Advanced Sequential Monitoring

Introduction

Traditional Shewhart charts are effective for large, sudden changes, but they may signal slowly when a process mean moves only slightly. In high-capability processes, safety-critical production, and low-defect services, a small persistent shift can create substantial cumulative loss before a conventional chart reacts. Sequential monitoring addresses this weakness by accumulating evidence over time. The two principal tools are the cumulative sum (CUSUM) chart and the exponentially weighted moving average (EWMA) chart.

These methods are not merely alternative plotting formats. They represent different evidence-accumulation models, require explicit choices about the shift of interest, and should be evaluated through run-length distributions rather than only average signal delay. Their design must also consider false alarms, inspection cost, investigation capacity, process autocorrelation, parameter uncertainty, and the consequences of delayed detection.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Explain why sequential charts detect small sustained shifts efficiently.
2. Construct and interpret standardized CUSUM and EWMA statistics.
3. Select reference values, decision intervals, smoothing constants, and control limits.
4. Compare charts using in-control and out-of-control run-length performance.
5. Configure a monitoring system according to operational risk and cost.
6. Diagnose limitations caused by autocorrelation, non-normality, estimation error, and adaptive operations.

1. Sequential monitoring as evidence accumulation

Let X_t denote the observation at sampling time t. Under a stable process, the usual model is

where μ₀ is the target mean and σ is the known or reliably estimated standard deviation. A mean shift of size δσ changes the mean to μ₁ = μ₀ + δσ. The standardized observation is

Under the ideal in-control model, Z_t has mean 0 and variance 1. Under the shifted model, its mean is δ. Standardization permits design parameters to be compared across processes, although the original units must be restored when setting operational limits.

A sequential chart uses current and previous observations. Its performance is commonly described by a run length: the number of observations taken before a signal. Important measures include:

If sampling occurs every h hours, a simple conversion is ATS ≈ h × ARL, provided sampling is regular and the initial timing convention is consistent. ARL is not a probability and should not be interpreted as a guaranteed number of observations. An ARL₀ of 370 means an average false signal interval of 370 observations, not that every 370th observation will signal.

2. CUSUM charts

2.1 Derivation and interpretation

The CUSUM is based on adding evidence that observations are above or below target. For a two-sided chart, the standardized one-sided recursions are

with C₀^+ = C₀^- = 0. The statistic C_t^+ accumulates evidence of an upward shift and C_t^- accumulates evidence of a downward shift. The reference value k is normally positive and expressed in standard-deviation units. A signal occurs when

where H is the decision interval or control limit, also in standardized units.

The maximum with zero is a reset: evidence in the wrong direction is discarded rather than allowed to cancel all future evidence indefinitely. This gives the chart sensitivity to a change that begins after a period of normal operation.

A likelihood-ratio argument explains the design. For normal observations with common variance and a target shift of δσ, the log-likelihood contribution, apart from a constant, is proportional to δZ_t − δ²/2. Dividing by δ gives Z_t − δ/2. Thus, for a chart optimized for a specified shift, the usual choice is

This relationship applies to the standardized normal-mean model with known variance. It is a design principle, not a universal rule for non-normal, autocorrelated, or variance-shift problems.

2.2 Choosing the reference value and decision interval

A smaller k makes the chart more responsive to smaller shifts but can increase false alarms or reduce performance for larger shifts. A larger H generally increases ARL₀ and delays detection. The pair (k,H) must therefore be selected jointly. Published tables and software can provide H values for a desired ARL₀, but the implementation must match the assumed chart: one- or two-sided, tabular or standardized, head-start or zero-state, and known or estimated parameters.

A tabular CUSUM may also be written in original measurement units. If the target shift is Δ units and K = kσ, then

and the signal rule is S_t^+ ≥ Hσ when H is standardized. The standardized form is usually less error-prone for design; the raw-unit form is convenient for shop-floor interpretation.

2.3 Worked CUSUM example

Suppose a filling process has target μ₀ = 500 ml and known standard deviation σ = 2 ml. Management wishes to detect an upward shift of one standard deviation, so δ = 1 and k = 0.5. Let H = 4.5 for illustration. For observations 501.4, 502.0, 500.8, 501.6 and 502.2 ml, the standardized values are 0.70, 1.00, 0.40, 0.80 and 1.10.

The upward recursion gives:

No signal occurs because 1.50 is less than 4.5. Notice that the statistic remains positive despite the third observation being below the target: it reduces, but does not erase, accumulated upward evidence.

3. EWMA charts

3.1 Definition and properties

The EWMA statistic is a weighted average that gives greatest weight to recent observations while retaining memory of earlier observations:

Usually Y₀ = μ₀. In standardized form, with Z_t as above and W₀ = 0,

Expanding recursively shows that older observations receive geometrically declining weights:

The weights sum to one only in the limiting sense; the final term represents initialization. Smaller λ produces stronger smoothing and greater sensitivity to small persistent shifts, but it also makes the chart slower to respond to abrupt changes and more dependent on initialization.

For independent standardized observations, the variance of W_t is

This is a dimensionless variance. The steady-state variance is λ/(2−λ). Consequently, time-varying control limits are

and the signal rule is |W_t| > L_t. In original units, the center line is μ₀ and the half-width is σL_t. These limits assume independent observations, a stable variance, correct target and standard deviation, and the stated initialization. Steady-state limits may be used after the start-up effect is negligible, but this choice must be documented.

3.2 Worked EWMA example

Use the five standardized observations above and choose λ = 0.20, with W₀ = 0. The successive values are:

The EWMA moves gradually toward the elevated observations. If a design limit were, for example, ±1.5 in standardized EWMA units, no signal would occur. The number is illustrative: a valid operational limit must be calibrated to a specified false-alarm performance, chart convention, and initialization.

4. CUSUM and EWMA: comparison and selection

CUSUM is explicitly directional and accumulates excess over a reference value. EWMA continuously smooths observations and retains a decaying memory. Neither dominates in every situation.

A fair comparison fixes the in-control objective, sampling frequency, parameter-estimation method, and signal convention. Comparing a CUSUM with one ARL₀ to an EWMA with another is not informative. For a range of possible shifts, plot ARL₁(δ) or ATS against δ. This reveals whether a method is robust or optimized narrowly for one shift.

5. Run-length analysis and design evaluation

Run length is random even when the process model is fixed. ARL alone can conceal highly variable performance. For a design study, estimate the run-length distribution using exact Markov-chain methods, numerical integration, or Monte Carlo simulation. A basic simulation procedure is:

1. Generate independent observations from the in-control or shifted model.
2. Update the chart recursively until a signal.
3. Record the run length.
4. Repeat many times using a fixed random seed for reproducibility.
5. Report ARL, median, selected percentiles, and SDRL with simulation uncertainty.

For an estimated mean from N independent simulation runs, the Monte Carlo standard error is approximately

where s_R is the sample standard deviation of simulated run lengths. Because run lengths can be highly skewed, report percentiles as well as means. Zero-state performance starts from a fresh chart at the shift point. Steady-state performance assumes the chart has operated in control before the shift, so the starting statistic has its in-control distribution. These are different questions and should not be conflated.

A practical comparison table might include δ = 0, 0.25, 0.5, 1.0, 2.0. The value at zero assesses false alarms; the other values assess detection. If the process can shift in either direction, evaluate both signs. If the process is sampled in subgroups, use the correct subgroup statistic and its standard error rather than treating individual observations as independent.

6. Configuring monitoring for risk and cost

Statistical sensitivity is not the only design criterion. A false alarm consumes investigation time, may cause unnecessary adjustment, and can itself destabilize a controlled process. A missed or delayed signal may create scrap, rework, warranty exposure, environmental harm, or patient risk.

Let C_FA be the cost of a false alarm, C_D the cost per sampling interval of operating after a shift before detection, and C_S the cost of sampling and routine monitoring. A simplified expected-cost model is

where f is the sampling frequency in observations per unit time. This model is deliberately approximate: it assumes a stationary shift scenario and does not include all consequences, such as investigation duration or finite production lots. It is useful for comparing candidate designs, not for claiming an exact economic optimum.

For high-consequence processes, constraints should precede cost minimization. For example, require a minimum ARL₀, a maximum ATS for a critical shift, and a maximum investigation workload. Then choose among feasible designs using cost, operator burden, interpretability, and maintainability. Sampling more frequently can reduce detection time but increases inspection cost and potential false-alarm opportunities per unit time.

A sound design specification records:

• quality characteristic, target, units, and rational subgroup;
• sampling interval and subgroup size;
• assumed distribution and independence structure;
• parameter-estimation data and uncertainty;
• chart type, sidedness, initialization, and head-start policy;
• k,H or λ,L and the design shift;
• in-control and out-of-control performance evidence;
• signal investigation and restart rules;
• authority for process adjustment and model revision.

7. Implementation in real operations

Phase 1: establish a trustworthy baseline

Use a stable reference period, remove only assignable causes supported by evidence, and avoid deleting inconvenient observations merely to improve chart performance. Estimate μ₀ and σ using methods appropriate to the data. If parameters are estimated from a small sample, account for estimation uncertainty through simulation or a validated design procedure.

Phase 2: select the monitoring unit

Sequential charts assume that the plotted observations have a meaningful order and comparable measurement conditions. Rational subgrouping should minimize within-subgroup special-cause variation while preserving sensitivity to between-subgroup changes. Mixing products, tools, shifts, or operating modes can create apparent shifts that are actually changes in population.

Phase 3: validate assumptions

Check serial dependence, seasonality, measurement-system stability, skewness, and changing variance. Autocorrelation violates the independent-observation variance formulas and can cause excessive signals or excessive masking. Options include modeling the process and charting residuals, changing the sampling interval, using a variance-adjusted design, or applying a method specifically developed for dependent data. A normal probability model is often a useful approximation, but it should not be accepted solely because a software default uses it.

Phase 4: define response rules

A signal is an investigation trigger, not automatic proof that the process is defective. The response should identify the assignable cause, contain potentially affected output, verify the measurement system, and document disposition. Repeatedly resetting a chart without investigation destroys the meaning of ARL and creates hidden risk. A restart policy should specify when the process is considered stable and whether the statistic returns to zero or retains a justified head start.

8. Illustrative case: precision dispensing line

The following is a fictional illustrative case.

Background and problem

A medical-device supplier dispenses adhesive into cartridges. The target mass is 12.00 g, the historical standard deviation is 0.06 g, and sampling occurs every 20 minutes. A shift of 0.12 g is economically important because it can impair bonding. The existing Shewhart chart has few false alarms but detects gradual nozzle wear late.

Analysis

The standardized shift is δ = 0.12/0.06 = 2. Engineering review also identifies smaller drift, around 0.5 to 1 standard deviation, as operationally valuable to detect. A CUSUM designed only for δ = 2 may be less sensitive to gradual wear than one designed for δ = 0.5; an EWMA can provide broader sensitivity. The team therefore simulates candidate designs at δ = 0, ±0.5, ±1, ±2, retaining the same sampling interval and a common target ARL₀.

Approach and solution

The selected system uses an EWMA as the primary drift detector, with a small smoothing constant chosen after run-length evaluation, and a CUSUM as a diagnostic chart optimized for the economically critical shift. Limits are calibrated by simulation rather than copied from a table because the baseline parameters are estimated and the measurement data show mild serial dependence. A Shewhart rule remains in place for very large individual excursions.

The response plan classifies signals into: immediate containment for large excursions; verification and nozzle inspection for sequential-chart signals; and engineering review when repeated near-signals occur without a formal alarm. The team also monitors false-alarm workload per shift, not just theoretical ARL₀.

Outcome and learning points

After a pilot period, the organization chooses the design with acceptable false-alarm workload and adequate ATS for the 0.5–1 standard-deviation drift. The main lesson is architectural: complementary charts can be justified when they have distinct purposes, but overlapping alarms must be governed to prevent confusion and multiple-testing inflation. The second lesson is that simulation must represent the actual sampling, dependence, parameter uncertainty, and restart policy.

9. Limitations, governance, and contemporary practice

Sequential charts can be misused when management treats every signal as operator failure, changes limits after inconvenient alarms, or optimizes only average detection delay. Data integrity is essential: timestamps, missing observations, manual overrides, and chart resets should be auditable. In regulated settings, validation should establish intended use, software version, parameter approval, access control, and change management.

AI and advanced analytics can support charting by detecting regime changes, estimating autocorrelation, prioritizing investigations, and simulating candidate policies. They should not replace transparent statistical controls without validation. Black-box alerts can be difficult to explain, reproduce, or defend in a quality investigation. A strong governance model retains a human-review path, monitors model drift, separates exploratory analytics from released control rules, and prevents target leakage from future information.

Contemporary implementations increasingly combine sequential charts with adaptive sampling, multivariate monitoring, Bayesian updating, and maintenance data. Such extensions can improve performance but also invalidate simple ARL formulas. Whenever sampling, limits, or models adapt in response to the data, performance should be evaluated under the complete adaptive policy rather than under an isolated chart equation.

Best-practice checklist

1. Define the harmful shift and its operational consequences before choosing a chart.
2. Standardize carefully, but preserve original units for decisions.
3. Calibrate k,H or λ,L to a documented ARL₀ and detection objective.
4. Compare ARL distributions across a range of shifts, not only one average.
5. Include initialization, estimated parameters, subgrouping, and dependence in evaluation.
6. Use signals to trigger structured investigation, not automatic adjustment.
7. Record resets, overrides, missing data, and limit changes.
8. Revalidate after product, equipment, measurement, sampling, or software changes.

Chapter Summary

CUSUM and EWMA charts improve sensitivity to small sustained shifts by accumulating evidence over time. CUSUM uses directional recursions with reference value k and decision interval H; its conventional normal-model design often sets k = |δ|/2 for a targeted standardized shift. EWMA uses smoothing constant λ and control multiplier L, with time-dependent variance during start-up and steady-state variance λ/(2−λ).

Neither chart is universally superior. Appropriate selection depends on the shift profile, false-alarm tolerance, sampling cost, consequences of delay, dependence structure, and investigation capacity. ARL, median run length, SDRL, percentiles, and ATS should be evaluated together, preferably through validated simulation or exact numerical methods. Operationally credible monitoring also requires stable baselines, reliable measurement, documented response rules, auditable governance, and periodic revalidation.

Key Terms

Sequential monitoring: Monitoring that uses current and historical observations jointly.

CUSUM: A chart that accumulates deviations from a reference value.

EWMA: An exponentially weighted moving-average chart.

Standardized observation: An observation expressed in standard-deviation units from target.

Reference value (k): CUSUM allowance parameter for the targeted shift.

Decision interval (H): CUSUM threshold for signaling.

Smoothing constant (λ): EWMA weight assigned to the newest observation.

Control multiplier (L): EWMA limit expressed in its standard-error units.

Run length: Number of observations until a chart signals.

ARL₀: Mean in-control run length.

ARL₁: Mean run length after a specified shift.

ATS: Average time to signal.

SDRL: Standard deviation of run length.

Zero-state performance: Run-length performance starting from a fresh chart.

Steady-state performance: Performance after the chart has reached its in-control distribution.

Rational subgroup: A sampling group designed to separate within- and between-process variation.

Head start: A nonzero initial chart statistic used to improve sensitivity under a defined policy.

Autocorrelation: Dependence between observations at different times.

Assignable cause: A specific, investigable source of process change.

Review Questions

Short Answer Questions

1. Why are CUSUM and EWMA charts generally more sensitive than Shewhart charts to small sustained shifts?
2. Define ARL₀ and explain why it is not a guarantee about the next signal.
3. What roles do k and H play in a CUSUM chart?
4. How does reducing λ affect the memory and responsiveness of an EWMA chart?
5. Why must autocorrelation be assessed before applying standard EWMA limits?

Long Answer Questions

1. Derive the standardized two-sided CUSUM recursions and explain the resetting operation.
2. Explain the EWMA variance formula, including the distinction between start-up and steady-state limits.
3. Compare CUSUM and EWMA for abrupt shifts, gradual drift, interpretability, and parameter selection.
4. Describe a rigorous run-length simulation study for selecting between two sequential charts.
5. Develop a governance and response policy for a sequential chart in a regulated production process.

Analytical / Application Questions

1. A process has μ₀ = 100 units and σ = 1.5 units. For a target upward shift of 0.75 units, calculate δ and the conventional CUSUM reference value k = |δ|/2. State the assumptions behind the calculation.
2. Design a comparison of a CUSUM and an EWMA for a process where false alarms are expensive but a 0.5-standard-deviation shift must be detected quickly. Specify performance measures, simulation scenarios, and decision criteria.
3. A chart produces many alarms during a period of seasonal demand. Propose a diagnostic investigation that distinguishes genuine process shifts from model misspecification, autocorrelation, changing variance, or inappropriate subgrouping.

Multiple Choice Questions with Answers

Question 1

For a standardized normal-mean shift of δ = 1, the conventional CUSUM reference value is:

A. 0.25

B. 0.50

C. 1.00

D. 2.00

Correct Answer: B. 0.50

Question 2

Which statistic accumulates upward evidence in a two-sided CUSUM?

A. C_t^+ = max(0,C_t−1^+ + Z_t − k)

B. C_t^+ = C_t−1^+ − Z_t − k

C. C_t^+ = λZ_t + (1−λ)C_t−1^+

D. C_t^+ = max(0,k − Z_t)

Correct Answer: A. C_t^+ = max(0,C_t−1^+ + Z_t − k)

Question 3

In an EWMA, decreasing λ generally:

A. Eliminates all memory of earlier observations

B. Increases sensitivity to a single abrupt observation

C. Creates stronger smoothing and more memory

D. Makes the statistic identical to a Shewhart statistic

Correct Answer: C. Creates stronger smoothing and more memory

Question 4

ARL₀ measures:

A. Mean detection delay after a specified shift

B. Mean in-control run length

C. Standard deviation of process measurements

D. Sampling cost per observation

Correct Answer: B. Mean in-control run length

Question 5

For independent standardized observations, the steady-state variance of a standardized EWMA is:

A. λ/(2−λ)

B. (2−λ)/λ

C. λ(2−λ)

D. 1/λ²

Correct Answer: A. λ/(2−λ)

Question 6

Why should run-length percentiles be reported in addition to ARL?

A. ARL is always a probability

B. Run-length distributions can be skewed and variable

C. Percentiles replace the need for an in-control design

D. ARL cannot be estimated by simulation

Correct Answer: B. Run-length distributions can be skewed and variable

Question 7

A signal from a sequential chart should normally be treated as:

A. Automatic proof of operator error

B. An investigation trigger

C. Permission to reset limits immediately

D. Evidence that the product must always be scrapped

Correct Answer: B. An investigation trigger

Question 8

If sampling occurs every 20 minutes and the average run length is 15 observations, the approximate ATS is:

A. 0.75 minutes

B. 15 minutes

C. 35 minutes

D. 300 minutes

Correct Answer: D. 300 minutes

Question 9

Which condition invalidates the usual independent-observation EWMA variance formula unless addressed?

A. Observations are recorded in chronological order

B. The target is expressed in engineering units

C. Substantial autocorrelation exists

D. The EWMA is plotted on a control chart

Correct Answer: C. Substantial autocorrelation exists

Question 10

A fair comparison between a CUSUM and EWMA should first ensure that they have:

A. The same chart color

B. The same number of parameters

C. A common in-control performance objective and sampling scheme

D. Identical recursive equations

Correct Answer: C. A common in-control performance objective and sampling scheme