EWMA Control Chart for Drift Detection

🏭

Industry Use

How this chart is applied in pharmaceutical manufacturing

The exponentially weighted moving average (EWMA) chart accumulates evidence across observations, which makes it more sensitive than a Shewhart individuals chart to small, sustained process shifts and gradual drift. Introduced by Roberts (1959), it is a standard tool for Stage 3 continued process verification, where slow degradation over months matters more than isolated spikes.

Where this is used:

Trending potency, moisture, environmental, or in-process data for slow drift.
Stage 3 continued process verification under the FDA process-validation lifecycle.
Complementing (not replacing) a Shewhart chart: EWMA catches small sustained shifts; Shewhart catches large abrupt ones.

∑

Calculation Explanation

Mathematical basis and choosing λ

EWMA recursion

zᵢ = λ·xᵢ + (1 − λ)·zᵢ₋₁ , z₀ = target (or historical mean)

Variance and control limits (time-varying)

Limits(i) = z₀ ± L·σ·√( [λ/(2−λ)]·[1 − (1−λ)^(2i)] )

Steady state (i → ∞): z₀ ± L·σ·√[λ/(2−λ)]. Use the time-varying limits for the first observations — the EWMA's variance grows toward its asymptote, so early limits are tighter than the steady-state pair, and comparing early points against steady-state limits alone can miss a start-up signal.

Choosing λ — the weight on the current observation. A small λ (0.05–0.2) carries long memory: maximum smoothing and the best sensitivity to small sustained shifts, at the cost of responding slowly to large abrupt shifts. A large λ weights recent data more, smooths less, and responds faster; at λ = 1 the chart degenerates to a Shewhart individuals chart. λ between 0.2 and 0.3 is the usual recommendation (Hunter 1986; NIST/SEMATECH e-Handbook), with L ≈ 2.7–3. Lucas and Saccucci (1990) tabulate (λ, L) pairs against average run length.

Roberts (1959) Hunter (1986) Lucas & Saccucci (1990) NIST e-Handbook §6.3.2.4

📋

Variable Definitions

All input parameters and their meaning

Symbol	Variable	Units	Description
xᵢ	Current observation	product units	Latest measurement (potency %, moisture %, …).
zᵢ	EWMA statistic	product units	Exponentially weighted average at time i — the plotted value.
λ	Smoothing parameter	dimensionless	Weight on the current observation, 0 < λ ≤ 1. Small λ = long memory, best for small sustained shifts; λ = 1 = Shewhart individuals. Typical 0.2–0.3.
z₀	Starting value	product units	The process target or historical mean.
σ	Process standard deviation	product units	A sound short-term estimate (e.g. from a moving range) — an estimate inflated by drift widens the limits and hides the drift.
L	Control-limit width multiplier	dimensionless	≈ 2.7–3, chosen with λ (Lucas & Saccucci tables).
i	Observation number	count	Drives the time-varying variance term [1 − (1−λ)^(2i)].

📖

Step-by-Step Tutorial

Worked example: granule moisture, time-varying limits applied

Scenario: granule moisture, target 2.2%, σ = 0.15%, λ = 0.2, L = 3. Eight daily readings: 2.1, 2.3, 2.0, 2.4, 2.2, 2.5, 2.3, 2.1%.

Compute the EWMA series (z₀ = 2.2).

Calculation
z₁ = 0.2×2.1 + 0.8×2.2 = 2.180
z₂ = 2.204 · z₃ = 2.163 · z₄ = 2.211 · z₅ = 2.208
z₆ = 2.267 · z₇ = 2.273 · z₈ = 2.239
Compute the time-varying limits for the early points.

Calculation
i = 1: 2.2 ± 3×0.15×√(0.1111×[1−0.64¹]) → 2.110 – 2.290
i = 2: 2.085 – 2.315 · i = 3: 2.071 – 2.329
By i = 8: 2.052 – 2.348 — essentially the steady-state pair.
Compute the steady-state limits.

Calculation
2.2 ± 3 × 0.15 × √(0.2/1.8) = 2.2 ± 0.15 → 2.05 – 2.35
Assess: every zᵢ is inside its own (time-varying) limits — in control — but z₆–z₈ sit well above center: a developing upward drift worth investigating before it signals.

✓ In control against time-varying limits · upward drift developing at z₆–z₈ — investigate the moisture source

⚠️

Interpretation, Assumptions and Limitations

What a signal means — and what the chart presumes

Signals: a point outside its EWMA limits signals a shift; because the statistic carries memory, the shift typically began before the signal. Investigate per SOP and re-initialize once the cause is corrected.
Tuning: match λ to the shift size that matters — smaller λ for smaller sustained shifts.
Independence: EWMA assumes serially independent observations; strong autocorrelation distorts the limits.
Mean shifts only: a companion chart is needed for variability.

📚

References

Primary sources for this guide

Roberts, S.W. (1959). Control Chart Tests Based on Geometric Moving Averages. Technometrics 1(3):239–250.
Hunter, J.S. (1986). The Exponentially Weighted Moving Average. Journal of Quality Technology 18(4):203–210.
Lucas, J.M. and Saccucci, M.S. (1990). Exponentially Weighted Moving Average Control Schemes: Properties and Enhancements. Technometrics 32(1):1–12.
NIST/SEMATECH e-Handbook of Statistical Methods, §6.3.2.4 (EWMA Control Charts). itl.nist.gov
FDA, Process Validation: General Principles and Practices (2011) — Stage 3 context. fda.gov