) times the previous EWMA. The smoothing constant lambda (between 0 and 1) sets how fast old data fades: with lambda = 0.2 the newest reading carries 20 percent of the weight, the one before it 16 percent, and so on. The chart starts at the process target, with control limits at L standard deviations of the EWMA statistic either side of it; that standard deviation equals the process sigma times the square root of lambda divided by (2 - lambda). Exact limits are slightly narrower for the first few samples while the statistic's variance builds up.

Choosing lambda and L

Lambda is the sensitivity dial: lambda = 1 reproduces a Shewhart chart exactly, while smaller values lengthen the memory and sharpen sensitivity to smaller shifts.

• Lambda 0.05 to 0.10: longest memory; best when a 0.5 sigma drift already costs money, and the most robust choice for non-normal readings.
• Lambda 0.20: the standard general purpose value, tuned for shifts around 1 sigma.
• Lambda 0.40: shorter memory for shifts near 1.5 sigma; beyond that a plain Shewhart chart is fast enough.

Pair lambda with L from published average run length (ARL) tables so the in-control false alarm rate matches a 3 sigma Shewhart chart: L = 2.62 for lambda 0.05, L = 2.81 for lambda 0.10, L = 2.96 for lambda 0.20.

Worked example: a filling line drifting one sigma

A bottling line targets 250.0 ml with sigma 2.0 ml, charting individual bottles. With lambda = 0.2 and L = 3, the EWMA standard deviation is 2.0 times the square root of (0.2 / 1.8), or 0.667 ml, so the limits are 248.0 and 252.0. After a changeover error the true mean shifts to 252.0 ml, a one sigma overfill. Starting the EWMA at 250.00:

1. Reading 252.8: EWMA = 0.2 x 252.8 + 0.8 x 250.00 = 250.56
2. Reading 251.5: EWMA = 250.75
3. Reading 253.1: EWMA = 251.22
4. Reading 252.2: EWMA = 251.42
5. Reading 251.9: EWMA = 251.52
6. Reading 253.4: EWMA = 251.90
7. Reading 252.6: EWMA = 252.04, above the UCL of 252.00. Signal.

Every raw reading sat between 251.5 and 253.4, well inside the Shewhart limits of 244 to 256, so an individuals chart would have stayed silent for dozens more bottles while the line gave away product on every fill.

EWMA versus Shewhart versus CUSUM

• Shewhart: fastest for shifts above about 2 sigma and simplest to read. Keep one alongside the EWMA, whose memory creates inertia: a statistic sitting near one limit reacts slowly to a sudden jump the other way.
• CUSUM: statistically comparable to EWMA for small shifts, but its tabular form with k and h parameters is harder to explain at the line.
• EWMA: near-CUSUM detection power, one intuitive tuning knob, works naturally on individual measurements, and tolerates moderate non-normality when lambda is small.

Making EWMA work on the shop floor

1. Validate the measurement system first with a gauge R&R study; the chart will accumulate measurement error just as faithfully as real drift.
2. Estimate the target and sigma from a stable in-control baseline, and confirm capability with Cp and Cpk before tightening surveillance.
3. Write the response into the control plan: who reacts to a signal, how fast, and with what containment steps.
4. Automate data capture. A chart designed to catch drifts within 10 samples is wasted if readings arrive on a clipboard at shift end.

Where Fabrico fits

An EWMA chart is only as good as the data feeding it, and that data layer is what Fabrico provides: real-time OEE and production monitoring, including computer vision on machines with no PLC, so engineers work from live, trustworthy machine data instead of end-of-shift transcription. When a chart signals, Fabrico's field-ready CMMS turns the alarm into action: a work order with asset history attached, preventive schedules adjusted, spare parts checked, all with an audit trail. Built in the EU with EU data residency, Fabrico is the real-time data foundation underneath whatever SPC tooling your quality team runs on top.

Frequently Asked Questions

What value of lambda should I start with?

Start with lambda = 0.2 and L close to 3. If your costly failure mode is a very slow drift near 0.5 sigma, drop lambda to 0.05 or 0.10 and set L from an ARL table to keep the false alarm rate acceptable.

Can I use an EWMA chart on individual measurements?

Yes, that is its most common use. Estimate sigma from the average moving range of a stable baseline. With small lambda the EWMA statistic is close to normal even when single readings are skewed, making it more robust than an individuals chart.

How is EWMA different from a simple moving average chart?

A simple moving average weights the last n points equally and drops the oldest abruptly, causing jumps when an extreme value leaves the window. EWMA weights decay smoothly and never fully drop old information, giving steadier detection at the same false alarm rate.

Ready to feed your control charts with live production data and turn every signal into an assigned work order? Book a Fabrico demo.