EWMA
Catches small, sustained shifts the raw chart misses.
An EWMA chart catches the slow, sustained drift a standard control chart misses. It weights recent points and accumulates the trend, so a small shift crosses the limit early, long before any single noisy point breaches.
What it measures
EWMA (exponentially weighted moving average) builds a running statistic that remembers the recent past with a tunable memory:
- The statistic: zₜ = λxₜ + (1 − λ)zₜ₋₁. Each new point xₜ is blended with the previous EWMA value, so the line is a smoothed, memory-weighted version of the process.
- The control limits: zₜ stays inside ±L·σ·√[λ/(2 − λ)], a band much tighter than the ±3σ of a raw chart because the EWMA's variance is reduced by averaging.
- The λ trade-off: small λ (0.05 to 0.2) gives long memory and high sensitivity to small, sustained shifts, but slower response to a large sudden jump. Larger λ behaves more like a raw chart. λ is chosen for the size of shift you most need to catch.
The reason it works: a 0.5σ drift hides easily inside the noise of individual points, but it biases every point slightly in one direction, and the EWMA accumulates that bias until the smoothed line crosses a limit the raw points never would.
How to read the output
Read the EWMA line, not the raw scatter. When it drifts steadily toward a limit, a small sustained shift is underway even though no individual point looks alarming; the crossing is the signal, and it typically arrives well before a 3σ breach on the raw chart. Match λ to the job: if you are chasing slow drifts (tool wear, gradual contamination, a creeping temperature), use a small λ for long memory; if you also need fast response to big jumps, EWMA is the wrong tool for those and a Shewhart or Western Electric overlay should run alongside. A flat EWMA hugging the centerline is a process genuinely on target, which is its own useful confirmation.
A real use case
An electrode coating line holds areal coating weight inside a tight window. A slipping pump gradually shifts the mean by about half a sigma over a shift, far too small to trip any single point past the 3σ control limit, so a standard chart stays green while every cell coated that shift drifts toward the low edge of spec. An EWMA chart on coating weight, tuned with a small λ, accumulates the bias and crosses its limit hours before the raw chart would have, pointing maintenance at the pump while the lot is still recoverable. The drift that a Shewhart chart only catches after it has become scrap is exactly the drift EWMA exists to see.
Common mistakes
- Using EWMA limits as if they were 3σ Shewhart limits. The EWMA band is tighter by construction; applying raw-chart limits defeats the method.
- Choosing λ without a target shift size. Too large and it misses small drifts; too small and it lags real step changes.
- Running EWMA alone and missing large sudden jumps, which a Shewhart or run-rule chart catches faster. The two are complements.
- Charting an EWMA on a measurement system that cannot resolve the shift, see Gauge R&R, so the drift you are smoothing for is below the gauge's own noise.
- Reacting to a single EWMA point near the limit instead of the sustained trend the chart is designed to surface.
Drift caught early, before the raw chart would breach
Niobia computes the EWMA statistic on the process signals you monitor, sets the memory-corrected control limits, and tracks the smoothed line against them so a small sustained shift is flagged as it develops rather than after it becomes a breach. It runs alongside the rest of the SPC layer, the Shewhart charts for sudden jumps, run rules for patterns, so slow drift and step changes are both covered, and a developing excursion becomes an alert to the right engineer while the lot is still recoverable.
Frequently asked
When should I use EWMA instead of a standard control chart?
When the failure mode is slow, sustained drift rather than a sudden jump: tool wear, gradual contamination, a creeping setpoint. EWMA accumulates the small bias those produce and crosses its limit early, where a Shewhart chart waits for a single point to breach 3σ.
How do I choose the smoothing factor λ?
By the size of shift you most need to catch. Small λ (0.05 to 0.2) gives long memory and high sensitivity to small drifts but slower response to large jumps; larger λ behaves more like a raw chart. A common starting point for small-shift detection is λ around 0.1 to 0.2.
Does EWMA replace the Shewhart chart?
No, it complements it. EWMA wins on small sustained shifts; the Shewhart chart and run rules win on large sudden changes. Running them together covers both failure modes, which is how Niobia's SPC layer is set up.